Damage prediction for wind turbines using wireless sensor and actuator networks

Journal of Network and Computer Applications 80 (2017) 123–140

Contents lists available at ScienceDirect

Journal of Network and Computer Applications journal homepage: www.elsevier.com/locate/jnca

Damage prediction for wind turbines using wireless sensor and actuator networks

MARK

⁎

Maicon Melo Alvesb,c, Luci Pirmeza, Silvana Rossettoa, Flavia C. Delicatoa,c, , Claudio M. de Fariasa, Paulo F. Piresa,c, Igor L. dos Santosa, Albert Y. Zomayac a b c

PPGI-DCC/IM, Universidade Federal do Rio de Janeiro, RJ 20001-970, Brazil Petróleo Brasileiro S.A., Petrobras, Brazil Centre for Distributed and High Performance Computing, School of Information Technologies, The University of Sydney, NSW 2006, Australia

A R T I C L E I N F O

A BS T RAC T

Keywords: Wireless sensor and actuator networks Structural health monitoring Damage prediction ARIMA model Adaptive Neuro Fuzzy Inference System (ANFIS)

The depletion of oil and gas reserves is bringing up economic, political and social issues which encourage the adoption of renewable, green energy sources. Wind energy is a major source of renewable energy because of the maturity and competitive costs of technological solutions to exploit this type of green energy. This kind of power generation is achieved through the use of wind turbines, which convert translational kinetic energy into rotational kinetic energy. The benefits already proven of this type of renewable energy source have motivated nations worldwide to adopt policies to improve the use of wind energy in order to minimize their dependence on oil and natural gas. However, the adoption of wind turbines poses several challenges. A key challenge is properly and timely identifying structural damages which affect the structural health of the wind turbine. In this context, we propose a damage prediction system for wind turbines based on wireless sensor and actuator network. The proposed system, called Delphos, is a decentralized system where all decision-making process is performed within the network, in a collaborative way by the nodes. The purpose of Delphos is to accurately predict when the turbine will reach a damage state, thus allowing timely actions on the turbine operation to prevent accidents, reducing maintenance costs and delays in the power generation. Delphos relies on a time series forecasting model, called ARIMA, and a fuzzy system to eliminate the influence of temperature in the process of damage prediction.

1. Introduction The depletion of oil and gas reserves is bringing up economic, political and social issues that encourage the adoption of renewable energy sources. In this context, wind energy is one of the most promising renewable energy sources because of its technological maturity and competitive costs. This kind of power generation is achieved through the use of wind turbines, which convert translational kinetic energy into rotational kinetic energy. The benefits already proven of this type of renewable energy source have motivated nations worldwide to adopt policies to improve the use of wind energy in order to minimize their dependence on oil and natural gas (Ackermann, 2005). However, the adoption of wind turbines poses several challenges. One major challenge is properly and timely identifying structural damages that affect the structural health of the wind turbine, as illustrated in Fig. 1. According to Swartz et al. Swartz et al. (2010), a wind turbine is an expensive asset that suffers, in average, three

⁎

incidents per year. Therefore, a damage prediction system, capable of identifying imminent incidents and performing control actions, is clearly applicable and useful in this context. Such system fits into the Structural Health Monitoring (SHM) application domain (Sohn et al., 2004). SHM systems are able to detect, locate and predict the evolution of structural damage, avoiding accidents as, for instance, a total breakdown of the structure. In the case of wind turbines, the use of damage prediction techniques allows (i) preparing, in advance, the required resources to perform the repair procedures; (ii) minimizing unnecessary component replacements; and (iii) avoiding the wind turbine inactivity, due to the equipment breakdowns. The vibration analysis is a well-known method that allows assessing the structure condition. In such method, the natural frequencies of the structure are analyzed in order to identify the occurrence of damage. In other words, the natural frequencies are the primary information used to detect, predict or localize damage in a structure when using vibration analysis. These natural frequencies are influenced by environmental

Corresponding author at: PPGI-DCC/IM, Universidade Federal do Rio de Janeiro, RJ 20001-970, Brazil. E-mail address: [email protected] (F.C. Delicato).

http://dx.doi.org/10.1016/j.jnca.2016.12.027 Received 21 June 2016; Received in revised form 20 November 2016; Accepted 15 December 2016 Available online 20 December 2016 1084-8045/ © 2016 Elsevier Ltd. All rights reserved.

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

of the natural frequencies of vibration of the monitored structure. Wireless sensor networks (WSN) (Akyildiz and Vuran, 2010) are a promising choice for performing structural health monitoring. Such networks are composed of smart sensors; one or more sink nodes and can also include actuators (in this case being called wireless sensor and actuator networks – WSAN). Smart sensors are small devices with (constrained) computational power, wireless communication and sensing capabilities (e.g. humidity, temperature and acceleration) (Lim, 2011). Unlike traditional sensors, these smart sensors can perform computational tasks as well as making intelligent decisions upon the collected data. A sink node is typically a more powerful device in terms of processing, storage and energy capacities, and is responsible for connecting the WSAN with external applications and/or networks (as the Internet, for instance). Actuators are devices able to act on the environment and perform control actions in response to decisions taken by the smart sensors. In the context of SHM applications, various sensors are attached to buildings/machinery/structures to acquire data containing physical or environmental states of these mechanical structures. Typical sensing units used for this type of application are accelerometers and strain sensors. The adoption of WSAN for monitoring wind turbines brings several advantages when compared to other approaches (e.g., in situ inspections and wired sensor networks). One advantage is to achieve different viewpoints about the state of the structure with minimal physical impact, since wireless sensors are typically light devices of reduced size and they do not require power and transmission cables. Another advantage is that each device is able to measure different physical quantities (e.g., acceleration, temperature, humidity and wind speed), and such measurements can be used together to ascertain the state of the turbine with potentially higher accuracy. Moreover, WSAN nodes are able of locally processing the data collected to make decisions about the state of the structure and to implement actions to locally control the

Fig. 1. Examples of structural damages in wind turbines (Chou et al., 2013).

parameters, such as temperature and humidity. Several works as (Xia et al., 2006; Moser and Moaveni, 2011) have reported that the values of natural frequencies can decrease or increase when these environmental parameters suffer variations. Therefore, the influences of such environmental parameters can be enough to mask any changes due to damages to a degree that it might not be detected (Sohn, 2006; Croxford et al., 2010). More specifically, the temperature not only modifies the material stiffness, but also alters the boundary conditions of a structure, thus hiding variations in the natural frequencies of the monitored structure (Sohn, 2006). Therefore, in order to achieve a high accuracy in the results of a vibration analysis, it is important to consider the influence of environmental parameters over the variation

Fig. 2. Overview of Delphos Operation.

124

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

et al., 2013) argue that, when compared with fuzzy inference systems and artificial neural networks, ANFIS not only has advantages of the two methods but also makes up for their shortcomings. On one hand, it has an effective self-learning mechanism. On the other hand, it has a variety of neural networks, it is able to optimize the control rules, and expresses the reasoning-function like human brain, while not requiring the presence of experts. It makes the system develop towards adaptive, self-organizing and self-learning, as discussed in (Zhang et al., 2014), which is a very appealing feature to be used in the context of SHM applications. After the filtering performed to reduce environmental interference, Delphos uses the natural frequencies to predict when damage will occur, by using the time series statistical model ARIMA. This model is used to predict the future values of the natural frequencies of the blade, based on the history of natural frequencies collected over time. Other methods such as Bayesian classification or decision tree may be used to perform the prediction of damage (Salfner et al., 2010). However, the limited resources of the smart sensors motivated the adoption of a statistical model. It is important to employ techniques demanding simple computation in order to save energy and suit the low computational power of WSAN devices. Therefore, ARIMA was chosen since it is based on a non-complex calculation that uses pre-identified coefficients (defined by a process that is executed outside the network). The main original contributions of this work are threefold. First, the decision process regarding the damage detection was conceived in a fully decentralized way in order to provide a timely response about potential damage, thus triggering control actions as fast as possible. Second, differently from other works, Delphos increases the accuracy of the prediction process since it reduces the effect of environmental interferences (temperature) on the natural frequencies collected from the monitored structure. Third, Delphos leverages the cooperation among WSAN nodes with the goal of increasing the reliability of the system response regarding the damage prediction. It is also important to mention that, as we will show in the performed evaluation, Delphos is a lightweight system, thus it does not impact negatively the WSAN operation in terms of resource consumption. The remainder of this paper is organized as follows. Section 2 discusses related work. Section 3 presents Delphos architecture and operation. Section 4 describes Delphos implementation. Section 5 presents the experiments performed to evaluate the proposed system. Finally, Section 6 concludes the work and points out future research directions.

behavior of the turbine. The main advantage of performing processing within the network (the so-called in-network processing) is the possibility of promptly acting upon the equipment operation (via actuators), since the system provides results more quickly than a solution based on centralized and remote control centers. In the case of wind turbines, the required action may be to reduce the rotation of the blades or to adjust the system so that the equipment remains in operation only in certain moments. In order to increase the reliability of the prediction process, the nodes of the WSAN may collaborate to achieve a single prediction result, since the conditions of the blade are observed from different perspectives or angles according to each of the sensors arranged on the blade. A further advantage is that the redundancy of the sensors embedded in the blade allows performing fault tolerance techniques, thus providing a robust solution. In this context, the purpose of this work is to introduce a structural damage prediction system, called Delphos, which is based on the use of wireless sensor and actuator networks (WSAN) and prediction models (Fig. 2). Delphos aims to make the automated and remote monitoring of wind turbines in order to predict damage and perform control actions upon the monitored structure as early as possible, thus potentially minimizing maintenance costs and avoiding disastrous consequences. Examples of control actions that can be triggered by the early identification of damage are the emergency shutdown of the monitored turbine or the reduction of the rotational speed of the rotor. For performing damage prediction, Delphos uses the data about the natural frequencies of the blade. Such natural frequencies are calculated from the vibration data sensed from the blade by using accelerometers (sensing units of the WSAN nodes). Delphos takes into account only the first natural frequency mode, which denotes the global behavior of the structure that is enough to predict when damage will occur (Sohn et al., 2004). Delphos is a decentralized system once the whole decision-making process of damage prediction is performed within the network, in a collaborative way among the sensor nodes, instead of sending collected data to a central station in charge of data processing and decision making. This decentralized approach has two major advantages: (i) reduction of the response time, consequently allowing faster execution of control actions and (ii) energy savings of the network nodes, since each node transmits only one final result, instead of the entire set of collected raw data. Delphos adopts two main techniques to achieve its goals: Adaptive Neuro Fuzzy Inference System (ANFIS) (Kurian et al., 2006) and Autoregressive integrated moving average (ARIMA) (Delurgio, 1998). ANFIS is a hybrid system built on Neural Networks and Fuzzy Logic techniques. It is adopted in Delphos before performing the damage prediction process (Fig. 3), in order to reduce the effect of environmental parameters (in the current version, temperature) over the natural frequencies thus increasing the accuracy of the damage prediction process. Therefore, Delphos performs an environmental interference filtering, where the filter is applied before the prediction of frequency values so that the system is able to analyze such values without being affected by the temperature influence. ANFIS was chosen to be used for two reasons. First, the Neural Networks technique allows extracting the knowledge about the behavior related to the environmental interference over natural frequencies, without the need to rely on a civil engineering expert. The Fuzzy Logic technique provides the capability to deal with imprecise terms such as “very”, “average”, and “little”, which are commonly associated to the influence of the temperature over the frequencies of a civil structure. Another reason for using fuzzy logic is that such technique fits the available computational resources provided by the constrained WSAN nodes. Thus, ANFIS is a hybrid of these two techniques, which exploits the advantages of both neural networks and fuzzy logic. In other words, ANFIS combines the low-level computational power of a neural network with the high-level reasoning capability of a fuzzy inference system. Moreover, several works, such as (Zhang et al., 2014; Bououden et al., 2013; Wang et al., 2013; Wu et al., 2012; Chiang

2. Related work In this section, we first present works that perform damage prediction in generic civil structures and wind farms using vibration data and prediction methods. Subsequently, we present studies that have proposed methods to identify and quantify the influence of environmental conditions in the natural frequencies of civil structures. Finally, we present the studies that use WSN as a basic communication and sensing infrastructure to realize, either partially or entirely, the decision process of damage detection in civil structures. Our work performs, similarly to the proposals described in (Pham and Yang, 2010; Pham et al., 2010; Garcia et al., 2010), damage prediction using vibration data and also uses ARIMA, a variation of the ARMA model to forecasting the state of a machine or structure. Concerning the methods to identify and quantify the influence of environmental conditions in the natural frequencies of civil structures, we highlight works (Xia et al., 2006; Moser and Moaveni, 2011; Liu and Dewolf, 2007). By analyzing the aforementioned studies, we concluded that it would not be feasible to take advantage of the methods proposed by them in the context of our work, since the monitored structures were built from materials with different physical properties (our focus is on wind turbines). For this reason, we employed a method capable of filtering the influence of environmental 125

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

Fig. 3. Delphos prediction process.

predicting damage. Delphos also performs the correlation of environmental variables, where the temperature is considered to help improving the interpretation of the natural frequencies behavior, in order to increase the accuracy of the damage prediction process. The proposed system uses WSAN nodes to assess the structure under different viewing angles (each sensor node performs the prediction damage based on their local read vibration) to increase the reliability of the prediction process. Finally, the system also employs actuator nodes to perform fast control actions on the turbine to prevent any catastrophic damage. Only a few works (Miguelanez and Lane, 2010; Kusiak and Li, 2010) perform damage prediction specifically on wind turbine blades. In these works, data mining algorithms are used to predict damage in blades. In both proposals, a pre-installed supervisory system on such blades provides data for the prediction system, located in a centralized computer. Therefore, both damage prediction systems get data from the supervisory system and execute algorithms to predict damage in order to issue warnings on several levels. Thus, differently from our work, such proposals do not perform localized control actions whenever damage is predicted, since they rely on a centralized data analysis and do not communicate with actuator devices. In this sense, the main advantage of Delphos concerns the fast response achieved when the damage occurs, allowing acting in the equipment (wind turbine) before the problem worsens.

conditions in the natural frequencies of any structure by means of a fuzzy system automatically generated by the Adaptive Neuro Fuzzy Inference System (ANFIS). In our case, the ANFIS uses a real database of the structure to extract the correlation between the temperature (physical quantity observed) and vibration (natural frequencies). In (Zhang et al., 2013) the authors proposed a scheme for structural health monitoring based on environmental effect removal. An interesting aspect of such work is the contextualization of the proposed scheme in an Internet of Things (IoT) environment. IoT (Zhang et al., 2013) is an emergent paradigm that combines environmental sensing with data transmission and processing through wireless communication techniques. SHM is one of the several application domains that can highly benefit from IoT features. Wireless sensor networks are major enablers of the IoT paradigm. In (Zhang et al., 2013), the statistical technique of Principal Component Analysis (PCA) was employed to eliminate environment effects from sensor data that contained both real vibration feature of architectural structure and environmental interferences. After the removal of environmental effects, Hilbert-Huang transformation (HHT), a classical method for signal analysis, was used for structural health analysis and monitoring. The technique adopted to remove the environmental influence was generic enough to be applied to diverse variables, such as temperature, humidity and wind. Simulation results showed that the proposed scheme was able to achieve high accuracy in structural health monitoring and robust performance against environmental interferences. The work described in (Zhang et al., 2013), although very relevant and related to our proposal, has the different goal of monitoring the structure, while we aim at predicting damages, so that preventive actions can be taken in a timely way. Several works (Swartz et al., 2010; Kim et al., 2007; Wu et al., 2009; Reyer et al., 2011) use a WSN as sensing and communication infrastructure for the purpose of detecting damage in different structures. In these works, sensor data are transmitted to a central device (the sink node) where a damage detection process is entirely executed. Other works (Hackmann et al., 2008; Bocca et al., 2011; Santos et al., 2014) propose the use of WSN not only to perform sensing and communication, but also to execute local tasks related to the process of damage detection in structures. However, as well as several others reported in the literature (Ji et al., 2012; Bhuiyan et al., 2015), these works focus on a process of damage detection, while our work aims at

3. Delphos: damage prediction system based on WSAN In this paper, we describe our proposal of a damage prediction system for wind turbines, called Delphos, which is able to predict damage and perform control actions so as minimizing resulting risks and losses. Delphos is a decentralized, reliable, and accurate system built on a WSAN (Wireless Sensor and Actuator Network). The decentralized damage prediction process has as one of its advantages to provide a fast response when performing control actions on the turbine. Another advantage concerns the energy savings for the network nodes, since each node transmits only one final result, instead of the entire set of collected raw data. Regarding the reliability, the system uses the different points of view obtained by each individual node to monitor the wind turbine blade, to make decisions about the state of the structure. Furthermore, since the temperature can interfere 126

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

on natural frequencies, the system performs a filtering of temperature values before performing the prediction process, thus increasing the accuracy of the final result. Delphos comprises two main mechanisms, namely: Abnormality Decision and Damage Prediction. The first one is responsible for verifying whether the blade is in a state of abnormality, that is, if the blade is not totally healthy, but at the same time there are no current occurrences of damage. The second mechanism is responsible for predicting when a damage will occur. In this section, we present the theoretical foundation about ANFIS and ARIMA (sub-Section 3.1), the logical and physical architecture of Delphos (sub-Section 3.2), and a description of the system operation (sub-Section 3.3).

Temperature of Ubatuba City (Brazil) 25 20 15

Dec-85

May-85

Oct-84

Mar-84

Jan-83

Aug-83

Jun-82

Nov-81

Sep-80

Apr-81

Feb-80

Jul-79

Dec-78

Oct-77

May-78

Mar-77

Jan-76

10 Aug-76

Temperature (Celsius)

30

Month Fig. 4. Example of time series regarding the seasonal variation of temperature in a Brazilian city (CPTEC/INPE, 2013).

3.1. Theoretical foundation This section presents the theoretical foundation needed for the understanding of the mechanisms used in our proposal. To perform the environmental interference filtering, Delphos makes use of a fuzzy system automatically generated using a technique referred to as ANFIS (described in sub-Section 3.1.1). To perform the real damage prediction, Delphos uses a statistical model for time series forecasting known as ARIMA, whose foundations are described in sub-Section 3.1.2.

known as time series prediction or forecasting (Morettin and Toloi, 2006). The prediction is the process of trying to anticipate the subsequent values based on the known data or past (history). With the information acquired from the prediction, a better planning can be achieved, or an undesirable state can be prevented from happening (such as the occurrence of damage in a civil structure). There are several models that can be used to perform a time series prediction, such as simple moving averages (SMA) and Single Exponential Smoothing (SES). The SMA method calculates the average of the last r observations. The term “moving average” is used because as soon as the next observation is available, the average of the observations is recalculated with the new data, i.e., including this observation in the set of observations and discarding the oldest observation. The SES is a weighted average that gives higher weight to more recent observations where this weight is determined by a parameter called smoothing constant (Faria et al., 2008). These methods are considered simpler methods of time series forecasting, and are not suitable for solving more complex problems (Morettin and Toloi, 2006). However, an advanced model for time series forecasting that has been gaining attention and has been widely discussed is the Autoregressive Moving Average (ARMA), which is also known as BoxJenkins model (Morettin and Toloi, 2006; Box et al., 1994). This classic model has some advantages compared with other methods, such as being less subjective and providing a probabilistic measurement of the forecasting error (Flores, 2009). As shown in the literature, the ARMA model is suitable for stationary time series. In the context of our work, after performing autocorrelation and partial autocorrelation tests of the data produced by the sensors, we concluded that the series turned out to be nonstationary. In this case, there are two options: applying a correction to the data or using a variation of the ARMA model called ARIMA (“Integrated ARMA”). The authors in (Morettin and Toloi, 2006) show that the ARIMA model is composed of one autoregressive (AR) term, one moving average term (MA) and a first order differentiation. Eq. (1) represents the resulting ARIMA model.

3.1.1. Adaptive Neuro Fuzzy Inference System (ANFIS) The Adaptive Neuro Fuzzy Inference System (ANFIS) is a kind of neural network that is functionally equivalent to Takagi–Sugeno fuzzy inference system (Loukas, 2001). Since it integrates both neural networks and fuzzy logic principles, it allows capturing the benefits of both in a single framework (Jang, 1993). Using a Takagi-Sugeno fuzzy inference system and a training data set, it is possible to create a fuzzy system able to automatically extract the intrinsic knowledge of the database (set of observations). Its inference system corresponds to a set of fuzzy IF–THEN rules that have learning capability to approximate nonlinear functions. The training process consists of adjusting the neural network parameters, in order to minimize the difference between the result provided by the system and the expected result. The training algorithm is said hybrid because it uses two different mathematical methods to optimize the adjustable parameters of the neural network (Jang, 1993). At the end of the training, the resulting neural network will represent exactly one full Takagi-Sugeno fuzzy system, i.e., the rules (with their weights) and fuzzy sets (with their positions in the universe of discourse) are already defined. However, some parameters to create an initial neural network are required such as (i) the number of fuzzy sets per input variable, (ii) the type of membership function that is used in the input variables and, finally, (iii) the type of membership function that will be used in the output variable. Different initial parameters can result in fuzzy systems that have different performance. Therefore, it is essential to perform a calibration, where the initial parameters are varied in order to generate the fuzzy system best suited to the problem. As previously mentioned, in this work we adopt ANFIS to generate a fuzzy system to perform the environmental interference filter, which will remove environmental factors from the natural frequency in order to increase the accuracy of the damage prediction process.

Xt=(Xt−1−Xt−2)*θ + Xt−1+, t−1*φ

(1)

Variables θ and φ are coefficients present in the equation of ARIMA to be adjusted in order to reflect the behavior of the time series. The estimation of coefficients (θ=0, 996866 and φ=−0, 938784 ) was performed using the Conditional Maximum Likelihood method. Also for this purpose, the value of the constant c must be chosen appropriately. The white noise, denoted by εt, is the difference between the predicted and the observed value at a given time t. Finally, the variable Xt is the value of the time series at time t. As mentioned, in this work we use the ARIMA model to predict the next natural frequency values of the monitored blade, in order to identify how much the blade may be compromised by the occurrence of damage.

3.1.2. Autoregressive integrated moving average (ARIMA) The ARIMA (Integrated ARMA) model is a statistical apparatus for time series prediction. Before presenting the model, it is necessary to make a brief conceptualization about time series. A time series may be defined as a set of observations of quantitative variables collected over time (Delurgio, 1998). Fig. 4 shows an example of a time series, plotting the average values of the environmental temperature measured in a Brazilian city between the years of 1976 and 1985. The analysis of such time series is based on the search of behavior patterns, and some of its properties offer subsidy to perform what is 127

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

Fig. 5. UML diagram component of Delphos architecture.

node. The Actuator Manager is responsible to take control actions on the environment. The Monitoring component coordinates data collection from the blade and from the environment around it by using the sensing units of the smart sensor nodes, as well as the information stored in the Monitoring Parameters database, such as the sampling frequency and the data monitoring period. Basically, this component is responsible for interacting with the sensing units to gather the collected temperature and vibration values and forwarding this information to the Data Handler and Environmental Interference Filter components. The Data Handler component processes raw vibration data in order to extract the natural frequencies of the blade by performing a Fast Fourier Transform (FFT) and a Curve Fitting procedure (as described in (Hackmann et al., 2008; Santos et al., 2014)). As a result, such procedures return the frequencies related to each vibration mode of the blade. As aforementioned, Delphos uses only the first natural frequency mode, which is enough to make accurate damage prediction. Besides the fact that first natural frequency mode is sufficient to know about blade state, the decision of using only this natural frequency mode comes from the need to reduce the processing load in each smart sensor node as much as possible, since they are devices with low processing capabilities. So, when relying on the first vibration mode only (the most reliable one), we avoid performing the whole prediction process for numerous modes of vibration, saving computational power at the costs of using a more simple prediction system. The Environmental Interference Filter component uses a fuzzy logic system to eliminate the influence of the environmental variables (currently only temperature) over the natural frequencies. As a result, this filter returns a “filtered natural frequency”. The fuzzy rules, as well as the input and output variables are stored in the Semantic and Fuzzy Rules database. The Abnormality Decision Maker component evaluates if the blade is in an abnormal state. Such an abnormal state is defined as a state in which the blade is not quite healthy, but it is not damaged yet. This

3.2. Delphos architecture In this work, we considered a WSAN composed of three types of nodes: (i) sensor nodes, (ii) sink nodes, and (iii) actuators. Sensor nodes are used to perform data collection and run the decision-making process of damage prediction. The sink is responsible for receiving alarms generated from the network and transmitting them to external networks/systems. The actuator has the function of receiving commands originating from the sensor network and make control actions appropriate to the predicted damage. The software components encompassing Delphos architecture are described in this section. The architecture (Fig. 5) comprises eight software components: Manager, Monitoring, Data Handler, Environmental Interference Filter, Abnormality Decision Maker, Damage Prediction, Actuator Manager, and Alarm. Additionally, five databases are used: Semantic and Fuzzy Rules, Monitoring Parameters, Coefficients, Time Series, and Control Actions. The databases are created and populated prior to the Delphos initialization. Such software components are deployed on physical devices as follows. The sensor nodes contain all the components, except the Actuator Manager and Alarm. The Actuator Manager component and the Control Actions base are deployed in the actuator node, while the Alarm component is located in the sink node. The Manager is the core component of the system and it is responsible for managing the operation of other components and coordinating control actions performed in the monitored blade. It gets the filtered natural frequency from the Monitoring component and sends this information to the Abnormality Decision Maker component in order to evaluate the abnormality state of the blade. If the blade is considered in a state of abnormality, the Damage Prediction component is activated by the Manager. The Damage Prediction component performs the damage prediction process by using the ARIMA model. Whenever damage is seen as imminent, the Manager contacts the Alarm and Actuator Manager components to perform two actions. The Alarm component is responsible to send an alarm signal to the sink

128

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

abnormality state. Whenever the blade is considered in an abnormal state, Delphos starts its damage prediction process by executing the second phase. In the third phase, the system executes a cooperative process to confirm the predictions made individually by each node of the WSAN, thus increasing the reliability of the damage prediction. In the fourth phase, the system takes the appropriated control actions to avoid the worsening of damage, and also sends alarms to external networks/systems to report about the performed prediction and control actions taken. It is important to note that both first and second phases are individually performed by each smart node of the WSAN, because the analysis performed by individual nodes (decision-abnormality and damage prediction) allows having different perspectives about the blade condition. In the third phase, the WSAN nodes exchange information about the individual predictions made by them. At last, the fourth phase is performed by a sink and an actuator node, where the first is responsible for sending the alarm and the second for performing control actions on the monitored environment.

component performs two tasks. Firstly, the natural frequency variation (Δω), which is defined as the difference between the initial (blade in healthy state) blade signature (ω0) and the blade signature at time “t” (ωt) is calculated (as described in (Santos et al., 2014)). Following, the found value Δω is evaluated to check whether it is within a normal range, which is defined as the normal natural frequencies variations when the blade is in its healthy state. Whenever Δω lies within its normal variation range, the blade is not yet considered in an abnormal state; thus the damage prediction process is not performed. Whenever the blade is considered in an abnormal state, the Damage Prediction component is triggered. It receives the values of Δω computed by the Abnormality Decision Maker and checks for the potential imminence of damage. The Damage Prediction uses the values of Δω to set up a history of the physical health of the blade and performs the forecasting of the future values of Δω based on this history, using the ARIMA model. A “Damage Imminence” is identified when the predicted values exceed a given threshold T. This threshold must be determined for each specific civil structure (Santos et al., 2014). As previously mentioned, an ARIMA model is used to predict the future values of the natural frequencies of the blade (Salfner et al., 2010). Therefore, two parameters of the time series need to be defined: Prediction Horizon (PH) and Error Margin (EM). PH is defined as the number of points to be predicted from the last known point of the time series. In the context of this work, the size of the PH is associated to the number of future data collections. Delphos indicates only the data collection in which damage will occur (i.e., the future data collection in which the values of natural frequency variations will be outside the threshold T). EM is used to determine the “data collection range”, the range in which the damage will occur. In other words, if the occurrence of damage was predicted in the hundredth data collection number, and the EM parameter is equal to 3, Delphos will indicate that damage will happen within the data collection range between 97 and 103. Finally, the Actuator Manager component is in charge of receiving the information about the damage prediction, identifying the most appropriated control action (by accessing the Control Actions database) and sending commands to actuator devices. The Alarm component informs external network, control centers or other systems about the outcome of the damage prediction.

3.3.1. First phase In the first phase, each smart sensor node performs the tasks of (i) monitoring the vibration and temperature values of the wind turbine, (ii) extracting the natural frequencies of the blade, (iii) submitting the natural frequencies to the environmental interference filter; (iv) computing Δω and, (v) checking whether the value of Δω is within the normality range. These tasks are performed by the Monitoring, Data Handler, Environmental Interference Filter and Abnormality Decision Maker components. In this phase, as shown in the UML sequence diagram of Fig. 6, the Manager component requests the Monitoring component to return the filtered natural frequencies. Then, the Monitoring performs the collection of temperature and vibration, and calls the Data Handler to extract natural frequencies. Hereafter, the newly extracted natural frequencies are returned to the Monitoring, which requests the Environmental Interference Filter component to run the filter to reduce the influence of temperature on the frequencies. Upon receiving this request, the Environmental Interference Filter forwards the temperature and the natural frequencies to the fuzzy system that effectively performs the filtering. The filtered natural frequency is sent back to the Monitoring that provides this filtered natural frequency to the Manager in response to the previously made request. Then, the Manager asks the Abnormality Decision Maker to determine whether the blade is in a state of abnormality or not. If the blade is in abnormality state, an indication of abnormality is properly configured and sent to the Manager, so that it becomes aware of this state. So, the smart sensor nodes that have identified an abnormality state start the execution of the second phase.

3.3. Delphos operation Delphos operates through successive cycles of monitoring and decision, called execution cycles. Each cycle comprises four major phases, where each phase is performed for a fixed period of time. To keep the processing of all the sensor nodes in the same execution phase, internal timers are implemented which determine the period of time each phase may take for completing its execution. Each period of the monitoring cycle should start at the same time on all sensors that are monitoring a same structure. The data collected by all sensors must be synchronized so that there is meaning in the comparison of data gathered and decisions taken by different nodes. There are currently several proposals for time synchronization protocols in WSN that could be adopted in Delphos. Our algorithm, by design, is agnostic to any particular protocol, i.e. the synchronization process can be performed by any synchronization protocol available in the literature that meets the requirement of keeping, for long periods of time, the desired degree of accuracy (in terms of temporal deviation among the clocks of the nodes) in the synchronization, where such degree is defined by an expert in the SHM application scenario. The proposals described in (Huang et al., 2014, 2012) are examples of time synchronization protocols for WSAN that meet such requirement. It is important to mention that, although the synchronization process in Delphos is performed for the first time in the setup phase, it must be reviewed (maintained and adjusted) during the monitoring cycle phase. In the first phase, the system verifies whether the blade is in

3.3.2. Second phase In the second phase, each smart sensor node that has identified a abnormality state performs: (i) inputting values of Δω calculated in the previous phase (abnormality decision) to a time series, (ii) calculating future values for Δω and (iii) checking if any expected value exceeds the threshold T (described in Section 3.2). If any damage is predicted, the system (in the fourth phase) has the task of contacting the actuators to perform actions that prevent the worsening of damage, for example by decreasing the blade rotation or even by emergently turning off the equipment. The Manager and Damage Prediction components are used in this phase. As shown in the sequence diagram of Fig. 7, the Manager asks the Damage Prediction to perform the prediction for the received value of Δω. Upon receiving this request, the Damage Prediction performs the prediction of damage for the monitored turbine. If damage is predicted, the Manager is informed about when the damage will occur. 129

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

Fig. 6. UML sequence diagram representing the first phase.

In this third phase, nodes exchange information between themselves, indicating whether or not any damage prediction was made. The prediction is said to be valid whenever more than one node reached to the conclusion that there is an imminent occurrence of damage. Given a set of forecasts made by several nodes, Delphos sends to the sink and actuator nodes the lowest damage prediction, i.e., that prediction in which the reaction time (difference between the current time and the time when the damage will occur) is the lowest. To realize this collaboration mechanism among the nodes (and the subsequent forwarding of the final result to the actuator and sink nodes), we adopt a data fusion strategy incorporated to the message

3.3.3. Third phase At this point of processing, Delphos has already ran the first and second phases, where each node of the network either found or not that there is imminent harm to the blade being monitored. Thereafter, Delphos performs a third phase, where nodes cooperate in order to increase the reliability of the damage prediction and to provide a single final result to the sink and actuator nodes. The third phase is performed by all network nodes, even if the node has not concluded that the blade is in an abnormal state or that damage will occur. This collaboration mechanism allows an increased reliability of the damage prediction process.

Fig. 7. UML sequence diagram representing the second phase.

130

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

Fig. 8. UML sequence diagram representing the fourth phase.

radio power signal to reach the next after neighbor. Another point worth mentioning is that the nodes keep waiting during a specified time (timeout) for a message to be received, where each node has a different time of expiration. If a timeout occurs, the node starts on its own all the collaboration process, temporarily assuming the designation of start_node. The nodes possess different expiration times, since in case of timeout only one node will temporarily assume the role of start_node a time, while the others must still be waiting for the reception of the message. If the nodes were assigned the same timeout value, every node would start at the same time the process as a start_node in case of not receiving any message.

routing algorithm used by the WSAN. In short, the nodes forward their prediction results to their direct neighbors (one hop) and these nodes assess whether the received forecast is lower than the forecast computed locally. At the end of this process, the lowest forecast is delivered to the sink and actuator nodes. Therefore, this collaborative process is executed by all sensor nodes, where each one communicates only with other two nodes, called before and after neighbors. The nodes are deployed in a flat and linear topology, because this topology is imposed by the physical shape of the wind turbine blades, which is basically a rectilinear structure. In this collaborative process, two sensor nodes play specific roles: start_node and end_node, where the former has the responsibility to start the process of cooperation and the later ends the process upon verifying the predictions made by the other network nodes. The nodes that perform the roles of start_node and end_node are manually selected during the system deployment process in the environment to be monitored. The remainder network nodes are called simply as nodes. In order to start the collaboration process, after the start_node receives a message from the sink node and performs the first and second phases of Delphos Operation, it sends a message to its after_neighbor which in turn forwards the message to its own after_neighbor until the message reaches the end_node. Thus, the message is routed through a static predefined route that begins and ends in the start_node and end_node, respectively. We adopt this static route because this fixed topology (flat and linear) is appropriate to the size and shape of the blade, since this structure has a shape that remains straight throughout its length. Another point is that this static route allows energy savings, since each node can adjust its signal strength to communicate only with its before and after neighbors. The messages sent through the statically configured route are of type DAMAGE_PREDICTION and always have their reception acknowledged by the destination node, i.e., the after neighbor. This type of message comprises two fields, where the first (PredictionsNumber) is used to account the amount of nodes which made the prediction and the second (indication) stores the indication of when the damage will occur (future monitoring cycle). All nodes (including start_node and end_node) verify if a damage prediction was made and, if so, increment the field PredictionsNumber and replace the damage indication value (in the field indication) if the indication made by the node is lower than the indication currently in the message. However, the end_node performs a particular task that concerns to verify if the field PredictionsNumber is greater than the number of predictions necessary to consider the prediction valid. This number is configured during the WSAN deployment process. If the prediction is considered valid, it is sent by the end_node to its after neighbor, which necessarily assumes the sink and actuator roles. During the collaborative process, there may be nodes that are unavailable and as a consequence affect the routing of messages through the static route. In this case, the node that sent a message and has not received confirmation of its after neighbor will increase its

3.3.4. Fourth phase After completion of the third phase, Delphos starts the fourth phase, but only in the cases where the system has sent a final result to the sink and actuator nodes. At this stage, the sink node issues an alarm and the actuator carries out control actions to prevent the occurrence of the predicted damage. Fig. 8 shows the UML sequence diagram with the actions performed in this phase. The Manager component asks the Actuator to execute proper actions (such as reducing the rotation of blade or stopping the operation of wind turbine) to prevent the occurrence of damage or its aggravation. Therefore, the Actuator needs to identify the control action that is most appropriate to the predicted damage and effectively triggers the actuators (physical devices) so that such devices perform the required action on the environment. By completing this cycle, the Actuator informs the Manager which action was executed. Then, the Manager calls the Alarm component, that issues the alarm and reports back to the Manager whether the alarm was successfully sent or not. 4. System implementation This section discusses the main issues related to the implementation of Delphos. Such issues regards (i) the pre-deployment phases (ARIMA estimation and ANFIS training), and (ii) the calibration of Delphos parameters (PH and EM). The implementation of Delphos required real vibration data, from which the ARIMA estimation, ANFIS training and Delphos calibration were performed. Therefore, the issues related to the acquisition of the vibration data used are also briefly discussed. 4.1. Vibration data collection In this section we describe the methods used for collecting the vibration data from the blade of a wind turbine. We also describe the existing limitations of the environment available for the vibration data collection, and how each one was overcome. The vibration data was organized within a database and expanded through the use of linear regression models, for further use in the procedures described in the remainder of this section and in Section 5. 131

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

natural frequency in the future, during its lifetime. In our work, the NFETS always present a linear ascending evolution profile (linear curve shape) to represent real situations in which the wind turbine blades suffer wear over the time. Therefore, this linear profile does not take into account the situations in which the blades present a sudden occurrence of damage (i.e. in extreme circumstances, such as during storms or wind speeds higher than 70 km/h). During the lifetime of the blade, the NFETS may present situations where it assumes a damaged or undamaged state. The damaged state is reached, in a given moment in time, whenever the NFETS reaches a value above the threshold T (described in Section 3.2). It is important to mention that this threshold T can be statistically determined after making a series of experimental samples. The purpose of adopting a tolerance value (denoted by T) is to prevent small random disturbances, which do not imply the occurrence of abnormal conditions, from being considered by the monitoring procedure as such. Such small random disturbances are attributed to imprecision in calculations and truncations in the determination of the modal frequencies, generated by the each sensor. So, we will assume the value of two standard deviations for T, and this value is the same for every node and every modal frequency. This is enough to assume that in a real situation less than 5% of frequency values will exceed the threshold in an undamaged situation (false-positive), assuming that the extracted values of frequencies follow a normal distribution and the average is well estimated using 7 frequency samples. Then, the threshold T assumes the value of 2% of the average value estimated for every mode of vibration and every sensor node. The evolution of the natural frequency values until the damaged state is reached occurs gradually, through the increasing of the severity of the damage in time, and following the trend of the ascending linear curve shape. The NFETS which presents a single damaged state is called Damaged NFETS, or briefly Damaged Evolution (DE). The NFETS which does not present a damaged state, i.e. the blade is healthy all the time, is called Undamaged NFETS, or briefly Undamaged Evolution (UE). Fig. 9 depicts an example of a Damaged and an Undamaged Evolution. In the procedures described in the remainder of Section 4 and in Section 5, the NFETS are the data further used to simulate the data collected by the sensors, and provided as input to Delphos prediction model, in different damaged and undamaged situations. A second limitation related to the experimental environment was that we did not have a controlled environment, such as a climatic chamber, to systematically vary the temperature in order to identify its influence over the natural frequencies of the blade. Therefore, to overcome such a limitation, we simulated the influence of the temperature over the collected natural frequency data for generating NFETS. Since we could not vary the temperature during the data collection, we had to collect it at the environment temperature, which was around 26 °C during the experiment execution. We call F the values of natural frequencies collected at the environment temperature, i.e. without the influence of a varying temperature. By using a linear regression (Eq. (2)) (Kurian et al., 2006; Griffith et al., 2006),

We collected real data from the natural frequencies of the blade of a wind turbine model Notus 112/138 (Enersud Energia Limpa, 2013). We used a pulse hammer (Sohn et al., 2004) to introduce excitation enough to the system is capable of evidencing the vibration modes. A smart sensor model Imote2, and a sensor board ITS410 equipped with an accelerometer, both from MEMSIC (MEMSIC, 2013), were used to acquire the vibration (acceleration) data from a single point directly on the surface of the blade. In addition to the sensor node Imote2, IPR2410 and IBB2410CA boards were used. The IPR2410 board contains the Marvell PXA 271 processor, 256 Kb of SRAM, 32MB of RAM and 32MB of SDRAM integrated with a radio 802.15.4 (CC2420). The IBB2410CA board provides power required for the sensor with three AAA batteries. The sensor Imote2 was used in the experiment of accelerometer calibration and experiment to collect real data of the blade. In these experiments, the standard communication protocol Imote2. Net was used, ie the standard Zigbee 802.15.4. The Imote2 was chosen to collect real data of the blade since it was the only sensor platform equipped with accelerometers available in our laboratory. The blade was installed in a controlled environment at the Laboratory of Networking and Multimedia (LabNet) of the Federal University of Rio de Janeiro (UFRJ). Each collected vibration data was converted from time domain to frequency domain by applying the FFT, as mentioned in Section 3.2. For every damage degree (amount of weight), 30 repetitions of data collection were performed (each collection comprising 1024 acceleration samples in time at a sampling rate of 100 Hz), to provide a reasonable confidence interval of 95% for every natural frequency value. To understand the meaning of damage degree, consider that an increase in such value means an increase in the severity of the damage inserted in the blade, and consequently, the natural frequencies variation increase as the damage degree increases. A limitation related to the experimental environment was the impossibility of causing real damage (for simulating every different values of damage degree) in the only turbine blade available to be assessed. To overcome this limitation, it was necessary to adopt a method of artificial damage insertion and to simulate the influence of the temperature over the collected natural frequency data. The adopted method of artificial damage insertion was based on the variation of the mass of the blade, which is similar to the approach used by Clayton et al. Clayton et al. (2006). According to this method, it is possible to simulate the insertion of damage (variation of the natural frequencies) of a blade by directly adding mass over it, without the need to actually damage (irreversibly) the blade. Thus, during our vibration data collection, different amounts of mass were gradually added to the blade, performing a data collection for each amount of mass added in order to assess the natural frequency variation. In each experiment, the blade presented a damage degree linked respectively to one of the following addition of weights: 0, 100, 250, 500, 750, 1000, 1250, 1750 and 2000 g. Based on the natural frequency data obtained, as previously described, we estimated a function relating the damage degree (amount of weight) with the frequency. We can estimate such a function by using linear, quadratic or cubic regression techniques. The functions estimated from these regressions are compared on the adjusted determination coefficient (R2 - adj) in order to identify the more successful function when trying to adjust the displayed points (obtained data) (Seber, 1977). Values close to 100% mean that there is a strong correlation between the estimated function and the points presented. Using such equation, we created 100 datasets of natural frequency versus time. Each dataset had the weight and the respective frequency calculated through this estimated equation. The weights were randomly obtained through a normal distribution, whose values ranged between 0 and 2000 g (the range of the experiments). The 100 datasets generated as result of this procedure are called Natural Frequency Evolution Time Series (NFETS), which are, therefore, a set of natural frequency data (free from the influence of temperature) versus time. This set of frequencies was used to represent the evolution of the blade

Natural Frequency Variation (Hz)

4.50 4.00 3.50 3.00 2.50 2.00 1.50 1.00 0.50 0.00 1

31

61

91 Time (days)

121

Fig. 9. Damaged and undamaged evolutions.

132

151

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

For training the ANFIS, it is necessary to build a database where each entry must contain: a given temperature value, natural frequency influenced by such temperature, and filtered natural frequency (without the influence of the temperature). ANFIS considers the original natural frequencies and temperature values as inputs. As output, ANFIS considers the filtered natural frequencies. So, using this information, ANFIS creates a fuzzy system capable of providing a natural frequency given a pair of filtered natural frequency and temperature. Before creating the fuzzy system, it is necessary to perform a variation of ANFIS parameters. In this process, the ANFIS parameters are varied to determine which fuzzy system performs better. In other words, for configuring the ANFIS parameters we need to determine which fuzzy system presents the smallest training error (training error around 122×10-6). The following parameters were varied: (i) the number of fuzzy sets per input variable (labels) and (ii) the types of member functions applied over the input and output variables. Table 1 shows the parameter configuration that presented the five smallest errors. So, based on this result, we choose a fuzzy system with the following parameters: (i) three fuzzy sets per input variable temperature and frequency variables, (ii) the Pi-Shaped function as the member function for each input variable and (iii) one linear function as membership function for the output variable.

subsequently to the data collection, we extrapolate F′ values, which are the values of natural frequency influenced by each temperature Tp. The constants 26 and 0078 are respectively related to (i) the reference environment temperature previously mentioned and (ii) the frequency variation assumed for every 10 °C of temperature variation, as described in (Griffith et al., 2006). Therefore, removing the influence of temperature from a NFETS means converting its frequency values to the expected values at the standard temperature of 26 °C.

⎛ 0, 0078 ⎞ F′=(Tp −26)* ⎜ − ⎟+ ⎝ 10 ⎠

F

(2)

By using Eq. (2) it was possible to add temperature values for each of the 100 NFETS. The range of temperature values was chosen based on the sensitivity of the MTS400 sensor board. This sensor board is used in MICAz platform to sense data from humidity, temperature, light and acceleration. According to (MEMSIC, 2013), the sensor board is able to observe variations in temperature within the range of −10– 60 °C. Once defined such range, the temperature values were added for each NFET according to a random sampling, which followed a normal distribution. This range of temperature (−10–60 °C) is suitable to make our experiments since wind turbines are located in sites with high variability of temperature such as deserts and offshore. Moreover, the operation of wind turbines itself, which is driven by wind velocity, can drastically decrease the temperature of the blade. Furthermore, the sampling of temperature can be made by Delphos within a time interval (12, 24 h or more) during which there may be significant difference of temperature values. From the 100 generated NFETS, 40 were used for Delphos calibration experiments C1 and C2 (described in Section 4.3), and the remaining 60 were used for the experiments described in Section 5. This number of NFETS is enough to cover a wind turbine lifetime, which is estimated at around 20 years (considering the case of the model Northern Power® 100 (Northern, 2013)). All the NFETS values were stored in a database called Experimental Database. It is important to mention that the Experimental Database is not part of Delphos logical architecture (depicted in Fig. 5), in the sense that it does not implement any logical functionally that is key for the system operation. Instead, it is a support database used only for the experimental settings of the performed system calibration. A third limitation related to the experimental environment was the impossibility of having faulty nodes in the experiments conducted with real sensor nodes. In this work, we considered faulty nodes as those nodes which have a malfunctioning during the system operation due to imprecision internal flaws in the accelerometer and temperature sensor, for instance. Since we could not have this behavior in practice (with real sensor nodes), to overcome such a limitation, we simulated the behavior of failure. In our simulation, each faulty node receives a Damage Evolution randomly chosen from the Experiments Base, where the Damage Evolution is necessarily different from the Damage Evolution that is being analyzed by the nodes with correct operation. In these experiments, the number of faulty nodes ranged between 0 and 4, one by one.

4.3. Delphos calibration This section presents Delphos calibration experiments (denoted as ‘C1′ and ‘C2′), conducted to identify the best values for PH and EM parameters (as described in Section 3.2). Such values should be chosen so as to minimize the number of false positives (FP) and false negatives (FN) returned by the system. In calibration experiment ‘C1′, the environmental interference filter is enabled before the process of damage prediction. In calibration experiment ‘C2′, the filter is disabled. Therefore, experiment ‘C1′ takes into account the influence of temperature over the natural frequency, while experiment ‘C2′ does not take such influence into account. Such approach was adopted so that the evaluation of Delphos according to the effectiveness of the environmental interference filter can be fair, assuring enough scientific rigor. In other words, the system was calibrated for achieving its best performance in both cases, with and without the environmental filter, so that the perceived difference in the system performance could be exclusively attributed to the presence of the environmental interference filter. The scenarios, metrics, environment setting, and results of both performed experiments are described as follows. 4.3.1. Calibration scenario and environment setting In calibration experiments ‘C1′ and ‘C2′ we used 40 (out of the 100) NFETS in the Experimental Database (as described in Section 4.1), and considered a neutral mix (Damage Evolution=50%; Undamaged Evolution=50%). Therefore, each simulation round was performed Table 1 Results of ANFIS parameter variation.

4.2. Delphos pre-deployment phases Before deploying Delphos in a given scenario (a specific wind turbine), it is necessary to perform two pre-deployment phases of the system operation, which are performed outside the WSAN. The first phase is responsible for creating, through the ANFIS, the fuzzy logic system which performs the environmental interference filter. In the second phase, the ARIMA estimation procedure is performed. In this work, the ANFIS training process was performed with Matlab (Kurian et al., 2006), while the ARIMA estimation was performed with the GNU Regression, Econometrics and Time-series Library (Gretl) (Shukla and Jharkharia, 2011). 133

Temp. labels

Freq. labels

Input membership function

Output membership function

Error

3 3 3

3 3 3

Linear Linear Linear

0.000001223945920 0.000001383552562 0.000001423556266

3

3

Linear

0.000001430142369

5

3

Pi-Shaped Trapezoidal Product of two sigmoid Difference between two sigmoidal Pi-Shaped

Linear

0.000001495273391

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

(Alberola and Pesch, 2008) to perform the simulated experiments described in Section 5. Delphos was implemented by creating eight new TinyOS components as described in Table 2. The ordered delivery of messages, which is a prerequisite of the collaborative process described in Section 3.3.3, is guaranteed through the protocol of message acknowledgments, which is natively available in TinyOS through component AMSenderC and interface Packet Acknowledgements. Delphos implementation can be found at http://ubicomp.nce.ufrj.br/ Delphos/.

with 20 DE and 20 UE. In both calibration experiments, only one sensor node was used, since the goal was assessing only the operation of ARIMA model in terms of hit prediction according to parameters PH and EM. The evaluation of Delphos according to the network topology, energy and other aspects are described in Section 5. In this work, we considered a WSAN composed of MICAz (MEMSIC, 2013) sensor nodes, whose radio supports 2.40–2.48 GHz band and 250 kbps data rate. Each MICAz node is endowed with 4 kB of RAM, 128 kB of flash memory for program storage and 512 kB for data storage, and is powered by two AA batteries, which provide up to 16 kJ of energy. Delphos was built by using TinyOS (Levis and Gay, 2009) development environment version 2.1.1, and nesC programming language (Levis and Gay, 2009). For dealing with communication at the physical layer, we adopted the standard 802.15.4 protocol implementation provided by TinyOS. For dealing with the transmission of pointto-point messages at link-level, we adopted the Active Message protocol implementation, also provided as an embedded feature of TinyOS. The maximum payload size allowed for the messages exchanged, limited by TinyOS 2.1, is 28 bytes. The implementation of our algorithm itself consists of a single program running inside each mote. Our algorithm runs over the BMAC protocol (Buonadonna et al., 2001) in the MAC layer. Using pure physical layer information, the BMAC is capable of detecting neighboring wake-up transmissions on the channel. Consequently, BMAC allows nodes to get into deep-sleep state until wake-up transmissions are detected. Our algorithm and the BMAC work independently. When a node in our algorithm becomes idle, it does not make transmissions anymore. Therefore, it is a matter of time until the BMAC protocol returns nearby nodes to deep-sleep mode. Besides, since the monitoring cycle starts when the sink node disseminates a message for the whole network, nodes in deep sleep will be woken up by the sink transmission during one of the periodic receive checks of BMAC. Therefore, the use of BMAC contributes for saving nodes’ energy during the idle time of our algorithm. The MICAz was chosen for two main reasons: (i) the existence of a reasonable amount of MICAz motes in our laboratory, acquired from research grants of our group, and (ii) the existence of simulators developed by third parties for this type of hardware platform. Therefore, choosing the MICAz platform allowed us to create a WSN with a fair number of nodes and also to use the same hardware platform in both simulated and real environments. Finally, the same implementation of our proposed algorithm was used in both the simulated experiments and the experiments using real motes, i.e. the software logic which incorporates the proposed algorithm was also installed in the actual motes. The data collected by the real Imote2 sensors were incorporated into the code of the MICAz sensors, thus eliminating the need of accelerometers in MICAz sensors. We used the WSN simulators TOSSIM (Levis and Gay, 2009) to simulate the calibration of the system parameters, and AVRORA

4.3.2. Calibration metrics The calibration experiments ‘C1′ and ‘C2′ used the following metrics: False Positives (FP), False Negatives (FN), True Positives (TP) and True Negatives (TN). TP metric counts the number of situations in which the system predicts a damage occurrence that actually occurs in the same moment indicated by the system. In other words, this metric indicates the desired behavior of the system, i.e. when the system was effectively able to predict the time at which damage will occur. The TN metric counts the number of situations where the system does not predict damage in a certain moment, and it does not occur. The cases recorded by the TN metric are desirable because Delphos did not perform a hasty control action, since in fact no harm would occur. FP counts the number of situations where the system predicts damage and it does not occur. Therefore, this metric allows knowing the cases when Delphos issued an unnecessary alarm, i.e., the system erroneously reported that there will be a damage at a given time, and this damage, in fact, would not occur. This behavior is not desired because the system would generate an alert or perform a countermeasure in a time when there would be no need for it. At the time of an unfounded control action, energy production could be negatively impacted by the reduction of the speed of the turbine, for example. The FN metric counts the cases in which the system does not predict damage, but it occurs. Undoubtedly, this is the worst situation that can happen, since the damage would occur without having been properly predicted by Delphos. In other words, we can say that in these cases Delphos would fail in its main function, which is to inform in advance when a damage will occur. The quantity (TP+TN) is referred in this work as the amount of trues and is used to denote the number of situations in which Delphos makes correct prediction, i.e. not predicting damage when it should not be predicted, and predicting damage when it would actually exist. 4.3.3. Results of calibration experiments The goal of the calibration experiments C1 and C2 is to identify the most suitable values for the system parameters, Prediction Horizon (PH) and Error Margin (EM) that attends both cases (with and without the influence of the temperature), in order to minimize the values of FP and FN. By providing experiments C1 and C2 in our paper, we intended to show that we can choose the same values of EM and PH in C1 and C2

Table 2 Description of Delphos components. Component name

Logical component

Description

AcoesControle.nc

Actuator manager and alarm

Comunicacao.nc

Abnormality decision maker

DecisaoAnormalidade.nc FiltroInterfAmbientais.nc

ModeloARMA.nc TratamentoDados.nc

Abnormality decision maker Environmental interference filter Environmental interference filter Damage prediction Data handler and monitoring

This component performs control actions to avoid blade breakdown and is also responsible to send alarms to sink node. This component executes cooperative process which nodes exchange their predictions with each other to increase the reliability of damage prediction This component verifies whether the blade is in an abnormality state. This component performs reduction of environmental interference using a Sugeno fuzzy logic system.

PredicaoFalhas.nc

Damage prediction

FuncoesMembro.nc

This component provides member functions used by the logical fuzzy inference system (Sugeno). This component implements ARMA model that is used to make the damage prediction. This component performs data processing in collected vibration in order to identify the blade natural frequencies. So, it executes processes such as FFT and peak extraction. This component uses ARMA model to make damage prediction based in blade natural frequencies variation.

134

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

in order to grants a higher accuracy in the performed predictions and to achieve a higher reaction time (time between the moment in which the prediction is performed and the moment in which the event is predicted to occur). Experiment C1 used an ARIMA model obtained from a time series without the influence of the temperature, while experiment C2 used an ARIMA model obtained from a time series with temperature influence. The outcome of the experiments allows choosing the smallest value of EM which grants a higher accuracy in the performed predictions, since the smaller the value of EM becomes, the less imprecise will be the prediction. Besides adopting the smallest EM value, it is also needed to choose the greatest value of PH (relative to the EM value chosen) to achieve a higher reaction time, since the higher the PH becomes, the higher will be the time between the moment in which the prediction is performed and the moment in which the event is predicted to occur. The experiments comprise three scenarios, where the value of EM is fixed, respectively, at 2, 3 and 4, while the values of PH were varied from 3 to 10, one by one. The experiments showed that Delphos does not performs well for PH values higher than 10 (TP lower than 18%), as ARIMA is suitable for short-term forecasts (Hyndman and Athanasopoulos, 2014). EM values higher than 4 are not feasible because they result in very inaccurate predictions. The experiments also showed that Delphos does not work properly for ME values smaller than 2 (TP below 40%), because ARIMA provides a future trend and not exact values. Results where EM values were equal to PH values were not considered, as in these cases the reaction time would be zero. Overall, it was observed that the system achieves excellent results of TN for almost all PH and EM values, while the amount of TP decays as the PH increases. This behavior is related to the ARIMA model and to the evolution without the presence of damage shown in Fig. 9. The ARIMA model makes the prediction of future values according to the trend of the time series being modeled. The tendency of time series without the presence of damage (Fig. 9) is to remain within the 95% confidence interval for the healthy state of the structure. That is, the values of the frequency variation do not approach the defined threshold. Thus, the system is prone to make predictions which do not approach the threshold and, therefore, are less likely to indicate the existence of damage when the structure is in a healthy state. Another point to be made regarding the experiment results is that the system reaches a higher amount of true (TP+TN) as the EM values increase. However, the use of higher values of EM implies a less accurate response, i.e., a large margin of error may provide an inaccurate prediction, since it will not be possible to identify the actual time interval in which the damage may occur. Otherwise, it appears that the system achieves higher values of true when PH is less than 6, confirming what was said about the short-term efficiency of the ARIMA model. Therefore, we need to find the PH and EM values so that Delphos performs well in terms of prediction hit and at the same time, does not present a high inaccuracy in its prediction response. For doing so, we utilized the metrics of specificity, sensitivity, and accuracy considering error margins of 2 and 3. Results for the margin of error equal to 3 are shown in Fig. 10. In both cases, the resulting ROC curves show that the system has a constant value of specificity for all PH values. However, the sensitivity and accuracy decreases as the PH values increase. This behavior is expected since a high PH value makes the prediction more inaccurate. From the analysis of the results, we choose adopting the lowest value of EM in order to ensure more accurate results and a higher PH value (related to the chosen EM value) to achieve higher reaction time. Thus, we used PH=6 and EM=2. In other words, Delphos can predict whether the damage will occur in up to 6 future data collections and the damage can occur in the collections of 4–8 where each future data collection represents one day.

Fig. 10. ROC curves for error margin=3.

5. Evaluation of the damage prediction system This section describes the experiments conducted with Delphos with the purpose of evaluating the following properties: (i) the system accuracy to predict damage with and without the environmental interference in terms of correct damage prediction; (i) the system accuracy to predict damage with and without the environmental interference in terms of correct damage prediction considering faulty sensors (iii) the gain, in terms of response time, of a decentralized decision making process (used by Delphos), in comparison to a centralized approach; and (iv) the system requirements in terms of memory (RAM and ROM) and energy consumption. All these properties were evaluated in a simulated environment using Avrora Simulator. Moreover, experiments in real sensor nodes were conducted to validate the simulated experiments performed to assess the system accuracy (item (i) above). Avrora was used since it was the only simulator available which provided an energy model for the MICAz platform, the adopted WSAN platform. 5.1. Environment setup and adopted metrics We considered three different scenarios in terms of the number of sensors, which are composed of six, twelve, and eighteen smart sensor nodes, respectively. One special node performing the roles of sink and actuator was added to each scenario. These three different scenarios allowed evaluating the impact (positive or negative) of the number of nodes over the provided accuracy (of the prediction process), network lifetime and response time. In all scenarios, we considered a WSAN composed of MICAz sensor nodes. The MICAz nodes were arranged in the surface with 15 m long. In this work, we used a plain, linear, static topology for the WSAN, which is a realistic setting for sensor networks monitoring this type of structure. The distance between each node was 2.5, 1.25, and 1 m for the scenarios with six, twelve, and eighteen smart sensors nodes, respectively. The sensor nodes were uniquely identified by a NODE_ID number. The NODE_ID 0 was assigned to the sink/actuator node and the NODE_ID 1 was assigned to the end_node (the sensor node nearest of the sink/actuator). All the remaining nodes were identified with the value of NODE_ID incremented by one until the start_node (the sensor node located farthest from the sink/actuator node) which was assigned with the highest NODE_ID in each scenario. Thus, considering a node with NODE_ID = N, its before_neighbor is identified by NODE_ID=(N – 1) and its after_neighbor is identified by NODE_ID=(N+1). Since in our experiments, the acceleration data were generated as described in Section 4.1, implementing data synchronization is not necessary. A complete simulation was divided in rounds. In each round a different arrangement of NFETS was used and the system behavior was evaluated in terms of correct predictions. The NFETS arrangement assumed three different configurations, namely: neutral (DE=50%; UE=50%), damaged (DE=75%; UE=25%) and undamaged (DE=25%; 135

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

UE=75%). The following metrics were evaluated in our experiments: TP, TN, FP and FN (as described in Section 4.3.2); the system response time (denoting the time elapsed between the beginning of an execution cycle in which a damage must be detected and the reception of a message by the actuator informing that a control action should be performed); the memory requirement (RAM and ROM); and the energy consumption. Based on the value of energy consumption measured in each simulation round and on the initial amount of energy in each network node, we estimated the network lifetime as the time elapsed from the beginning of the system execution until the moment in which the WSAN is not able to achieve its main goal. In the context of this work, the main goal of the WASN is to perform the damage prediction, which must be done by at least 2 nodes, according to the collaboration process described in Section 3.3.3.

Table 4 Results of experiment ‘E2′ (not considering temperature influence). Mix of NFETS

Smart sensor nodes

Undamaged

Neutral

Damaged

Metrics

6

12

18

TP TN FN FP TP TN FN FP TP TN FN FP

5.0% 75.0% 0% 20% 7.5% 50.0% 0% 42.5% 12.5% 25.0% 0% 62.5%

5.0% 75.0% 0% 20% 7.5% 50.0% 0% 42.5% 10.0% 22.5% 2.5% 65.0%

5.0% 75.0% 0% 20% 7.5% 50.0% 0% 42.5% 10.0% 22.5% 2.5% 65.0%

5.2. Evaluating the system accuracy 5.3. Evaluating the system accuracy considering faulty sensors We begin our analysis by evaluating the system accuracy to predict damage, in terms of correct damage prediction, with and without considering the environmental interference. Two experiments were conducted for this purpose. The experiment named ‘E1′ was performed considering the influence of temperature on the variation of the natural frequencies of the blade. Experiment ‘E2′ was performed without considering such influence. In both experiments, we adopted the same metrics presented in Section 4.3.1, and the same parameters selected in the calibration experiments described in Section 4.3.3 (PH=6 and EM=2). Table 3 and Table 4 show respectively the results (in terms of FP, FP, TP, and TN) obtained in Experiments ‘E1′ and ‘E2′ when the number of sensors is set to 6, 12, and 18. From Table 3, when the number of sensors is equal to 6, we can observe that the system achieved a total number of true occurrences for the Neutral, Undamaged and Damage configurations equal to 100%, 100% and 97.5%, respectively. From Table 4, when the number of sensors is equal to 6, we can observe that the system achieved a total number of true occurrences for the Neutral, Undamaged and Damage configurations equal to 57.5%, 80% and 37.5%, respectively. So, for each configuration of the mix of NFETS (neutral, damaged and undamaged), and for each number of sensor nodes (6, 12 and 18), the results presented in Table 3 and Table 4 showed that the temperature can significantly influence the process of predicting damages and that Delphos is able to filter such interference as expected. Therefore, the temperature negatively influences the damage prediction, since the series of frequency under its influence is more difficult to be modeled by the ARIMA model. On the other hand, the series consisting of frequency values without the influence of temperature define a smoother curve (with less noise), thus allowing a prediction with higher accuracy.

In this section, we evaluate the system accuracy to predict damage, in terms of correct damage prediction, with and without the environmental interference, and considering the presence of faulty sensors in the network. Two experiments were conducted for this purpose. The experiment named ‘E3′ was performed considering the influence of temperature on the variation of the natural frequencies of the blade. Experiment ‘E4′ was performed without considering such influence. In both experiments, we adopted the same metrics presented in Section 4.3.2, and the same parameters selected in the calibration experiments described in Section 4.3.3 (PH=6 and EM=2). Table 5 e 6 show respectively the results (in terms of FP, FP, TP, and TN) obtained in Experiments ‘E3′ and ‘E4′ for each configuration of the mix of NFETS (neutral, damaged and undamaged), and for each number of faulty nodes (ranging from 0 to 4, one by one), when the number of sensors in the network is set to 6. From Table 5, we can observe that the system performs well with respect to the number of true occurrences (sum of TP and TN) for all the three configurations of NFETS when we varied the number of faulty nodes in the WSAN. For Neutral and Undamaged configurations of NFETS, Delphos achieved a total number of 100% of true occurrences and for Damage configuration a total number of true occurrences equal to 97.5%. We can also observe in Table 5 that Delphos is more susceptible to FPs (false positive) than FNs (false negative) since the ARIMA model strongly contributes for providing better results when assessing the presence of a greater number of damage samples. We can observe such fact in the Damage configuration where Delphos achieved a total number of 0% of FN and 2.5% of FP. This behavior is acceptable in the context of SHM, since it is preferable to issue warnings earlier Table 5 Experiment E3 results for WSN composed of 6 sensors nodes.

Table 3 Results of experiment ‘E1′ (considering temperature influence). Mix of NFETS

Undamaged

Neutral

Damaged

Mix of NFETS

Faulty nodes

Smart sensor nodes Metrics

6

12

18

TP TN FN FP TP TN FN FP TP TN FN FP

25.0% 75.0% 0% 0% 50.0% 50.0% 0% 0% 72.5% 25.0% 0% 2.5%

25.0% 75.0% 0% 0% 50.0% 50.0% 0% 0% 70.0% 25.0% 2.5% 2.5%

25.0% 75.0% 0% 0% 50.0% 50.0% 0% 0% 70.0% 25.0% 2.5% 2.5%

Neutral

Undamaged

Damaged

136

Metrics

0

1

2

3

4

FN FP VP VN FN FP VP VN FN FP VP VN

0.00% 0.00% 50.00% 50.00% 0.00% 0.00% 25.00% 75.00% 0.00% 2.50% 72.50% 25.00%

0.00% 0.00% 50.00% 50.00% 0.00% 0.00% 25.00% 75.00% 0.00% 2.50% 72.50% 25.00%

0.00% 0.00% 50.00% 50.00% 0.00% 0.00% 25.00% 75.00% 0.00% 2.50% 72.50% 25.00%

0.00% 0.00% 50.00% 50.00% 0.00% 0.00% 25.00% 75.00% 0.00% 2.50% 72.50% 25.00%

0.00% 0.00% 50.00% 50.00% 0.00% 0.00% 25.00% 75.00% 0.00% 2.50% 72.50% 25.00%

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

Table 6 Experiment E4 results for WSN composed of 6 sensors nodes. Mix of NFETS

Neutral

Undamaged

Damage

Faulty nodes Metrics

0

1

2

3

4

FN FP VP VN FN FP VP VN FN FP VP VN

0.00% 42.50% 7.50% 50.00% 0.00% 20.00% 5.00% 75.00% 0.00% 62.50% 12.50% 25.00%

0.00% 42.50% 7.50% 50.00% 0.00% 20.00% 5.00% 75.00% 0.00% 62.50% 12.50% 25.00%

0.00% 42.50% 7.50% 50.00% 0.00% 20.00% 5.00% 75.00% 0.00% 62.50% 12.50% 25.00%

0.00% 42.50% 7.50% 50.00% 0.00% 20.00% 5.00% 75.00% 0.00% 62.50% 12.50% 25.00%

0.00% 42.50% 7.50% 50.00% 0.00% 20.00% 5.00% 75.00% 0.00% 62.50% 12.50% 25.00%

Fig. 11. Results of the response time.

reduction of 23%, 21% and 19% in the response time, for WSANs with 6, 12 and 18 nodes, respectively. Such difference is related to the fact that the sink node needs to perform the process of damage prediction for each received message, e.g, it must maintain a time series for each sensor in the network. For each prediction made, the sink waits for two indications of damage for triggering the actuator. In the distributed approach, the nodes perform the prediction process in parallel and only wait for collaborating over the final result obtained by each node. Faster actions can be very important to increase the safety of the operation and also to improve the performance of the energy production system that is built upon the wind turbines.

than not informing about the imminence of a damage that may lead to the occurrence of accidents. It is observed in Table 5 that the system performance is not affected by the increase in the number of faulty sensors. This behavior is related to the collaborative process of the sensor nodes, since Delphos ensures that the nodes communicate only with their neighbors nodes (before and after sensor) in order to validate a damage prediction, not being influenced by the total number of faulty sensors. Another point to mention is that Delphos has mechanisms such as message confirmations and timeout verifications which allow keeping the reliability of the damage prediction process even when increasing the number of faulty sensors. From Table 6, we can observe that the system is not efficient with respect to the number of true occurrences when the influence of temperature on the variation of the natural frequencies of the blade is not considered. We can observe that the system achieved a total number of true occurrences for the Neutral, Undamaged and Damage configurations equal to 57.5%, 80% and 37.5%, when the number of faulty sensors varied 1–4, respectively. Again, the system performance is not directly affected by the increase in the number of faulty sensors. As we can notice from the results presented in Tables 5, 6, the temperature can significantly influence the process of predicting damages and Delphos is able to filter such interference as expected. Therefore, as already seen, the temperature negatively influences the damage prediction, since the series of frequency under its influence is more difficult to be modeled by the ARIMA model. Additionally, Delphos also has a higher number of false positives than false negatives due to the use of the ARIMA model, what is considered an appropriate behavior to the problem in question.

5.5. Evaluating the system requirements We evaluated Delphos requirements in terms of memory and energy consumption by considering four different operation modes, named as: (i) Collab/Ack; (ii) Collab/Wack; (iii) WCollab/Ack; and (iv) Wcollab/Wack. In mode (i), nodes collaborate during the step of transmitting the results to the sink node and all messages are confirmed (ACKed). In mode (ii), nodes collaborate on the step of transmitting results to the sink node and messages are not confirmed. In mode (iii), nodes do not collaborate on the step of result transmission and all messages are confirmed (ACKed). Finally, in mode (iv), nodes do not collaborate on the step of results transmission and messages are not confirmed. Such operation modes allow assessing whether the use of collaboration and message acknowledgements causes a significant impact in terms of WSAN resource consumption. The high communication reliability is an important requirement in SHM applications (Kim et al., 2007). Therefore, using an approach based on message acknowledgements for assuring high communication reliability is of great relevance, but it is also important to assess the impact of such approach over the WSAN resource consumption, resulting in a trade-off that needs to be wisely managed. Fig. 12 shows the memory consumption (ROM and RAM) accord-

5.4. Evaluating the gains of decentralization In order to evaluate the gains of a decentralized system (a WSAN with a set of smart sensor nodes, where each one performs the complete prediction cycle) compared to a centralized system (a WSAN with a set of sensor nodes, where each node performs only data collection, natural frequency identification and then transmits data to a central point to be processed), we conducted experiments to measure the response time of the system in both versions (decentralized and centralized). We designed a centralized version of Delphos, in which the sensors only collect data from vibration and temperature, identify the natural frequency and transmit them directly to the sink/ actuator node. In both experiments, the response time of Delphos to perform actions in the turbine was recorded in the same way and was computed as the total time required for completing the execution of a cycle, in which a damage was forecasted. Fig. 11 shows the obtained results. As the figure shows, the decentralized version of Delphos achieves a shorter response time to perform the action in the turbine. There was a

Fig. 12. Results of the memory consumption.

137

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

(SCADA) system. The SCADA system allows monitoring and control of industrial processes. In (Kusiak and Li, 2010), the SCADA system has the function of monitoring, besides the vibration and temperature parameters, also the wind and the energy conversion parameters in order to indicate if a wind turbine is properly working. In our work, we adopted a WSAN composed of six sensor nodes, with an additional sensor performing both roles of sink and actuator. The distribution of damage evolution was set at 75% damaged and 25% without damage. This distribution was chosen since in (Kusiak and Li, 2010) the authors used a sample distribution in which the amount of damage situations was greater than the amount of situations without damage. In (Kusiak and Li, 2010), the accuracy, specificity and sensitivity metrics were used to evaluate the performance of their solution. The accuracy metric has already been defined. The sensitivity metric (TP/(TP+FN)) is used to calculate the proportion of damage that is correctly predicted, that is, this metric measures the system's behavior to predict damage when it actually exists. The specificity metric (TN/(TN+FP)) is used to calculate the ratio of undamaged predictions that are correctly predicted, that is, this metric measures the system's behavior to predict undamaged cases when the damage actually does not exist. Delphos and the work described in (Kusiak and Li, 2010) both obtained results equal to 97.50% and 75.10% for the accuracy metric; 96.67% and 86.34%, for the sensitivity metric and 100% and 64.43% for the specificity metric, respectively. Both works present good sensitivity results and they are close to each other. In both works the sensitivity is relatively high, implying that most damage has been correctly identified. In other words, both works are able to correctly damage prediction when damage will occur. As explained earlier, Delphos has a good result regarding specificity, because the ARIMA model follows the behavior of the series of evolution without damage and this series does not exceed the threshold of abnormality.

Fig. 13. Results of the network lifetime.

ing to each operation mode and considering the MICAz sensor platform. We can observe that in all operational modes, Delphos requires around 30% of total RAM memory and 26% of total ROM memory. This high memory consumption was expected since Delphos operation depends on a number of additional logical components to implement the fuzzy system and the time series forecasting model. As an example, we investigated the memory consumption of the TakagiSugeno fuzzy system alone and it was about 24.14% of RAM and 22.48% of ROM. Nevertheless, we can conclude that the proposed system is feasible since it can run smoothly on the target platform (and still leaving enough space in memory to accommodate code of other protocols and applications). In order to evaluate the system requirements in terms of energy, we used the values of energy consumption in each node given by Avrora to estimate the network lifetime. Fig. 13 shows the estimated network lifetime by considering the four operational modes considered. We can observe that the collaboration process reduced the lifetime of the network. It was an expected result since the distributed algorithm (described in Section 3.3.3) determines that all nodes must generate a message after each execution cycle. When the collaboration process is not active, the nodes generate messages only when a damage prediction is made. However, we can also observe that the use of the collaboration mechanism reduced the network lifetime in only 5 days, a decrease of 0,01% (the lifetime of the network without the collaboration was estimated at about 425 days, while in the case of collaboration it was estimated at about 420 days). This small increase in power consumption is due to the reduction in the signal strength of each node, since each node needs only to communicate with its before_neighbor and after_neighbor. When the collaboration process is not active, on the other hand, the nodes need to send their messages directly to the sink/actuator. Fig. 13 also shows that message acknowledgments cause a minimum decrease in the network lifetime. This justifies the use of acknowledgements for assuring high communication reliability, an important requirement of SHM applications. Another issue is that increasing the number of nodes in the network does not impose a significant reduction in the network lifetime, allowing configuring Delphos to use a greater number of sensors (for example to cover larger areas) without restricting the total network lifetime.

5.7. Evaluating the system execution in real WSAN platforms In that last experiment, we aimed to validate the results obtained by simulations (in terms of accuracy) by comparing them with the results obtained with a real WSAN platform. This experiment was performed with MICAz platforms (as described in Section 4.3.1) in a controlled environment (our research laboratory at UFRJ). In this case, the nodes were kept stationary and disposed on the floor, spread over a distance of 15 m. The distance between each node was 2.5 m for the scenarios with six smart sensors nodes. It is important to note that each node adjusts its signal strength to communicate only with its before and after neighbors, thus minimizing collisions. So, we did not consider the occurrence of collisions in this experiment. We adopted the same metric presented in Section 5.1. We used the same ‘E1′ experiment presented in Section 5.3 (which is responsible for assessing the accuracy of Delphos in terms of correct damage prediction, considering the influence of temperature), except that we only used 10 random damage evolutions samples, from which 5 were Damaged Evolutions samples and 5 were Undamaged Evolutions samples. We adopted a WSAN composed by 6 MICAz sensors and one additional device playing both roles of sink and actuator. In both cases Delphos achieved 100% of true occurrences in terms of prediction damage accuracy. Therefore, these results successfully validate the experiments performed in the simulated environment, since the experiment with real nodes presented the same results.

5.6. Discussion This section presents a discussion regarding a comparison between the results obtained by Delphos (decentralized approach) and the ones obtained by the work of Kusiak and Li (Kusiak and Li, 2010) (centralized approach) in terms of damage prediction in wind turbines. The authors in Kusiak and Li (Kusiak and Li, 2010) make use of data mining algorithms to predict damage in wind turbines. These algorithms were used to model the relationships between the identified parameters and wind turbine vibrations. The data sets used in such work were collected by the Supervisory Control and Data Acquisition

6. Conclusions and future directions In this paper we described our proposed damage prediction system called Delphos. Delphos uses a WSAN to predict damage in wind turbines, and performs control actions to minimize or even avoid accidents. Delphos filters the influence of the environmental temperature over the natural frequencies of the blade to increase the accuracy 138

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al.

RTSS, Barcelona, Spain. Huang, G., Zomaya, A., Delicato, F., Pires, P., 2012. An accurate on-demand time synchronization protocol for wireless sensor networks. J. Parallel Distrib. Comput. 72, 1332–1346. Huang, G., Zomaya, A., Delicato, F., Pires, P., 2014. Long term and large-scale time synchronization in wireless sensor networks. Comput. Commun. 37, 77–91. Hyndman, R., Athanasopoulos, G., 2014. Forecasting: Principles and Practice, OTexts. Jang, R., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 23, 665–685. Ji, S., Liang, Q., Li, L., Huang, L., Wang, J., 2012. Structural damage detection using wireless intelligent sensor networks. Int. J. Smart Home 6, 29–38. Kim, S., Pakzad, S., Culler, D.. et al., 2007. Health Monitoring of Civil Infrastructures using Wireless Sensor Networks, IPSN’07, Cambridge, Massachusetts, USA. Kurian, C., George, V., Bhat, J., Aithal, R., 2006. ANFIS model for the time series prediction of interior daylight illuminance. Int. J. Artif. Intel. Mach. Learn 6, 35–40. Kusiak, A., Li, W., 2010. The prediction and diagnosis of wind turbine faults. Renew. Energy 36, 16–23. Levis, P., Gay, D., 2009. TinyOS Programming. Cambridge University Press, USA. Lim, A., 2011. Smart sensor networks. In: Cutsuridis, V., Hussain, A., Taylor, J. (Eds.), Perception-Action Cycle: Models, Architectures, and Hardware. Springer New York, New York, NY, 675–709. Liu, C., Dewolf, J., 2007. Effect of temperature on modal variability of a curved concrete bridge under ambient loads. J. Struct. Eng. 133, 1742–1751. Loukas, Y., 2001. Adaptive neuro-fuzzy inference system: an instant and architecturefree predictor for improved QSAR studies. J. Med. Chem. 44, 2772–2783. MEMSIC. Imote2 Hardware Bundle, 〈http://www.memsic.com〉, (accessed 28.01.13). MEMSIC. MICAz Wireless Measurement System, 〈http://www.memsic.com〉, (accessed 28.01.13). Miguelanez, E., Lane, D., 2010. Predictive Diagnosis for Offshore Wind Turbines Using Holistic Condition Monitoring. OCEANS, Seattle, WA, USA. Morettin, P., Toloi, C., 2006. Análise de Séries Temporais. Edgard Blucher, São Paulo, Brazil. Moser, P., Moaveni, B., 2011. Environmental effects on the identified natural frequencies of the Dowling Hall Footbridge. Mech. Syst. Signal Process. 25, 2336–2357. Northern. Northern Power® 100 Wind Turbine General Specifications, 〈http://www. northernpower.com/pdf/Northern-Power-100-Product-Brochure-EU-A4.pdf〉, (accessed 28.01.13). Jan. Pham, H., Yang, B., 2010. Estimation and forecasting of machine health condition using ARMA/GARCH model. Mech. Syst. Signal Process. 24, 546–558. Pham, H., Tran, V., Yang, B., 2010. A hybrid of nonlinear autoregressive model with exogenous input and autoregressive moving average model for long-term machine state forecasting. Expert Syst. Appl. 37, 3310–3317. Reyer, M., Hurlebaus, S., Mander, J., Ozbulut, O., 2011. Design of a Wireless Sensor Network for Structural Health Monitoring of Bridges. ICST, Palmerston North, New Zealand. Salfner, F., Lenk, M., Malek, M., 2010. A survey of online failure prediction methods. ACM Comput. Surv. 42, 1–42. Santos, I., Pirmez, L., Lemos, E., Delicato, F., Pinto, L., Souza, J., Zomaya, A., 2014. A localized algorithm for Structural Health Monitoring using wireless sensor networks. Inf. Fusion 15, 114–129. Seber, G., 1977. Linear Regression Analysis. Wiley, New York, NY. Shukla, M., Jharkharia, S., 2011. Applicability of ARIMA Models in Wholesale Vegetable Market: an Investigation. IEOM, Kuala Lumpur, Malaysia. Sohn, H.. et al., 2004. A Review of Structural Health Monitoring Literature: 1996–2001, Los Alamos National Laboratory, 301 p. 〈http://institute.lanl.gov/ei/shm/pubs/LA_ 13976_MSa.pdf〉, Sohn, H., 2006. Effects of environmental and operational variability on structural health monitoring. Philos. Trans. R. Soc. 365, 539–560. Swartz, A., Lynch, J., Sweetman, B., Rolfes, R., Zerbst, S., 2010. Structural monitoring of wind turbines using wireless sensor networks. Smart Struct. Syst. 6, 1–14. Wang, H., Chen, B., Lin, C., 2013. Adaptive neural tracking control for a class of stochastic nonlinear systems with unknown dead-zone. Int. J. Innov. Comput. Inf. Control 9, 3257–3269. Wu, H., Zonta, D., Pozzi, M.. et al., 2009. Real-Time Health Monitoring of Historic Buildings with Wireless Sensor Networks. 7th IWSHM, Stanford, California, USA. Wu, Z., Shi, P., Su, H., Chu, J., 2012. Exponential synchronization of neural networks with discrete and distributed delays under time-varying sampling. IEEE Trans. Neural Netw. Learn. Syst. 23, 1368–1376. Xia, Y., Hao, H., Zanardo, G., Deeks, A., 2006. Long term vibration monitoring of an RC slab: temperature and humidity effect. Eng. Struct. 28, 441–452. Zhang, H., Guo, J., Xie, X., Bie, R., Sun, Y., 2013. Environmental Effect Removal Based Structural Health Monitoring in the Internet of Things, 7th IEEE IMIS, Taiwan, pp. 512–517. Zhang, L., Liu, J., Lai, J., Xiong, Z., 2014. Performance analysis of adaptive neuro fuzzy inference system control for MEMS navigation system. Math. Probl. Eng. 2014, 7.

of the damage prediction process. The results of the experiments to evaluate the efficiency of Delphos show that it performs well when the Environmental Interference Filter is enabled (i.e., when the system takes into account the influence of the temperature). The experiments to evaluate the energy consumption show that the WSAN presents a lifetime of around 400 days. This is satisfactory, because wind turbines may suffer, in average, one damage incident at every 120 days (Swartz et al., 2010). The experiments performed with a real smart sensor platform show the feasibility of our proposal, since Delphos performed successfully, considering the limitations of memory and processing capabilities of the WSAN smart sensors. Future work in this research comprises: (i) the inclusion of other Damage Evolution profiles, such as an exponential curve shape, which could allow the system to deal with the sudden occurrence of events (such as bird crashes on the blades of the wind turbine); (ii) the use of a reliable communication protocol between the smart sensor and actuator or sink nodes; (iii) the improvement in the quality of the real data collection step, using a climate chamber with the appropriated control of the environmental temperature and humidity variation; and (iv) the addition of other environmental variable, such as the humidity, in the Environmental Interference Filter. Acknowledgments This work was partially supported by Brazilian Funding Agencies FAPERJ (Grant 213967) and CNPq (under Grants 200757/2015-6, 307378/2014-4, 310958/2015-6, 457783/2014, and 200758/2015-2). The work of Igor L. dos Santos is supported by a scholarship from CAPES Foundation, Ministry of Education of Brazil. References Ackermann, T., 2005. Wind Power in Power Systems. John Wiley & Sons Ltd, Chichester, UK. Akyildiz, I., Vuran, M., 2010. Wireless Sensor Networks. John Wiley & Sons Ltd, Chichester, UK. Alberola, R., Pesch, D., 2008. AvroraZ: extending Avrora with an IEEE 802.15.4 compliant radio chip model, PM2HW2N '08, Vancouver, Canada. Bhuiyan, M., Wang, G., Wu, J., Cao, J., Liu, X., Wang, T., 2015. Dependable structural health monitoring using wireless sensor networks. IEEE Trans. Dependable Secur. Comput., 1. Bocca, M., Toivola, J., Eriksson, L., Hollmén, J., Koivo, H., 2011. Structural Health Monitoring in Wireless Sensor Networks by the Embedded Goertzel Algorithm. ICCPS, Chicago, IL, USA. Bououden, S., Chadli, M., Allouani, F., Filali, S., 2013. A new approach for fuzzy predictive adaptive controller design using particle swarm optimization algorithm. Int. J. Innov. Comput. Inf. Control 9, 3741–3758. Box, G., Jenkins, G., Renseil, G., 1994. Time Series Analysis: Forecasting & Control. Prentice Hall, Upper Saddle River, NJ, USA. Buonadonna, P., Hill, J., Culler, D., 2001. Active message communication for tiny networked sensors. In: Proceedings of the XX Annual Joint Conference of the IEEE Computer and Communications Societies, IEEE, Anchorage, Alaska, USA, p. 11. Chiang, H., Chen, Y., Lee, T., 2013. Multi-Stage with neuro-fuzzy approach for efficient on-road speed sign detection and recognition. Int. J. Innov. Comput. Inf. Control 9, 2919–2939. Chou, J.-S., Chiu, C.-K., Huang, I.-K., Chi, K.-N., 2013. Failure analysis of wind turbine blade under critical wind loads. Eng. Fail. Anal. 27, 99–118. Clayton, E., Qian, Y., Orjih, O., Dyke, S., Mita, A., Lu, C., 2006. Off-the-Shelf Modal Analysis: Structural Health Monitoring with Motes, XXIV IMAC, Dallas, USA. CPTEC/INPE. Historical Database. 〈http://bancodedados.cptec.inpe.br/〉. (accessed 28. 01.13). Croxford, A., Moll, J., Wilcox, P., Michaels, J., 2010. Efficient temperature compensation strategies for guided wave structural health monitoring. Ultrasonics 50, 517–528. Delurgio, S., 1998. Forecasting Principles and Applications. McGraw-Hill, Singapore. Enersud Energia Limpa . 〈http://enersud.com.br〉, (accessed 28.01.13). Faria, E., Albuquerque, M., Alfonso, J., Cavalcante, J., 2008. Previsão de Séries Temporais utilizando Métodos Estatísticos, CBPF technical note: CBPF-NT-003. Flores, J., 2009. Comparação de Modelos MLP/RNA e Modelos Box-Jenkins em Séries Temporais Não-Lineares, MSc. diss. Universidade Federal do Rio Grande do Sul. Garcia, F., Pedregal, D., Roberts, C., 2010. Time series methods applied to failure prediction and detection. Reliab. Eng. Syst. Saf. 95, 698–703. Griffith, D., Casias, M., Smith, G., Paquette, J., Simmermacher, T., 2006. Experimental Uncertainty Quantification of a Class of Wind Turbine Blades, XXIV IMAC, Dallas, USA. Hackmann, G., Sun, F., Castaneda, N., Lu, C., Dyke, S., 2008. A Holistic Approach to Decentralized Structural Damage Localization Using Wireless Sensor Networks.

139

Journal of Network and Computer Applications 80 (2017) 123–140

M.M. Alves et al. Maicon Melo Alves has a Master degree in Computer Science from Federal University of Rio de Janeiro (UFRJ), Brazil in the area of Computer Networks and Distributed Systems. He is an expert in Computer Networks from ESAB, a Bachelor of Computer Science from UCAM and a Technician in Data Processing. He is currently pursuing his PhD degree at the Fluminense Federal University and is employed by Petrobras, where he works in the administration and support of infrastructure for a scientific system of oil exploration. He has 15 years of experience in his field, published two books (Linux and Linux Sockets & Performance Monitoring) and has conducted presentations on Linux and networks. He is a Red Hat Certified Virtualization Administrator (RHCVA) at Red Hat.

Claudio Miceli de Farias received a M.Sc. degree on Computer Science in 2010 and his doctorate degree in 2014 from the Federal University of Rio de Janeiro, Brazil. He is nowadays professor at the Tercio Pacitti Institute for Applications and Computational Research. His research interests include nanonets, wireless sensor networks, network security, VOIP, real-time communications and video processing.

Paulo F. Pires is an associate professor at the Department of Computer Science at Federal University of Rio de Janeiro (Brazil), and member of the Centre for Distributed and High Performance Computing, University of Sydney (Australia). His main research interests are model driven development, Internet of Things, infrastructures for Web service composition, and application of Software Engineering techniques in the development of software systems for emerging domains, such as embedded, ubiquitous and pervasive systems. He holds a technological innovation productivity fellowship from the Brazilian Research Council (CNPq) since 2010 and is a member of the Brazilian Computer Society (SBC).

Luci Pirmez is a professor at the Institute of Informatics of the Federal University of Rio de Janeiro (UFRJ), Brazil. She received her M.Sc. and Ph.D. degree, both in computer science from the Federal University of Rio de Janeiro in 1986 and 1996, respectively. She is a researcher at the Computer Centre of Federal University of Rio de Janeiro. Her research interests include wireless sensor networks, Internet of Things and security. She is one of 300 researchers in computer science from all over Brazil selected as a CNPq Fellow (CNPq is the technology research branch of the Brazilian government). She is currently involved in a number of research projects with funding from Brazilian government agencies, in the areas of wireless networks, wireless sensor networks, network management and security.

Igor Leão dos Santos received his Master degree in 2013 from the Federal University of Rio de Janeiro (UFRJ), where he is currently pursuing a Doctorate degree in Informatics. He is also a lecturer at UFRJ. His research interests include Wireless Sensor and Actuator Networks, Cloud Computing, Information Fusion and Structural Health Monitoring.

Silvana Rossetto was graduated in Computer Science from Federal University of Espirito Santo (1998), received her Master degree in Computer Science from Universidade Federal do Espirito Santo (2001) and D.Sc. in Computer Science from the Catholic University of Rio de Janeiro (2006). Her current areas of interest are distributed systems and wireless sensor networks. She holds the position of Professor in the Department of Computer Science (DCC), Institute of Mathematics (IM), Federal University of Rio de Janeiro (UFRJ).

Albert Y. Zomaya is the chair professor of high-performance computing and networking in the School of Information Technologies at Sydney University. His research interests are in the areas of complex systems, parallel and distributed computing, and green computing. Zomaya has a PhD in control engineering from Sheffield University, UK. Zomaya is a Fellow of AAAS, IEEE, and IET (UK).

Flávia C. Delicato is an associate professor at the Department of Computer Science at Federal University of Rio de Janeiro (Brazil), and member of the Centre for Distributed and High Performance Computing, University of Sydney (Australia). She is a Level 1 Researcher Fellow of the National Council for Scientific and Technological Development (CNPq) and a Young Researcher Fellow from FAPERJ Brazilian Funding Agency. She has more than 150 papers published in journal and conferences. She has been participating in several research projects with funding from International and Brazilian government agencies. Her main research interests are: wireless sensor networks and ubiquitous computing; adaptive middleware; model-driven development; System of Systems and internet of things.

140