An anti-jamming artificial immune approach for energy leakage diagnosis in parallel-machine job shops

Computers in Industry 101 (2018) 13–24

Contents lists available at ScienceDirect

Computers in Industry journal homepage: www.elsevier.com/locate/compind

An anti-jamming artificial immune approach for energy leakage diagnosis in parallel-machine job shops

T

⁎

Jianhua Guoa, , Haidong Yangb a

School of Computer Science, Guangdong Polytechnic Normal University, No. 293, Zhongshan Avenue, Guangzhou, Guangdong, China School of Mechatronics Engineering,Guangdong University of Technology, No. 100 Waihuan Xi Road Guangzhou Higher Education Mega Center, Guangzhou, Guangdong, China b

A R T I C LE I N FO

A B S T R A C T

Keywords: Artificial immune system Danger model Artificial immune network Fault detection Energy leakage Tyre vulcanization

Energy leakages lead to enormous economic losses and environmental pollution for many manufacturing systems. Thus, energy leakage diagnosis becomes an essential requirement for an economical and environmentfriendly production. However, for parallel-machine job shops, it is still a challenge to isolate leaking machines from a jamming environment with limited energy measurements, stochastic parallel jobs, and a fluctuating energy supply. To address this challenge, an anti-jamming artificial immune approach that combines the danger model with an immune network is proposed in this study. The proposed approach involves strategies, namely a danger-model-inspired framework, an anti-jamming antigen feature, and an anti-jamming aiNet (AJ-aiNet) algorithm, which are specifically dedicated to overcome the jamming factors. The danger-model-inspired framework realizes the collaboration between danger (energy loss) detection at shop level and antigen (process behaviour) isolation at machine level. The anti-jamming antigen feature, called the difference in process behaviour fluctuation (DPBF), acquires characteristics that are sensitive to energy leakage from the process parameter time series. Antibody ageing and antigen killing strategies are embodied within aiNet to mitigate the disturbance of jamming antigens between leaking and normal clusters. In order to evaluate the proposed approach, several experiments were performed in a tyre vulcanization shop floor to diagnose the steam leakage of steam traps. The results show that the proposed approach achieved the isolation of leaking machines without energy consumption measurements at the machine level. As anti-jamming strategies, DPBF can separate leaking and normal process behaviours effectively, AJ-aiNet can achieve clustering of DPBF samples with jamming antigens correctly, and the collaboration of danger detection can significantly suppress false diagnoses and improve the time efficiency of diagnosis.

1. Introduction With the increasing level of complexity and automation in manufacturing processes, manufacturers need more effective and efficient techniques to monitor the operation status and diagnose process faults of machines to enable a sustainable, economical, and environmentfriendly production [1–3]. Energy leakage, such as steam, compressed air, and cool air leakage from broken pipelines or worn valves, is a typical fault in job shops with high-energy consumption. This fault tends to cause enormous economic losses and environmental pollution [4–6], and hence, its diagnosis becomes an essential requirement for an economical and environment-friendly production. Energy leakage diagnosis is a combination of fault detection (which identifies if there is an energy leakage) and isolation (which determines the location of the energy leakage) [7]. Owing to the complexity of the

⁎

Corresponding author. E-mail address: [email protected] (J. Guo).

https://doi.org/10.1016/j.compind.2018.05.004 Received 1 December 2017; Received in revised form 29 May 2018; Accepted 30 May 2018 0166-3615/ © 2018 Elsevier B.V. All rights reserved.

industrial environment, energy leakage diagnosis is a very challenging activity. Bayar et al. [1] and Wang et al. [8] pointed out that in order to develop a fault diagnosis system, several limitations (e.g. data acquisition and processing, tolerances and sensitivity to change, and alarm frequency and quality) should be considered. For energy leakage diagnosis in many industrial environments, energy consumption is measured at the shop/group level rather than at the machine level owing to the issues of technology, such as too small branch pipe, too narrow installation space, strong vibration of machine, etc. The limited energy measurements increase the difficulty in isolating the leaking machines. Moreover, stochastic parallel jobs with multi-process and random operation rhythm reduce the sensitivity of energy consumption to energy leakage. Furthermore, the fluctuating energy supply and environmental dynamics confuse the fluctuation of process parameters caused by energy leakage. In particular, for parallel-machine job shops, independent

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

energy measurements, stochastic parallel jobs, and a fluctuating energy supply. In this study, three hypotheses are under consideration. First, an energy leakage is associated with individual machine in a parallelmachine job shop. Second, energy consumption at shop/group level and process parameters at machine level can be acquirable. Third, energy leakage can’t be accurately isolated by analysing jammed process parameters. Thereafter, an anti-jamming artificial immune approach (AJAI approach) that combined the danger model with an immune network is proposed. The main contributions of this paper rely on three main suggestions to overcome the jamming environment: a dangermodel-inspired framework, an anti-jamming antigen feature, and an anti-jamming aiNet (AJ-aiNet) algorithm. The danger-model-inspired framework realizes the collaboration between danger (energy loss) detection at the shop level and antigen (process behaviour) isolation at the machine level. This dual activation strategy aims to suppress false alarm rates caused by the jamming environment. An anti-jamming antigen feature, called the difference in process behaviour fluctuation, is presented to enhance the separability between leaking and normal process behaviours. Antibody ageing and antigen killing strategies are embodied within aiNet to mitigate the disturbance of jamming antigens between leaking and normal clusters. The paper is organized as follows. Section 2 overviews the main principles and applications of the danger model and artificial immune network. The proposed approach is presented in Section 3. The computer-based implementation is presented in Section 4. Section 5 introduces an application case. The experiments and results are described in Section 6. The advantages and limitations of the proposed approach are discussed in Section 7, and the conclusions drawn are presented in Section 8.

machines and parallel jobs make energy leakage isolation more complicated. Therefore, for parallel-machine job shops, an energy leakage diagnosis approach needs to overcome a jamming environment with limited energy measurements, stochastic parallel jobs, and a fluctuating energy supply. In recent years, many approaches for energy leakage diagnosis have been presented based on fluid mechanics. These approaches, such as pressure-based diagnosis [9–11], negative-pressure-wave-based diagnosis [12], electromagnetic-wave-based diagnosis [13], or hybrid techniques [14], focus more on identifying a disruption or mutation of parameters in energy transmission pipelines with regular topology and stable flow, rather than on isolating a potential leaking machine from a jamming industrial environment. These approaches are not suitable for energy leakage diagnosis in parallel-machine job shops because they ignore the fluctuating energy supply, environmental dynamics, and limited data acquisition. In addition, based on energy consumption mechanisms, various energy consumption models were proposed to diagnose energy leakage. Duflou et al. [15] proposed an organizational energy consumption model including five levels: device/unit process, line/cell/multi-machine system, facility, multi-factory system, and enterprise/global supply chain. Rahimifard et al. [16], Wang et al. [17], and Li et al. [18] proposed productive energy consumption models including three levels: process, product, and production. Bi and Wang [19] and Pfefferkorn et al. [20] suggested technological energy consumption models including three levels: theoretical, technical, and real. The energy leakage diagnosis approaches based on these models focused on evaluating energy consumption baselines at different levels. The authors [21,22] suggested to build energy consumption baselines from historical data using artificial neural networks [23] and support vector machines [24], or to evaluate leakage risks from energy balances and mass balances [25,26]. However, the authors found that these approaches have two deficiencies in real applications. First, the feasibility of isolating leaking machines is heavily dependent on energy measurements at the machine level. Second, the false alarm rate is high owing to the insensitivity of energy consumption caused by stochastic parallel jobs. Biological immunity inspired the design of promising approaches for fault diagnosis in manufacturing fields [1,27–29]. These approaches focused on designing conceptual frameworks for fault diagnosis with the inspiration from immune mechanisms (e.g. collaboration, pattern recognition, learning, and memory) [28,30,31], or on suggesting artificial immune algorithms for a specific machine or system fault diagnosis [32–35]. However, only few immune approaches for hidden faults such as energy leakage or dealing with a jamming environment are found. Therefore, for parallel-machine job shops, it is still a challenge to isolate leaking machines from a jamming environment with limited

2. Danger model and artificial immune network The proposed approach is inspired from the danger model and artificial immune network. In this section, the principles and applications related to these two immunity concepts are overviewed. 2.1. Danger model Before the proposal of the danger model, immunologists focused their thoughts and applications on the functions of the immune system by making a distinction between self and non-self (SNS model) [36,37]. However, this paradigm has failed to explain problems associated with self-change. The danger model proposed by Matzinger [38–40] outlines a model of immunity based on the idea that the immune system is more concerned with entities that do damage than with those that are foreign. One of the important affirmations of the danger model is that the

Fig. 1. Diagram of danger model [29]. 14

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

fault detection and concluded that the methods are generally applied to classify certain fault samples with no particular focus. They have been widely used to develop fault detection applications, such as for power transformer systems [53,54], gearbox systems [55], and RFID readerto-reader collision [56]. For the energy leakage diagnosis in parallel-machine job shops, artificial immune networks provide methods to classify process behaviours into leaking and normal clusters. Nevertheless, aiNet and its improvements still need to solve the issue that they are sensitive to jamming environments. The issue lies in the fact that an antibody population accepts an individual antigen independently and neglects the cumulative effect of the antigen population. From the viewpoint of information processing, the learning algorithm stresses the space distribution of the dataset but neglects the density distribution.

immune system is activated by danger signals from injured cells, such as those exposed to pathogens and mechanical damage instead of foreign molecules [28,37]. Fig. 1 depicts an immune response of the danger model [29]. A cell that is in distress sends out a danger signal, whereupon antigens in the neighbourhood are captured by antigenpresenting cells (APC) such as macrophages, which then present the antigens to lymphocytes. Essentially, the danger signal establishes a danger zone around itself. Thus, B cells producing antibodies that match antigens within the danger zone get stimulated and undergo a clonal expansion process. Those that do not match or are too far away do not get stimulated [29]. In this sense, the danger model suggests collaboration with other immune principles, such as negative selection, clonal selection, and immune network. With the collaboration between antigen recognition (signal one) and co-stimulation (danger signal), the immune system effectively protects against the chance of accidentally reacting to self [29]. This provides the underlying analogy to suppress false detections and improve the time efficiency for a fault diagnosis. Inspired from the danger model, Bayar et al. [1] suggested a conceptual danger model for improving the effectiveness of a fault diagnosis in industrial applications. Laurentys et al. [28] proposed an approach for a rapid fault detection and achieved a low false rate and high time efficiency, and the authors proposed a collaborative framework to realize an implicational radio frequency identification (RFID) intrusion detection [41]. For an energy leakage diagnosis in parallel-machine job shops, the danger model provides an analogy to realize collaboration between information at the shop level and that at the machine level. This collaboration is helpful to mitigate the impact of a jamming environment, particularly with limited energy measurements, and to suppress false diagnoses and improve the time efficiency of diagnosis.

3. Proposed approach 3.1. Framework For parallel-machine job shops, energy leakage will be observed on both the energy consumption of the shop and process parameters of the machines, which are supposed to be available. With the inspiration from the danger model, the motivation of the proposed AJAI approach is to overcome the jamming environment by building a dual activation mechanism that combines the sensitivity of energy consumption and the isolation capability of process behaviours. In analogy with the danger model proposed by Matzinger [40], energy is considered as the protected body cells. Accordingly, energy consumption can be considered as cell death. It is important to distinguish two types of cell deaths:

• Apoptosis death: This represents the technology energy consumption. • Necrosis death: This represents the energy loss including energy

2.2. Artificial immune network The main idea of the immune network theory [42] is that the immune system maintains an idiotypic network of interconnecting cells for antigen recognition. These immune cells interconnect with each other in certain ways that lead to stabilization of the network. In such ways, two immune cells are connected if the affinities they share reach a certain threshold, and the strength of the connection is directly proportional to the affinity they share. The interactions of the immune cells result in a network with a natural eigen-behaviour, whose state will be disturbed by antigens. Artificial immune networks stress clonal selection and affinity maturation [43] as well as the immune network theory, and are aimed at solving problems related to pattern recognition and data clustering. Farmer and Packeard [44] proposed a continuous immune network model that represented immune cells and molecules as binary strings in a Hamming shape-space. Strings were allowed to match complementarily in any possible alignment, modelling the fact that two molecules may react in more than one way. Timmis and Neal [45] and Castro and Zuben [46] proposed an immune network learning for acquiring the internal image of a dataset and then detecting clusters, named resource limited artificial immune network (RAIN) and artificial immune network (aiNet), respectively. The RAIN learning algorithm produces a topological representation of the antigenic patterns, and the aiNet learning algorithm acquires a reduced spatial distribution of the antigenic universe. In order to interpret the resultant aiNet, various graph concepts and hierarchical clustering techniques can be utilized, such as minimum spanning trees and dendrograms [27]. Afterwards, Castro and Timmis [47] redefined the Ab-Ag affinity and proposed optaiNet for function optimization and copt-aiNet for combinatorial optimization. Recently, some researchers improved aiNet to quickly capture a better solution by adding learning mechanisms such as social learning (AINET-SL) [48], qualitative model learning [49], user-item rating [50], local search strategy [51], and particle swarm optimization [52]. Bayar et al. [1] summarized the immune network-based methods for

leakage.

Energy leakage points are considered as antigens of which features are interpreted by certain process behaviours of machines. In a detection period, a danger signal could thus be interpreted as an event in which the energy loss exceeds baselines. Without energy leakage, energy is released at very low levels; otherwise, it is supposed to be released at significant levels. Based on the above concepts, the framework of AJAI approach is designed as shown in Fig. 2. The approach consists of two subsystems: shop detection subsystem and machine isolation subsystem. They are in charge of the function of APCs and lymphocytes (B cell and T helper), respectively. The shop detection subsystem determines if there is energy leakage in the job shop (danger signal) and the machine isolation subsystem determines the machines where the energy leakage is located (antigen recognition signal). Both subsystems consist of presentation, learning, and monitoring modules. Considering the jamming environment and the operating characteristics in parallel-machine job shops, three anti-jamming strategies are embedded in the framework.

• Energy loss is separated from energy consumption and used as the source of danger signal. • As an anti-jamming feature, the difference in process behaviour fluctuation (DPBF) is presented from parameter time serials. • The AJ-aiNet algorithm is presented to enable a cluster antigen sample with noise.

It is important to illustrate the dual activation mechanism of the framework. Normally, only the monitoring at the shop level keeps working. Once the danger is determined, the monitoring module establishes a danger zone based on the measurement scope and detection 15

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Fig. 2. Framework of AJAI approach.

transmission for a specific job process. In [22], the authors presented the physical and technology steam consumption models for a tyre vulcanization job. It provided a reference methodology for other job processes, and thus, this study will not discuss in detail the modelling method for the physical and technology energy consumption any more. The predictive baseline model aims to evaluate the normal energy loss based on historical statistical data. The authors suggest building the model in three steps. First, input factors related to the job process should be collected. Then a predictive mode between input factors and energy loss (output) gets trained using historical statistical data. Finally, the baseline mode of energy loss is evaluated based on the residual between the evaluated and predicted energy loss. Artificial neural networks and support vector machines can be used to train the predictive mode. In [22], the authors suggested to build a steam loss baseline for tyre vulcanization using the Levenberg–Marquardt back propagation algorithm, and the total time of vulcanizers running, total time of vulcanization job, shop air temperature, amount of vulcanizing process, and total mass of vulcanized tyres were the input factors. While monitoring, a danger signal will be sent if the evaluated energy loss goes beyond the baseline.

period associated with the energy loss, and then the function of determining antigen recognition is activated. A warning response is only sent out if both danger and antigen are positive. 3.2. Energy loss presentation model and predictive baseline model The energy loss presentation model is built on organizational energy consumption models [15] and technological energy consumption models [19,20]. In this model, the energy system in a parallel-machine job shop is considered as a hierarchical structure as shown in Fig. 3. For generality, the machine here is a logical concept, an entity with independent manufacturing capability that can be abstracted as a logical machine. A physical machine with concurrent process capability may be divided into several logical machines. A parallel-machine job shop is considered as a group of identical logical machines. In a detection period, a machine performs multiple jobs and each job consumes a certain amount of energy. Therefore, from the viewpoint of organization, the energy consumption is decomposed into shop level, machine level, and job level. Then, from the viewpoint of technology, the energy consumption at the job level can be split into technology energy consumption and energy loss, and finally the former is split into physical energy consumption and discharged energy. The energy loss presentation model for a parallel-machine job shop is shown in Fig. 4. Physical energy consumption is the required energy from physical laws, and it is generally used to heat, compress, and form. Discharged energy is energy within emissions. Technology energy consumption is the expected energy consumption according to the theory of energy transmission. Thus, energy loss, which is the difference between the measured energy consumption and technology energy consumption, has been separated from energy consumption in the energy decomposition model. The physical energy consumption and technology energy consumption need to be evaluated from the energy conversion and

3.3. Antigen feature presentation model For energy consuming jobs, such as heating, cooling, and pressing, the temperature and pressure at key locations are the main parameters and their time series generally show periodical fluctuation characteristics. Typical examples are jobs that require a stable temperature or pressure. In automatic machines, the energy input will start when the observed temperature or pressure is lower than a given value, and will pause when the observed temperature or pressure reaches the given value. Thus, the parameters fluctuate around their given values. Energy leakage would reduce energy efficiency and then generally results into a change of period and amplitude of fluctuation. Therefore, the period 16

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Fig. 3. Hierarchical structure of energy system in a parallel-machine job shop.

pr11 pr12 ⎛ pr22 21 Pr = ⎜ pr ... ... ⎜ pr ⎝ Nm,1 prNm,2

... pr1, Nr ⎞ ... pr2, Nr ... ... ⎟ ... prNm, Nr ⎟⎠

(1)

where Pr is the matrix of process parameters, pr is an element of Pr, Nm is the number of machines, and Nr is the number of process parameters. The continuous values of pr in a detection period form a time series. For a process parameter prij, the time series is represented by Eq. (2).

prTij = (prij1, prij2, ..., prijT )

(2)

prTij

where is the time series of prij, and T is the length, which is also the sensed times of prij in a detection period. The average period and amplitude of fluctuation of prTij can be calculated from the peaks and valleys. A prTij can be split into an upper part and lower part by a line between peaks and valleys, and each part is split into multiple continuous sub-series. The maximum element of each sub-series of the upper part is a peak, and the minimum element of each sub-series of the lower part is a valley. The time interval between two adjacent peaks is the fluctuation period, and the difference between a peak and its adjacent valley is the fluctuation amplitude. The average fluctuation period τijT or amplitude aijT is the average of all fluctuation periods or amplitudes of prTij . If the ith machine and the kth machine are an intercomparable machine pair, the DPBF can be calculated using Eqs. (3)–(5).

Fig. 4. Energy loss presentation model for a parallel-machine job shop.

and amplitude of fluctuation of process parameters may be a suitable information source for antigen (leakage point) features. In a parallel-machine job shop, multiple similar processes are run on multiple replaceable machines in parallel. Therefore, there generally exist two or more machines with similar process settings and similar machining speed. These machines can be grouped as intercomparable machines, and they should have similar curves of process parameters. Practitioners usually recognize faults by comparing the parameters between two intercomparable machines. This comparison can help to ignore the impact of stochastic parallel jobs and mitigate other disturbances caused by environmental dynamics. In this model, two intercomparable machines are configured as an intercomparable machine pair, and then the novel process feature, DPBF, is defined based on the difference in fluctuation of process parameters between the intercomparable machine pair. The obtainable process parameters in a parallel-machine job shop can be represented by Eq. (1).

δikj = < δτikj,

δaikj >

(3)

δτikj = τij-τkj

(4)

δaikj = aij-akj

(5)

prTij

prTkj ,

where δikj is the DPBF between and andδτikj and δaikj are the differences in average period and amplitude between prTij and prTkj , respectively. An antigen feature consists of DPBFs of all process parameters and is represented by Eq. (6). 17

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Agik = (δik1,δik 2 , ..., δik, Nr )

1.1.6 Reselect ζ% of the highest affinity cells and create a partial Mp memory cell matrix; 1.1.7 Eliminate those cells, whose affinity is inferior to a threshold σd, yielding a reduction in the size of the Mp matrix; 1.1.8 Calculate the network Ab-Ab affinity, sij; 1.1.9 Eliminate sij < σs (clonal suppression); 1.1.10 Concatenate M and Mp, (M⟵[M;Mp]); 1.2 Determine S, and eliminate those cells whose sij < σs (network suppression); 1.3 Determine G and then sort by descending order; eliminate those cells from the last element with gradi > σg, and eliminate those recognized Ags by those cells from X (Ab ageing and Ag killing); 1.4 Replace r% of the worst individuals; 2. Test the stopping criterion.

(6)

where i and k are the ith and kth machine, which are grouped as an intercomparable machine pair, Ag is the antigen feature, and δ is the element associated with the process parameter. The antigen feature presentation model can be summarized in the following steps: Step I: Setting an intercomparable machine pair: set two intercomparable machines as an intercomparable machine pair. Generally, two machines sharing one control system or one operator team are grouped as a pair. Step II: Calculating the average period and amplitude of fluctuation of each process parameter; Step III: Calculating the DPBFs for each intercomparable machine pair.

C = C − α(C − X) 3.4. Anti-jamming artificial immune network algorithm

where C is the matrix of network cells, X is the matrix of antigens, and α is the learning rate or mutation rate. α is set according to the Ab-Ag affinity: the higher the affinity, the smaller the α. Compared with the classical aiNet, AJ-aiNet just added step 1.3. The Ab ageing strategy eliminates the disturbance of sparse Ag and enhances the boundary among clusters. The Ag killing strategy avoids the repeated interference of sparse Ag and improves the computing efficiency. Based on the immune network M and distance matrix D, the succeeding steps will follow the cluster analysis method of aiNet in Castro and Zuben [46], which considers M and D as a graph, and then generates the minimal spanning tree (MST) of the graph. Finally, the MST is divided into a few subtrees by pruning the longest branches. Each substree is a cluster and represents the feature space of a certain process state (leaking or normal), and one cluster will be labelled with the state with which most of its recognized Ag is labelled. While monitoring, a new Ag will be matched by the nearest Ab, and the Ag will be identified by the state label of the cluster in which its matched Ab belongs.

In a biological immune system, mature Abs will naturally die due to ageing if they do not receive stimulation of Ag for a long time, and some Ags will not appear after being killed. With this inspiration, the antijamming aiNet (AJ-aiNet) is introduced to mitigate the disturbance of noise points. Compared with the classical aiNet, AJ-aiNet only adds Ab ageing and Ag killing strategies in the learning process. In AJ-aiNet, the suppression threshold σs is also defined as the recognition distance of Ab to Ag; an Ab will recognize the Ags within the recognition distance and receive their stimulation. Ab ageing is based on the number of recognized Ags. In order to accommodate different Ag distribution densities, the Ab ageing strategy is designed on a recognition gradient. The recognition number of Abs is calculated at first and then is sorted by descending order, and the recognition gradient is represented by Eq. (7).

gradi =

Nrgi − 1 − Nrgi Nrgi

(8)

(7)

where gradi is the recognition gradient of the ith Ab, and Nrg is the recognition number of Ab. The recognition gradient reflects the change of Ag density, and a considerable recognition gradient means a mutation of Ag density. The Ags after the mutation should be considered as noise points. Then the subsequent Abs are aged and moved from the immune network, and the recognized Ags are moved from the Ag population. The recognition gradient is a relative concept, and it is easy to set an empirical value without considering a sample distribution density. The following notation is adopted in the AJ_aiNet algorithm: X: the learning sample of Ag, a dataset composed of Np vectors, C: matrix containing all the Nt network cells; M: matrix of the N memory cells, (M ⊆ C); Nc: number of clones generated by each stimulated cell; D: dissimilarity matrix with elements dij(Ab-Ag); S: similarity matrix with elements sij (Ab-Ab); n: n highest affinity cells selected to clone and mutate; ζ: percentage of the matured cells to be selected; σd,s,g: natural death, suppression, and ageing threshold, respectively; and G: recognition gradient vector with elements gradi. The AJ-aiNet learning algorithm works as follows: 1. At each iteration step, do: 1.2 For each antigen i, do: 1.1.1 Determine its affinity, dij, to all the network cells according to a distance metric; 1.1.2 Select the n highest affinity network cells; 1.1.3 Clone these n selected cells. A higher cell affinity means a larger Nc; 1.1.4 Apply Eq. (8) to these Nc cells; 1.1.5 Determine D for these improved cells;

4. Computer-based implementation The AJAI approach is driven by real time shop floor and machine data streaming, which are big data sets for large plants. In order to meet the real-time requirement of energy leakage diagnosis, an implementation is presented on the basis of computer cluster and concurrent streaming processing. 4.1. Technical infrastructure The infrastructure of AJAI implementation is shown in Fig. 5. It consists of Supervisory Control and Data Acquisition (SCADA), Hadoop Distributed File System (HDFS), Apache Cassandra, Apache Kafaka and Apache Spark. Cassandra together with HDFS supports distributed storage, management and access for big data sets on a computer cluster. As a message queue, Kafaka provides a distributed streaming platform to transfer data from one application to another. Spark is a fast and general engine for large-scale data processing, and enables MapReduce programming and stream processing of live data streams. Live data streams are sourced from energy consumptions and parameters of machines or machine groups. Raileanu et al. proposed an agent-based approach for measuring parameters of resources [57] and then used this information for weighting operation types of resource [58]. The approach can be used as a reference for high frequency data acquisition of SCADA. The modules of AJAI approach are implemented as MapReduce applications. Energy loss presentation, antigen feature presentation, energy leakage detection and energy leakage isolation are driven by the live data streams from SCADA. With the support of Spark engine, they 18

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Fig. 5. The infrastructure of AJAI implementation.

process the streams in time-based batch mode, and generate the resulting streams in batches. Inputs and outputs are consolidated as discretized streams, and transferred through Kafaka in real time. The resulting streams of energy leakage are reported to the alarm interface in real time. Source stream, middle stream (energy loss, antigen) and resulting stream persist in Cassandra database. Predictive baseline learning and immune network learning can be executed in a continuous batch mode, such as week-based batch or month-based batch. It is necessary for the two learning modules to be implemented as MapReduce mode since they are time-consuming. Learning results will be transferred to database and detection modules (energy leakage detection and energy leakage isolation). Predictive baseline learning can use linear regression, neural network, support vector machine, etc., to seek a pattern in training sample, and allow energy consumption prediction in short-term, such as a job period. Immune network learning can be run as an incremental mode. The resulting immune cells can be added in the existing immune network, and some inactive immune cells can be removed. The learning mode in continuous batch can adapt to the changes of environment, machine and technology, and reduce false and missing detections.

Table 1 Job 1: computing the average of fluctuation periods or amplitudes of process parameters. Map function: transfer process parameter stream for shuffle and sort. Input: a batch of process parameters from SCADA, Key: machine identification (ID), value: time+ process parameter vector. Body: only transfer input stream. Output: a batch of process parameters to Reducer, Key: machine ID, value: time + process parameter vector. Reduce function: computing the average period and amplitude of fluctuation of process parameters Input: process parameters aggregated by machine ID and sorted by time, Key: machine ID, value: a matrix with time+ process parameter vector. Body: computing the average period and amplitude of fluctuation of process parameters group by key. Output: the average of fluctuation periods or amplitudes of process parameters to Job 2, Key: machine ID, value: the vector of average fluctuation periods and amplitudes.

Table 2 Job 2: computing the DPBFs of intercomparable machine pairs. Map function: matching intercomparable machine pairs. Input: the output from Job 1, Key: machine ID, value: the vector of average fluctuation periods and amplitudes. Body: get the other machine ID matched with input key. Output: matched machine pairs, Key: < machine ID1, machine ID2 > , value: the vector of average fluctuation periods and amplitudes.

4.2. An example for module implementation One module of AJAI approach can be split into one or more MapReduce jobs. Taking antigen feature presentation as examples, the module is split into 2 jobs as Tables 1 and 2.

Reduce function: computing the average period and amplitude of fluctuation of process parameters. Input: two vectors of average fluctuation periods and amplitudes aggregated by key. Key: < machine ID1, machine ID2 > , value: two vectors of average fluctuation periods and amplitudes. Body: computing DPBFs for each key. Output: the DPBFs of machine ID1 and machine ID2, to energy leakage isolation and database. Key: < machine ID1, machine ID2 > , value: DPBFs.

5. Application case Steam leakage in tyre vulcanization shops is a typical case of energy leakage in parallel-machine job shops. The tyre vulcanization shop floor we studied in [22] is still adopted in this case. It is located in Guangzhou China, has 104 double mould vulcanizers, and vulcanizes rubber tyres of passenger vehicles and trucks. A vulcanization job forms a green tyre to the desired shape and converts it into a strong and elastic material under an elevated temperature and pressure [59,60]. All vulcanization jobs are accomplished by vulcanizers automatically. Double 19

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

due to wear or breaking, which results in steam leakage. The leakage is hidden inside the recycling pipe; thus, it is not observable. It does not lead into failure of vulcanization but reduces the energy efficiency. Thus, it should be diagnosed as soon as possible for economical production. As shown in Fig. 7, the steam consumption is only measured at the shop level. It is difficult to be measured at the mould or vulcanizer level owing to the steam resistance of the steam meter, lack of installation space, mechanical vibration, etc. In the studied shop floor, the temperature of the upper hot plate, lower hot plate, and bladder, and the pressure of the inner steam and outer steam are collected by the process control system (PCS) every 2 s, and production data are also recorded in real time. These data are obtainable from a SCADA system, and the authors [22] have instantiated an energy loss presentation model and a predictive baseline model. In a steam trap, the discharge is automatically triggered when the condensate reaches a given weight. The cyclical discharges of a steam trap cause periodical fluctuation of steam humidity in the hot plates, leading to periodical temperature fluctuation of the hot plate. Fig. 8 shows the temperature curves of hot plates of a vulcanizer with a normal left steam trap and a leaking right steam trap. It is found that the temperature curves of the lower hot plates show obvious periodical fluctuation, but those of the upper hot plates do not have. The reason for this result is that the lower hot plates are at a lower location and nearer to the steam trap, and steam humidity in them is more sensitive to the accumulated condensate in the steam traps than in the upper hot plates. The leakage reduces the speed of steam condensation and thus, changes the fluctuation characteristics of temperatures. By comparing the temperature curves of the two moulds, it is found that the leakage reduced the frequency and amplitude of the temperature fluctuation. Therefore, only the sensitive temperature of the lower hot plate needs to be selected to present the antigen feature.

Fig. 6. Structure schematic of a vulcanizer mould.

mould vulcanizers are very prevalent in actual tyre plants owing to their high efficiency. A double mould vulcanizer has two steel moulds, which can be considered as two logical machines sharing one control system. In an antigen feature presentation model, the two moulds of a vulcanizer are naturally considered as an intercomparable machine pair. Fig. 6 illustrates the structure schematic of a vulcanizer mould. The major parts of the vulcanizer mould include the steel mould, upper hot plate, lower hot plate, and bladder. The outer steam pipe is connected to a steam channel of hot plates, and the inner steam, which has a higher pressure and temperature than the outer steam, is connected to the bladder. During vulcanization, a green tyre is fixed between the steel mould and the bladder, and then steam is injected into the steam channel and the bladder to heat and form the green tyre. In the bladder, the steam transfers heat to the green tyre through a capsule, and is vacuumed at the end of the process. In the steam room, the steam transfers heat to the green tyre through the upper and lower hot plates, and steam circulation is achieved by discharging condensate through a steam trap connected to the steam channel of the lower hot plate. The schematic of the steam pipe system in the vulcanization shop floor is shown in Fig. 7. Boiler steam is converted to inner steam in the steam converter valve, and then a part of the inner steam is converted into outer steam in the steam reducing valve. The main inner/outer steam pipe is connected to all vulcanizers through steam branches. It is an example of the hierarchical structure of the energy system shown in Fig. 3. The steam leakage during tyre vulcanization refers to the leakage of the steam trap. The steam trap may lose the function of steam resistance

6. Experimental results The proposed approach was developed using MATLAB 2013 Ra. Tests were performed on a PC Intel(R) Core(TM) i7 2.5 GHz with 8G of RAM, and data samples were selected from the application case mentioned in Section 5. The clustering performance of the AJ-aiNet algorithm and the diagnosis performance of the AJAI approach were evaluated. The steam consumption of the studied vulcanization shop was

Fig. 7. Schematic of a steam pipe system in the vulcanization shop floor. 20

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Fig. 8. Temperature curves of hot plates.

measured using a steam meter. The evaluation of the physical steam consumption, technology steam consumption, and steam loss baseline followed the models that the authors suggested in [22]. At the machine level, only the temperature of the lower hot plate was selected to present the antigen feature. The detection period was set as 1 h. Hourly energy loss, fluctuation characteristics of process parameters, and DPBF were calculated. A month’s (720 detection periods) data were extracted for experiments. To meet the requirement for learning and validation, the leakage events of steam traps were collected from the maintenance database of enterprise to label the state of detection period. The detection periods of steam traps were labelled with NORMAL, LEAK, and UNSURE. For each steam trap, the time span between two adjacent checks was labelled with NORMAL if no leakage was found, the time span from check time to repair time was labelled with LEAK, and the other time span was labelled with UNSURE. Then, detection periods contained in a NORMAL time span were labelled with NORMAL, those overlapping with a LEAK time span were labelled with LEAK, and the rest were labelled with UNSURE. For each vulcanizer, detection periods with two normal steam traps were labelled with NORMAL, those with one or two leaking steam traps were labelled with LEAK, and the rest were labelled with UNSURE. The leak detection periods were classified into LEFT LEAK (leak in left steam trap), RIGHT LEAK (leak in right steam trap), and BOTH LEAK (leak in two steam traps). At the shop level, the detection periods with only normal steam traps were labelled with NORMAL, those with at least one leaking steam trap were labelled with LEAK, and the rest were labelled with UNSURE. At vulcanizer level, the month’s sample includes 28212 NORMALs, 3106 UNSUREs, 1442 LEFT LEAKs, 1310 RIGHT LEAKs, and 0 BOTH LEAK. At shop level, the month’s sample includes 453 NORMALs, 165 UNSUREs and 102 LEAKs.

Fig. 9. Distribution of the evaluated sample with DPBFs.

Then the evaluated sample was split into learning sample and test sample evenly. The former was used to train an immune network, and the latter would be used to test the diagnosis performance, which will be discussed in the next subsection. The two samples were normalized in the range [0,1] to eliminate the impact of dimension range. The clustering stability and quality were adopted to evaluate the clustering performance of the AJ-aiNet algorithm, and the classical aiNet was used for comparison. The clustering correctness was judged by observing the consistency between the cluster centres of Abs and the learning sample. The clustering stability and quality were assessed against the correct rate of repeated tests and the cell size of the immune network. The parameters n, ξ, and σd were given empirical values of 4, 10, and 1, respectively. With σs = 0.01, σg was tried from 1 with a step value of 1, and each value was repeated 10 times. It is found that the correct rates of the immune network are 20%, 40%, 80%, and 100% with the σg = 1, 2, 3, and 4, respectively. Thus σg = 4 was supposed to be a suitable value for this learning sample. Then with σg = 4, σs was tried from 0.01 with a step value of 0.01; each clustering test was repeated 10 times. The correct rate and average

6.1. Clustering performance test and evaluation of AJ-aiNet The evaluated sample excluded UNSURE records, included LEAK records, and randomly selected 1400 NORMAL records to balance the sample size of state clusters. Fig. 9 shows the two-dimensional distribution of the evaluated sample. It is found that there exists a fuzzy boundary between cluster NORMAL and cluster LEAK. 21

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Table 3 Correct rate and average cell size of immune network. σs

0.01 0.02 0.03 0.04 0.05 0.06

aiNet

Table 4 Detection performance.

AJ-aiNet

Correct rate

Average size

Correct rate

Average size

20% 0% 0% 0% 0% 0

206 150 111 81 64 46

100% 100% 100% 100% 40% 0%

178 116 82 63 51 41

Systems

False rate

Missing rate

Average time (s)

Minimum aiNet Minimum AJ-aiNet Dual Activation

3.75% 3.31% 0.93%

6.51% 1.56% 1.62%

27.50 6.96 3.28

anti-jamming strategies. It is also found that there is no significant difference in missing rate between dual activation and minimum AJaiNet, but the false rate of dual activation is much smaller than minimum AJ-aiNet. This result shows that dual activation helps to suppress the false rate effectively. From the average detection times, it is found Minimum AJ-aiNet detects faster than Minimum aiNet, and Dual Activation detects faster than Minimum AJ-aiNet. The reasons are that the cell size of Minimum AJ-aiNet is smaller than Minimum aiNet, and Dual Activation detects only a little bit of data at machine level but the other two needs to detect all data. Dual Activation can meet the real time requirement from the detection time 3.28 s. In practice, the data at shop level is much less than the data at machine level, it can be supposed that detection time of Dual Activation is proportional to abnormal data at machine level. In this case, detection time of Dual Activation is approximately 3% (leaking rate) of minimum AJ-aiNet. In order to simulate an actual diagnosis, a continuous diagnosis was performed on the month’s data using dual activation. The sample contains 72 continuous LEAK records (72 h), which were confirmed as steam leakage from the steam traps of 5 vulcanizers. The part of diagnosis around the LEAK points, from the 200th hour to the 500th hour, is shown in Fig. 11. The energy loss detection identified a group of continuous UNSURE points near the start of a continuous LEAK section. Then, minimum AJ-aiNet isolated two leaking vulcanizers by determining continuous UNSURE detection periods. These alarms should not be simply considered as false diagnosis. By contrast, they should be thought of as an early alarm for the leaks at a low level because the UNSURE points are coherent. In this sense, the results show the AJAI can achieve energy diagnoses in parallel-machine job shops effectively.

cell size of the immune network are presented in Table 3. It is found that aiNet has a low probability of obtaining a correct immune network even with very small σs, but AJ-aiNet can stably output correct immune networks with σs from 0.01 to 0.04. Meanwhile, the results in Table 3 show that under the same σs, the average size of the immune network acquired by AJ-aiNet is less than that of the classical aiNet. The minimum correct immune network acquired by AJ-aiNet and classical aiNet were named minimum AJ-aiNet and minimum aiNet, respectively, of which structures are shown in Fig. 10. It is found that the cell size of the former is approximately 1/3 of the latter. This shows that AJ-aiNet achieves a better quality of clustering than aiNet by adding anti-jamming strategies. 6.2. Diagnosis performance test of AJAI approach A steam loss baseline model was obtained using the approach suggested in [22]. It is combined with minimum AJ-aiNet into an instance of dual activation. Minimum AJ-aiNet and minimum aiNet were used as two instances of single activation. With the three instances, the diagnosis experiments were performed on the test sample generated in Section 6.1. False rate and missing rate [61,62] were used to evaluate the diagnosis performance. False detection refers to a false LEAK detection, missing detection refers to a false NORMAL detection, false rate is the rate of false LEAK detections to true NORMAL detections, and missing rate is the rate of false NORMAL detections to true LEAK detections. Average detection time was used to evaluate the time performance. The diagnosis results are presented in Table 4. It is found that there is no significant difference in false rate between minimum AJ-aiNet and minimum aiNet, but the missing rate of minimum AJ-aiNet is much smaller than that of minimum aiNet. This result shows that minimum AJ-aiNet reduced the missing rate by adding

7. Discussion In parallel-machine job shops, stochastic parallel jobs and a fluctuating energy supply lead to random fluctuation of parameters and produce serious jamming information for energy leakage diagnosis. The experimental results showed that the AJAI approach overcame the

Fig. 10. Immune network structure. 22

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

Fig. 11. Continuous detection test of shop detection subsystem.

overcome a jamming environment with limited energy measurements, stochastic parallel jobs, and a fluctuating energy supply, the proposed approach used anti-jamming strategies, namely a danger-model-inspired framework, an anti-jamming antigen feature, and an anti-jamming aiNet (AJ-aiNet) algorithm. The approach was developed using MATLAB 2013 Ra, and as an application case, the steam leakage of steam traps in a tyre vulcanization shop floor was diagnosed. The experimental results showed that the approach achieved isolation of leaking vulcanizers in the absence of energy consumption measurements at the machine level and attained a 0.93% false detection rate and 1.62% missing detection rate. The results show the effectiveness of the anti-jamming strategies. It is found that the transient parameter, average parameter, and average fluctuation of parameter were unable to separate the leaking and normal process behaviours. However, DPBF can separate leaking and normal process behaviours effectively, AJ-aiNet can mitigate disturbance and achieve clustering of DPBF samples correctly, and dual activation can suppress false detections and improve the time efficiency of diagnosis significantly. In comparison, the classical aiNet obtained a low probability of clustering DPBF samples correctly, and the approach with single activation obtained a false detection rate that is 2.34% higher than the proposed approach with dual activation. Compared with the approaches based on fluid mechanics and those based on energy consumption models, the proposed approach combines the isolation capability of the former and the detection capability of the latter, and overcomes the jamming environment by using anti-jamming strategies. To continue this research, we suggest to perform case study on the implementation of this approach on the basis of computer cluster and concurrent streaming processing, and to embed the intelligence of this approach at shop floor end points using edge computing technology.

jamming environment and achieved an energy leakage diagnosis with acceptable diagnosis performance. As a new feature, DPBF mitigates the impact of random fluctuation caused by jamming factors. Fig. 9 shows that DPBF make boundaries between NORMAL and LEAK undergoing statistical and comparative analysis. Fig. 10(b) shows that AJ-aiNet generated a correct immune network from the sample. Tables 3 and 4 show that AJ-aiNet achieved a better stability and quality of clustering than the classical aiNet. Meanwhile, these results showed that the classical aiNet has a low probability of generating a correct immune network from the sample with a few jamming points. These findings show that adding antibody ageing and antigen killing strategies allows isolating a leak antigen feature from a normal antigen feature. Table 4 indicates that the dual activation helps to suppress false rates and improve effectively. It should be attributed to the presentation of energy loss. In [22], the authors have validated that energy loss is a very sensitive index to energy leakage. The time performance in Table 4 showed that dual activation also helps to improve the time efficiency of diagnosis, because only the information at the shop level needs to be monitored in normal condition. Moreover, the implementation on the basis of computer cluster and concurrent streaming processing supports the real time performance of energy leakage diagnosis in large plants. To a certain degree, the danger-model-inspired framework combined the approaches based on fluid mechanics and those based on energy consumption models. Compared with the former, DPBF and AJaiNet make this approach suitable to a jamming environment. Moreover, the evaluation of DPBF depends on the statistics of a parameter time series instead of an analysis on high real-time parameters, such as negative pressure wave, and electromagnetic wave in [9,12,13], which are difficult to acquire in an industrial environment. Compared with the latter such as [15–17], this approach depends on process parameters instead of energy measurements to isolate leaking machines, and is suitable to energy measurements only at the shop level. Moreover, the suggested energy loss is much more sensitive to energy leakage than energy consumption. The proposed approach achieved the energy leakage diagnosis in the studied vulcanization shop floor, which is a typical parallel-machine job shop. It is noted that building an energy loss presentation model may be a hindrance for extensive applications of this approach, because it needs to consider the complex energy mechanism of a job process when evaluating physical and technology energy consumption. This study ignored the complexity by quoting the author’s previous work in [22]. In the studied case, only one parameter is adopted to evaluate DPBF and isolate the leaking machine; the parameters may be more complicated in other applications. This approach is suggested for parallel-machine job shops with limited energy measurements. If energy consumption measurements at the machine level are available, the energy loss may be more effective for isolating the leaking machine than process parameters.

Acknowledgements The authors would like to thank the information management group of the vulcanization shop floor studied, who provided the experimental data and access to the database system. This work was supported by the National Natural Science Foundation of China (NSFC) (Grant No: 61702119 and 71401044). References [1] N. Bayar, S. Darmoul, S. Hajri-Gabouj, H. Pierreval, C. Zhansheng, Fault detection, diagnosis and recovery using Artificial Immune Systems: a review, Eng. Appl. Artif. Intell. 46 (2015) 43–57. [2] Y. Zhang, S. Ren, Y. Liu, T. Sakao, D. Huisinghe, A framework for Big Data driven product lifecycle management, J. Clean. Prod. 159 (15) (2017) 229–240. [3] Y. Zhang, J. Wang, Y. Liu, Game theory based real-time multi-objective flexible job shop scheduling considering environmental impact, J. Clean. Prod. 167 (20) (2017) 665–679. [4] R. Saidur, S. Mekhilef, Energy use, energy savings and emission analysis in the Malaysian rubber producing industries, Appl. Energ. 87 (2010) 2746–2758. [5] C.-W. Park, K.-S. Kwon, W.-B. Kim, B.-K. Min, Energy consumption reduction technology in manufacturing –a selective review of policies standards, and research, Int. J. Precis. Eng. Manuf. 10 (5) (2009) 151–173. [6] Y. Zhang, G. Zhang, T. Qu, Y. Liu, R.Y. Zhong, Analytical target cascading for optimal configuration of cloud manufacturing services, J. Clean. Prod. 151 (10) (2017) 330–343.

8. Conclusion This study presented an anti-jamming artificial immune approach to achieve energy leakage diagnoses in parallel-machine job shops. To 23

Computers in Industry 101 (2018) 13–24

J. Guo, H. Yang

[7] R.-E. Precup, P. Angelov, B.S.J. Costa, M. Sayed-Mouchaweh, An overview on fault diagnosis and nature-inspired optimal control of industrial process applications, Comput. Ind. 74 (11) (2015) 75–94. [8] W. Wang, H. Yang, Y. Zhang, J. Xu, IoT-enabled real-time energy efficiency optimisation method for energy-intensive manufacturing enterprise, Int. J. Comp. Integr. Manuf. (2017) 1–18. [9] Y. Kim, S.J. Lee, T. Park, G. Lee, J.C. Suh, J.M. Lee, Robust leak detection and its localization using interval estimation forwater distribution network, Comp. Chem. Eng. 92 (9) (2016) 1–17. [10] P. Ostapkowicz, Leak detection in liquid transmission pipelines using simplified pressure analysis techniques employing a minimum of standard and non-standard measuring devices, Eng. Struct. 113 (7) (2016) 194–205. [11] S. Tian, J. Du, S. Shao, H. Xu, C. Tian, A study on a real-time leak detection method for pressurized liquid refrigerant pipeline based on pressure and flow rate, Appl. Therm. Eng. 95 (2) (2016) 462–470. [12] Y. Zhang, S. Chen, J. Li, S. Jin, Leak detection monitoring system of long distance oil pipeline based on dynamic pressure transmitter, Measurement 49 (3) (2014) 382–389. [13] A. Cataldo, E.D. Benedetto, G. Cannazza, A. Masciullo, N. Giaquinto, G.M. D'Aucelli, N. Costantino, A.D. Leo, M. Miraglia, Recent advances in the TDR-based leak detection system for pipeline inspection, Measurement 98 (2) (2016) 347–354. [14] T. Zhang, Y. Tan, X. Zhang, J. Zhao, A novel hybrid technique for leak detection and location in straight pipelines, J. Loss Prev. Process Ind. 35 (5) (2015) 157–168. [15] J.R. Duflou, J.W. Sutherland, D. Dornfeld, C. Herrmann, J. Jeswiet, S. Kara, M. Hauschild, K. Kellens, Towards energy and resource efficient manufacturing: a processes and systems approach, CIRP Ann. – Manuf. Technol. 61 (2) (2012) 587–609. [16] S. Rahimifard, Y. Seow, T. Childs, Minimising Embodied Product Energy to support energy efficient manufacturing, CIRP Ann. – Manuf. Technol. 59 (1) (2010) 25–28. [17] Q. Wang, F. Liu, C. Li, An integrated method for assessing the energy efficiency of machining workshop, J. Clean. Prod. 52 (8) (2013) 122–133. [18] Y. Li, Y. Hea, Y. Wang, P. Yan, X. Liu, A framework for characterising energy consumption of machining manufacturing systems, Int. J. Prod. Res. 52 (2) (2014) 314–325. [19] Z.M. Bi, L. Wang, Optimization of machining processes from the perspective of energy consumption: a case study, J. Manuf. Syst. 31 (4) (2012) 420–428. [20] F.E. Pfefferkorn, S. Lei, Y. Jeon, G. Haddad, A metric for defining the energy efficiency of thermally assisted machining, Int. J. Mach. Tools Manuf. 49 (5) (2009) 357–365. [21] H. Yang, J. Guo, G. Liu, Energy anomaly detection in tire curing by using data integration and forecasting techniques, J. Sci. Ind. Res. 71 (6) (2012) 385–391. [22] J. Guo, H. Yang, A fault detection method for heat loss in a tyre vulcanization workshop using a dynamic energy consumption model and predictive baselines, Appl. Therm. Eng. 90 (11) (2015) 711–721. [23] W. Yaïci, E. Entchev, Performance prediction of a solar thermal energy system using artificial neural networks, Appl. Therm. Eng. 73 (1) (2014) 1346–1357. [24] L. Wang, Z. Liu, C.L.P. Chen, Y. Zhang, S. Lee, Support Vector Machine based optimal control for minimizing energy consumption of biped walking motions, Int. J. Precis. Eng. Manuf. 13 (11) (2012) 1975–1981. [25] B. Widarsson, E. Dotzauer, Bayesian network based early warning for leakage in recovery boilers, Appl. Therm. Eng. 28 (7) (2008) 754–760. [26] Y. Zhao, S. Wang, F. Xiao, A statistical fault detection and diagnosis method for centrifugal chillers based on exponentially-weighted moving average control charts and support vector regression, Appl. Therm. Eng. 51 (1–2) (2013) 560–572. [27] J. L.N. d. Castro, Timmis Artificial immune systems as a novel soft computing paradigm, Soft Comput. 7 (8) (2003) 526–544. [28] C.A. Laurentys, R.M. Palhares, W.M. Caminhas, Design of an artificial immune system based on Danger Model for fault detection, Expert Syst. Appl. 37 (7) (2010) 5145–5152. [29] U. Aickelin, S. Cayzer, The danger theory and its applications to artificial immune systems, in, Proceedings of the first international conference on artificial immune systems, Hewlett-Packard Company, Canterbury UK, 2002, pp. 141–148. [30] C.A. Laurentys, G. Ronacher, R.M. Palhares, W.M. Caminhas, Design of an artificial immune system for fault detection: a negative selection approach, Expert Syst. Appl. 37 (7) (2010) 5507–5513. [31] N. Bayar, S. Darmoul, S. Hajri-Gabouj, H. Pierreval, Z. Cai, Using immune designed ontologies to monitor disruptions in manufacturing systems, Comput. Ind. 81 (9) (2016) 67–81. [32] Z. Chilengue, J.A. Denteb, P.J.C. Branco, An artificial immune system approach for fault detection in the stator and rotor circuits of induction machines, Electr. Power Syst. Res. 81 (1) (2011) 158–169. [33] F.P.A. Lima, M.L.M. Lopes, A.D.P. Lotufo, C.R. Minussi, An artificial immune system with continuous-learning for voltage disturbance diagnosis in electrical distribution systems, Expert Syst. Appl. 56 (9) (2016) 131–142. [34] G.C. Silva, R.M. Palhares, W.M. Caminhas, Immune inspired Fault Detection and

[35]

[36]

[37]

[38] [39] [40] [41] [42] [43] [44] [45] [46]

[47]

[48]

[49] [50] [51]

[52]

[53] [54]

[55] [56]

[57]

[58]

[59] [60]

[61]

[62]

24

Diagnosis: a fuzzy-based approach of the negative selection algorithm and participatory clustering, Expert Syst. Appl. 39 (11) (2012) 12474–12486. D.A. Azzawi, H. Moncayo, M.G. Perhinschi, A. Perez, A. Togayev, Comparison of immunity-based schemes for aircraft failure detection and identification, Eng. Appl. Artif. Intell. 52 (C) (2016) 181–193. S. Forrest, A.S. Perelson, L. Allen, R. Cherukuri, Self-Nonself discrimination in a computer, Proceedings of 1994 IEEE Symposium on Research in Security and Privacy Oakland CA (1994) 202–212. R. Deaton, M. Garzon, J.A. Rose, R.C. Murphy, S.E. Stevens, D.R. Francheschetti, A DNA Based Artificial Immune System for Self-nonself Discrimination, IEEE International Conference on Systems, Man, and Cybernetics Computational Cybernetics and Simulation, Orlando, FL, 1997, pp. 862–866. P. Matzinger, Tolerance danger and the extended family, Annu. Rev. Immunol. 12 (1) (1994) 991–1045. P. Matzinger, An innate sense of danger, Semin. Immunol. 10 (5) (1998) 399–415. P. Matzinger, The danger model: a renewed sense of self, Science 296 (5566) (2002) 301–305. H. Yang, J. Guo, F. Deng, Collaborative RFID intrusion detection with an artificial immune system, J. Intell. Inf. 36 (1) (2011) 1–26. N.K. Jerne, Towards a network theory of the immune system, Ann. Immunol. 125C (1–2) (1974) 373–389. G. Ada, G. Nossal, The clonal selection theory, Sci. Am. 257 (2) (1987) 50–57. J.D. Farmer, N.H. Packeard, The immune system, adaptation, and machine learning, Physica 22 (1-3) (1986) 187–204. J. Timmis, M. Neal, A resource limited artificial immune system for data analysis, Knowl. Based Syst. 14 (3–4) (2001) 121–130. L.N. d. Castro, F.J.V. Zuben, An evolutionary network for data clustering, Proceedings of 6th Brazilian Symposium on Neural Networks IEEE Computer Society Press, Rio De Janeiro, RJ, Brazil (2000) 84–89. L.N.d. Castro, J. Timmis, An Artificial Immune Network for Multimodal Optimisation, In, In 2002 Congress on Evolutionary Computation, Part of the 2002 IEEE World Congress on Computational Intelligence, IEEE, Honolulu, Hawaii, USA, 2002, pp. 182–196. Z. Li, C. He, H.-Z. Tan, AINet-SL: Artificial immune network with social learning and its application in FIR filter designing, Appl. Soft Comput. 13 (12) (2013) 4557–4569. W. Pang, G.M. Coghill, QML-AiNet: an immune network approach to learning qualitative differential equation models, Appl. Soft Comput. 27 (2) (2015) 148–157. A.M. Acilar, A. Arslan, A collaborative filtering method based on artificial immune network, Expert Syst. Appl. 36 (4) (2009) 8324–8332. A.M. Acilar, A. Arslan, A novel approach for designing adaptive fuzzy classifiers based on the combination of an artificial immune network and a memetic algorithm, Inf. Sci. 264 (1) (2014) 158–181. R.J. Kuo, C.M. Pai, R.H. Lin, H.C. Chu, The integration of association rule mining and artificial immune network for supplier selection and order quantity allocation, Appl. Math. Comput. 250 (2015) 958–972. H. Xiong, C. Sun, Artificial immune network classification algorithm for fault diagnosis of power transformer, Autom. Electr. Power Syst. 22 (2) (2007) 930–935. F.Z. Wang, S.M. Shao, P.F. Dong, Research on transformer fault diagnosis method based on artificial immune network and fuzzy C-Means clustering algorithm, Appl. Mech. Mater. 574 (2014) 468–473. W.K. Jiang, Y.J. Chen, J. Zhang, A fault diagnosis method based on artificial immune network, Appl. Mech. Mater. 385–386 (2013) 658–662. Z. Li, J. Li, S. Wang, C. Tang, M. Xia, Comparative performance analysis of various artificial immune networks applied to RFID reader-to-reader collision avoidance, Appl. Soft Comput. 32 (7) (2015) 553–567. S. Raileanu, F. Anton, A. Iatan, T. Borangiu, S. Anton, O. Morariu, Resource scheduling based on energy consumption for sustainable manufacturing, J. Intell. Manuf. 28 (7) (2017) 1519–1530. S. Raileanu, T. Borangiu, F. Anton, Establishing optimal energy working parameters for a robotized manufacturing cell, in: T. Borangiu (Ed.), Proceedings of the 24th International Conference on Robotics in Alpe-Adria-Danube Region (RAAD), Springer International Publishing Switzerland, Bucharest, Romania, 2016, pp. 13–25. X. Yan, A numerical modeling of dynamic curing process of tire by finite element, Polym. J. 39 (10) (2007) 1001–1010. M. Pandy, R.N. Pate, S. Amarnath, Determination of time delay and rate of temperature change during tyre curing (vulcanizing) cycle, Procedia Eng. 51 (2013) 828–833. P. Saurabh, B. Verma, An efficient proactive artificial immune system based anomaly detection and prevention system, Expert Syst. Appl. 60 (10) (2016) 311–320. T.S. Powers, J. He, A hybrid artificial immune system and self organising map for network intrusion detection, Inf. Sci. 178 (15) (2008) 3024–3042.