A smart multisensor system for volcanic ash fall-out monitoring

Sensors and Actuators A 202 (2013) 13–22

Contents lists available at ScienceDirect

Sensors and Actuators A: Physical journal homepage: www.elsevier.com/locate/sna

A smart multisensor system for volcanic ash fall-out monitoring B. Andò ∗ , S. Baglio, V. Marletta, S. Medico Dipartimento di Ingegneria Elettrica, Elettronica e Informatica (DIEEI), University of Catania, V.le A. Doria 6, 95125 Catania, Italy

a r t i c l e

i n f o

Article history: Available online 27 March 2013 Keywords: Volcanic ash Air transport safety Multisensor system Particle granulometry ROC analysis

a b s t r a c t Ash fall-out following the explosive activity of volcanoes represents a factor of risk for people and a serious hazard for air traffic, especially if airports are close to the volcanic area. This is the case of the Fontanarossa airport located in Catania, in the south of Italy, close to the Etna volcano. Researchers at DIEEI are facing the challenge of developing a low-cost smart multisensor system to monitor the ash fall-out phenomenon by measuring the average granulometry of ash particles and the ash flow rate. Moreover, the system must be selective in respect to volcanic ash against others sediments such as dust, sand or soil. This paper is particularly focused on the methodology to be adopted for ash granulometry detection. The main idea is to use a piezoelectric transducer to convert ash impacts into electrical signals, which should provide information about the ash granulometry. Experimental results showing the suitability of the proposed approach are presented. Moreover, Receiver Operating Characteristic (ROC) analysis has been proposed as a theoretical support to properly implement the threshold mechanism aimed at ash granulometry classification. © 2013 Published by Elsevier B.V.

1. Introduction The atmospheric dispersion of ash released by the explosive activity of an active volcano is a relevant risk factor, such as hazard posed by the Etna volcano to the eastern area of Sicily in the south of Italy. Ash clouds cause extensive damage to roads, sanitation systems [1], agriculture [2], health [3] and the daily activities of people living on the slopes of the volcano, and also forms a substantial hazard to air traffic (civil, commercial and military) [4]. Monitoring of ash fall-out is of great interest to meet logistical needs, as well as to achieve a better understanding of such phenomena and their ruling mechanisms. Due to the worldwide increase in volcanic explosion phenomena with consequent atmospheric ash dispersion, and in particular after the eruption of the Icelandic Eyafjallajokull volcano in 2010, the need of regulation for the safety of air transport has arisen. In response to this need, the International Civil Aviation Organization (ICAO) has issued guidance on managing flight operations into, or near, areas of known or forecast volcanic clouds [5]. During the eruptions of 2001 and 2002, the Fontanarossa airport in Catania (at that time the third major airport in Italy in terms of numbers of passengers) was repeatedly declared

DOI of original article: http://dx.doi.org/10.1016/j.sna.2013.01.056. ∗ Corresponding author. Tel.: +39 095 7382601; fax: +39 095 7387965. E-mail addresses: [email protected] (B. Andò), [email protected] (S. Baglio), [email protected] (V. Marletta), [email protected] (S. Medico). 0924-4247/$ – see front matter © 2013 Published by Elsevier B.V. http://dx.doi.org/10.1016/j.sna.2013.03.027

inappropriate for take-offs and landings due to volcanic ash fall-out, creating great inconvenience to passengers along with financial loss for airlines and airport operators. During such crises, the decision to open or close the airport by the Italian Civil Aviation Authority resulted from subjective assessments, and therefore contained high levels of uncertainty and risk. As a consequence a scientific and technical commission has been established with the aim of regulating flight operations at airports in the presence of volcanic ash [6]. The National Institute of Geophysics and Volcanology (Istituto Nazionale di Geofisica e Vulcanologia, INGV) was designated a primary source of information about eruptions that produce volcanic clouds. To cope with this service INGV is equipped with systems for monitoring ash clouds and forecasting their space-time evolution [7]. The latter provides decision support for aeronautical authorities in order to significantly reduce the factors of unpredictability, and therefore the impact of this phenomenon on airport infrastructure and flight operations. Traditional approaches for the monitoring of volcanic ash employ high-cost instrumentation, typically based on satellites [8] or X-Band dual-polarization radars [9]. Reliable instruments for ash granulometry classification, generally based on high-cost infrared cameras [10] and providing information on ash suspended in atmosphere, have been developed and deployed. These solutions allow discrete measurements at a restricted number of monitoring stations, thus guaranteeing a low spatial resolution. Moreover, such systems are difficult to install and maintain and are often used to perform spot measurements.

14

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

Fig. 1. (a) Schematisation of the multisensor node; (b) schematisation of the on-board electronics.

In order to properly confront surveillance activities of volcanic clouds and ash fall-out the need for a distributed network of monitoring stations has emerged. Researchers at DIEEI in Catania are developing an early warning distributed network of self-powered, low-cost, ␮-controller based multisensor nodes for the monitoring of ash fall-out in the area spreading from the main craters of the Etna volcano to Fontanarossa airport [11]. The sensor network will provide information on ash fall-out (presence, average granulometry, ash flow rate) that is useful to predict the time-space evolution of such phenomenon. The idea behind the early warning approach is the possibility to provide fast and well spatially distributed information on incoming phenomena at the expense of high accuracy furnished by high-cost instrumentation. Furthermore, in cases of specific need, accurate measurements can be performed by dedicated high-cost instrumentation at critical sites evidenced by the early warning system. Fundamentally, it can be confirmed that this approach in the field of volcanic ash monitoring is innovative, in particular when taking into account its capability to provide experts with a time continuous awareness of the ash fall-out phenomenon with a high degree of spatial resolution. The latter, along with its element of low-cost, are the main advantages of the proposed solution. The research activity is conducted within the SECESTA project, the Italian acronym for “A sensor network for the monitoring of volcanic ash fall-out for the safety of air transport”. The project is funded under the POR FESR Sicily 2007–2013 program and it exploits the synergic operations of research institutes (among these DIEEI of Catania) and small-medium sized enterprises. A number of monitoring nodes will be installed along the direction between craters and the airport to build a Wireless Sensor Networks aimed to monitor volcanic ash fall-out. The WSN will be developed, by a small-medium sized enterprise with the collaboration of the National Institute of Volcanology, Catania Section. This paper focuses on the sensing methodology adopted for ash granulometry detection. The main idea is to use a piezoelectric transducer to convert ash impacts into electrical signals whose properties are strictly dependent on ash granulometry. Apart from the already-mentioned advantages of the general approach followed by the overall monitoring system, the sensing solution proposed for ash granulometry classification is truly lowcost and it offers high reliability due to the sensing architecture adopted.

Concerning the importance of the monitoring of the ash granulometry, it must be remarked that an awareness of ash dimensions and its distribution throughout the area of interest are fundamental information in the process of forecasting the time-space evolution of ash clouds and fall-out. Forecast mechanisms are usually implemented by models taking into account both the meteorological quantities (e.g. wind speed and direction) and the characteristics of ash fall-out (flow and granulometry) [7]. In the next section an overview of the multisensor node is given. A detailed description of the sensing strategy developed for ash granulometry estimation is provided in Section 3 along with theoretical considerations and experimental results. The methodology for particle classification based on Receiver Operating Characteristic (ROC) analysis is presented in Section 4, while concluding remarks are given in Section 5.

2. Multisensor node Although this paper mainly addresses solutions adopted for the classification of ash particles, in this section basic information on the whole structure of the developed multisensor node are given for the sake of completeness. The goal of the multisensor architecture developed at DIEEI laboratories in Catania is the implementation of a self-powered, ␮-controller platform for the monitoring (presence, granulometry, flow rate) of ash fall-out in areas extending from volcano craters to airports. Schematisations of the multisensor node and the on-board electronics are shown in Fig. 1, while a true image of the laboratory prototype developed is displayed in Fig. 2. Basically, the system consists of a funnel-shaped structure to convey falling ash to a tank. A dedicated array of coupled infrared (IR) diode-phototransistors allows for the monitoring of ash levels in the tank in order to indirectly estimate the ash flow-rate. A piezoelectric sensor is placed within the funnel-shaped structure with the purpose of converting the impacts of ash grains into electrical signals. As outlined in Fig. 3, the piezoelectric transducer is placed with a slope of 45◦ with respect to the direction of fall (vertical axis) to reduce multiple impacts of the same ash particle due to undesired bounces. More details about the proposed sensing solution and the conditioning electronic will be given in Section 3.

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

15

hostile agents (water, dust, sand, soil or bird droppings) that could damage the system. The opening of this shutter can be remotely controlled by operators, or automatically triggered when explosive phenomenon are observed. To this end the possibility of using the same piezoelectric transducer as used for ash granulometry classification could be investigated for the detection of explosive activities. Although the developed system is a lab-scale prototype, the lowcost features are related to the sensing technology adopted, as is easily demonstrated from a rough comparison with other monitoring systems cited in the SOTA review [8–10]. The engineered system can be shrunk to reduce cost, and mass production can also contribute to this aspect. In any case, the system does not require complicate installation efforts, and it has been developed to be robust against environmental factors. For cost comparisons it could be taken into account that (apart from expensive non real-time solutions, e.g. dedicated satellites) an instrument based on X-band radar costs tens of thousands of dollars, an IR camera costs several thousand dollar plus the cost of the supporting network to transmit the data, while mass produced sensors, of the type described in this paper, would cost on the order of 300–700 dollars apiece including the necessary data network.

Fig. 2. Real image of the multisensor node prototype developed at DIEEI Laboratories with the purpose of investigating suitable sensing solutions for the monitoring of ash fall-out phenomenon.

A shutter system, placed at the bottom of the tank and driven by a digital driver, has been developed to empty the tank when it is full or before the start of a new measurement. To exploit the magnetic properties of volcanic ash, a digital magnetometer is used to discriminate volcanic ash from other sediments [12]. Wind and rain sensors are also used to provide information on meteorological quantities. The latter are used in models in order to predict the space-time evolution of the observed phenomenon. Signals furnished by sensors are preprocessed by a ␮-controller unit as shown in Fig. 1(b). In particular a Droids Multi Interface Board (MuIn) equipped with a PIC18F2520 Microchip running at 40 MHz has been employed due to its versatility. Data are then transferred by a wireless transmission protocol (IEEE 802.15.4) to a dedicated PC station. Nodes are supplied by the supply network, if available, or by alternative energy sources (photovoltaic and/or aeolic). Finally, a further shutter system is placed on top of the ash collector with the purpose of shielding the system against

Fig. 3. Schematisation of the interaction forces between the volcanic ash particle and the piezoelectric transducer.

3. Sensor developed for ash granulometry classification As previously mentioned, the main target of the presented activity is the development of cheap solutions for ash granulometry classification to be integrated in low-cost distributed sensor networks. The proposed methodology is based on the use of a piezoelectric transducer to convert ash impacts into electrical signals, which will provide information about ash granulometry. The underpinning idea is to exploit the relationship between particle granulometry and the physical force provided by their impact on the sensor. It is important to note that the ash density (which fixes the relationship between granulometry and the mass of each particle) is an intrinsic characteristic of the volcano, and it can be considered as a known quantity. The use of a piezoelectric sensing strategy, with respect to other sensing methodologies (e.g. exploiting capacitive or inductive sensors), provides a direct measure. It avoids the indirect estimation of the average particle size, e.g. obtained by measuring the volume of collected ash. Conversely, the proposed piezoelectric approach can be used to perform the granulometry estimation of each particle, in order to achieve the granulometry classification of volcanic ash over the observation time interval. A typical piezoelectric response to a single ash particle impact is shown in Fig. 4. As it can be observed, the device starts oscillating at its characteristic natural frequency fn following a typical decreasing exponential envelope. As will be demonstrated in the follow-up, the amplitude of the sensor’s response is strictly related to the particle mass (impact force) or to its size (for a known particle density). From a statistical point of view, the probability of concurrent impacts on the piezoelectric is very low, in any case consecutive impacts of particles on the piezoelectric will cause consecutive impulsive solicitations and then consecutive but observable signature in the piezoelectric response. The analysis of these signatures can be exploited for the sake of granulometry classification. The probability of sediments laying on the piezoelectric sensor is very low due to the intrinsic topology of the tilted piezoelectric sensing tool. In any case, the device’s responsivity (in terms of voltage impulse amplitude as a consequence of the ash impact force) can be considered independent on the sensor mass (which can change for the effect of sediments on the sensor surface)

16

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

Fig. 4. (a), (c) and (e) Experimental impulse responses of the piezoelectric sensor and fitting obtained through model (2) for the three ash sizes (small, medium and large). A natural frequency of 3.49 kHz has been estimated for the piezoelectric sensor. (b), (d) and (f) Details of the device impulse response for the three ash classes.

as stated by model (2–3) presented in Section 3.1. Moreover, masking effects of the sensing surface due to deposited sediments can be considered negligible due to the high ratio between the sensor section and the particle dimension. On the basis of such considerations, the classification strategy implemented through the developed system is based on the effect

of ash impacts on the amplitude of the piezoelectric response. The effect of sediments on the natural frequency does not affect the classification methodology. In the following section, the modelling, design, synthesis and characterisation of the piezoelectric readout strategy developed will be discussed.

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

3.1. Modelling and design of ash granulometry sensor

The minimisation method adopted is the Nelder–Mead optimisation algorithm [17] with the following function J:

In this section an analytical model of the granulometry sensor is presented. In particular, a model for the collision force produced by ash particles impacting on the piezoelectric is discussed together with a model describing the impulse response of the piezoelectric transducer. With reference to Fig. 3, ash grains impact on the surface of the piezoelectric transducer with a force F thus generating free charges and consequently a voltage signal. The interaction force F can be resolved into two forces: a tangential component Ft , whose effect on the piezoelectric sensor can be assumed to be negligible, and a normal (to the slope) component Fp . Moreover, assuming both ash particles and the piezoelectric transducer are linearly elastic solids and modelling the ash particle as a sphere of radius RA , the perpendicular force Fp can be determined by the Hertz law [13,14]:

3/5

Fp = 1.2644A



1 kP + kA

2/5

6/5

RA2 vA

(1)

where A = 2 kg/m3 , kA and A represent the density, the stiffness and the terminal settling velocity [15] of the ash particle, respectively, and kp represents the stiffness of the piezoelectric transducer. The impact of ash particles produces an impulsive stress on the piezoelectric transducer which will start to oscillate at its natural frequency fn , thus producing an oscillating voltage with the same frequency and whose amplitude decreases with an exponential rate as shown in Fig. 4(a). The predicted trend for the impulse response is given by: Vout = A ·

e−ωn t



1 − 2

sin (2fn t)

(2)

where  is the system damping, fn is the natural frequency and A depends on the impulsive stress generated by the Fp force by the relationship [16]: A=

g33 Fp d S

17

(3)

where g33 = 22e−3 Vm/N is the piezoelectric coefficient in the stress direction, d = 0.23e−3 m and S = 4.155e−4 m2 are the thickness and the surface area of the transducer, respectively. Three classes of particle have been considered within this paper: small, medium and large, with an average radius of around 0.4 mm, 1 mm and 2 mm, respectively. As clearly evidenced in Fig. 4, the amplitude of the response generated by the piezoelectric sensor depends on the ash granulometry. This relationship is also evincible by Eq. (2), which states that the voltage generated by the piezoelectric sensor depends on the ash granulometry through the term A, which is related to the impact force and hence to the particle radius. In fact, the impact (and consequently the A parameter) will also be related to the velocity and the density of the falling particles. Nevertheless, ash density is usually a characteristic of the volcanic area while, close to the ground, ash particles reach a constant terminal velocity [15]. The latter property is strictly related to the ash dimensions, and hence will not randomly affect the classification procedure. By fitting model (2) to the observed responses for different particle sizes (small, medium and large) the following parameters have been estimated:  = 0.0098, fn = 3.49 kHz and As = 2.30 V, Am = 3.20 V, Ab = 11.5 V for small, medium and large particles, respectively.

J=

3  size=1

⎛ ⎝

˙iN (Vout − Vout, pred )2 N

⎞

size :

⎠

1 = small 2 = medium

(4)

3 = large

Function J is the sum of the root mean square of the residuals between the observed responses Vout and predicted responses Vout, pred for the three particle size classes considered. Fig. 4(a), (c) and (e) shows the fitting between the predicted and experimental impulse responses for the three ash sizes, while details are presented in Fig. 4(b), (d) and (f). Eqs. (1)–(3) have been used for the design of the sensing system, including the conditioning electronics. It must be born in mind that volcanic ash granulometry depends on the monitoring site, wind conditions and is strictly related to the specific volcano. With the general view of using the proposed approach at different sites and for different volcanoes, model (1–3) becomes a fundamental tool to predict the system’s behaviour under different operating conditions and to develop a suitable design for the sensing system. 3.2. The real device: synthesis and characterisation The low-cost piezoelectric transducer adopted in the granulometry sensor prototype is the 7BB-35-3L0 by Murata, which has a round piezoelectric diaphragm of diameter 35 mm. Dedicated conditioning electronics consisting of a charge amplifier followed by an inverting amplifier has been developed. A schematic of the electronics is shown in Fig. 5(a), where Ceq = 14 nF and Req = 67 k are the equivalent components of the piezoelectric transducer. The charge amplifier furnishes a voltage signal proportional to the charge generated by the piezoelectric transducer. Values of the feedback resistance Rf = 1.2 M and capacitance Cf = 22 nF of the charge amplifier have been chosen to fix a low cut-off frequency, ftchA = 6 Hz, which is compatible with the specific application, as shown in Fig. 5(b). The cut-off frequency, ftchA , was defined according to experts in order to make the piezoelectric system also sensitive to seismic activity, with the purpose of implementing a trigger system for the automatic opening of the top shutter. It must be observed that volcanic or seismic activities take place at lower frequencies (from Hz to some hundreds Hz) with respect to the device’s natural frequency, fn , thus making the effect of ash fall-out distinguishable from ground vibrations. The electronic gain of the amplifier has been conveniently tuned to meet the specifications of the ␮-controller device. A power consumption of the sensing system (including the conditioning electronics for the piezoelectric transducer and the ␮C board) of about 10 mA @ 3.3 V has been measured. A higher power budget is required by the wireless transmitter module. The radio module adopted is a XBee-PRO embedded RF module by Digi @ 2.4 GHz, whose power consumption in transmission mode is 215 mA @ 3.3 V. Experiments presented in the following are aimed to validate the sensing methodology and to assess prototype performances when executing the ash granulometry classification task. To such aim, experiments have been performed by repeating single particle falls, while future efforts will be dedicated to the characterisation of the system’s behaviour in the case of continuous falling flows of volcanic ash. Real experimental investigations have been performed in laboratory by using volcanic ash from Etna volcano with three different

18

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

Fig. 5. (a) Schematic of the conditioning circuit used for the piezoelectric transducer. Values of components are Ceq = 14 nF, Req = 67 k, Cf = 22 nF, Rf = 1.2 M; R1 and R2 are trimmers of 10 k and 100 k, respectively, used to appropriately tune the amplifier gain. (b) Frequency response of the charge amplifier.

typical size (small, medium and large) ranging from 0.4 mm to 2.0 mm, as shown in Fig. 6. Ash particles have been dropped on the sensor from a height of at least 4 m to ensure they reach the terminal velocity. Wind conditions have not been taken into account also due to the consideration that wind will not compromise the functionality of the sensing node while it will define the ash fall out distance. A typical output signal from the readout electronics is shown in Fig. 7(a). A Matlab® algorithm has been developed to automatically window the raw signal in order to extract information due to the impact of ash particles on the piezoelectric. The algorithm uses a threshold mechanism to detect the useful part of the signal due to ash impact. Examples of signals obtained for the three granulometries investigated are shown in Fig. 7(b)–(d). As can be observed, for the examples considered the signal peak value increases with ash particle size. Through this paper, the maximum signal amplitude has been chosen as the granulometry classification parameter. This choice is strictly related to the simplicity of extracting such a quantity from the observed signals, which is also essential for the implementation of automatic classification features by embedded architectures. In any case, future efforts will be dedicated to investigating the classification features of other signal signatures (e.g. fall time). Distributions of voltage peak amplitudes, A, for the three granulometries investigated are shown in Fig. 8. As can be observed,

distributions partially overlap with the consequence that small sized ash particles could be classified as medium or vice versa, as well as medium grains potentially being recognised as large and vice versa. The classification of ash particles size is not a conclusive task, due also to their irregular shape, and the definition of belonging to one granulometry class rather than the closest one is not trivial. Anyway, as shown in Fig. 8, the average values of the considered output quantity, A, are well separated. Moreover, as shown in Fig. 9 and according to model (1–3), A increases with the size of the ash particles, following the square law reported in the figure itself. The situation depicted in Fig. 8 is typical of two-class prediction problems, also named binary classification. The case under investigation can be approached as two separate binary classification problems which require two thresholds, Th1 and Th2. A useful approach to properly fix threshold values can be based on ROC analysis [18,19]. In the next section a brief introduction to the ROC curve theory will be given together with its application in the ash granulometry classification problem. 4. ROC analysis as a methodology for ash granulometry classification In a simple two-class prediction problem (positive and negative classes) the classifier is required to map each instance (i.e. samples to be classified) to one of the two classes. The best case rarely exists, i.e. two classes are perfectly separated with a well-defined cut-off value. Indeed, a situation like the one shown in Fig. 10(a) with an overlap between two classes is typically observed. For every possible cut-off point or criterion value adopted to discriminate between two populations, there will be four possible outcomes:

Fig. 6. Volcanic ash with the three different granulometries used to perform experimental investigations.

• some instances correctly classified as positive (TP = True Positive fraction). • some instances wrongly classified as negative (FN = False Negative fraction). • some instances correctly classified as negative (TN = True Negative fraction). • some instances wrongly classified as positive (FP = False Positive fraction).

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

19

3

5 4

small size

2

Raw signal

Output voltage Vout (V)

Output voltage Vout (V)

3 2 1 0 -1

1

0

-1

-2

-2 -3 -4 1

1.5

2

-3 0

2.5

time (s)

0.005

0.01 time (s)

(a)

(b) medium size

3

large size

10

2

Output voltage Vout (V)

Output voltage Vout (V)

0.02

15

4

1 0 -1

5

0

-5

-10

-2 -3 0

0.015

0.005

0.01 time (s)

0.015

0.02

-15 0

0.005

0.01 time (s)

(c)

0.015

0.02

(d)

Fig. 7. (a) Raw output voltage signal from the conditioning electronics for the piezoelectric; (b), (c) and (d) The output voltage signal after windowing for the case of small, medium and large sized ash particles, respectively.

Two indexes of interest can be defined: the True Positive Rate (TPR) or Sensitivity as the ratio between the number of true positives and the total positive instances:

and the False Positive Rate (FPR) or 1-Specificity (with Specificity defined as the True Negative Rate) as the ratio between the number of false positives and the total negative instances: FPR =

Th1

TP = Sensitivity TP + FN

FP = 1 − Specificity TN + FP

A ROC curve is a 2D plot of the TPR (sensitivity) vs. FPR (1specificity) for a binary classifier system, as its discrimination threshold varies. An example of a ROC curve is shown in Fig. 10(b). ROC graphs are used to depict the trade-off between the hit rates (benefits) and false alarm rates (costs) of classifiers [20,21]. To choose a point on the ROC curve with a higher TPR means to move to the left of the threshold (towards lower values) with the consequence that on one hand more positive instances will be classified as positive (TP increases and FN decreases) but on the other hand more negative instances will be wrongly classified as positive (FP increases and TN decreases).

Th2

7

medium

small

large

6

Frequency

TPR =

8

5 4 3 2 1 0 0

2

4

6 8 Output Voltage (V)

10

Fig. 8. Distribution of the output voltage peaks.

12

20

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

Similarly, when the two distributions of medium and large particles are taken in account and a threshold value Th2 is fixed, the four fractions will be defined as:

12 A=2.6605*R2A+1.1906 Average output voltage A (V)

10

TP = fraction of A quantity due to large ash particles correctly classified as large. FN = fraction of A quantity due to large ash particles classified as medium. TN = fraction of A quantity due to medium ash particles correctly classified as medium. FP = fraction of A quantity due to medium ash particles classified as large.

8

6

4

2

0 0.4

0.6

0.8

1 1.2 1.4 1.6 Radius RA of the ash (mm)

1.8

2

Fig. 9. Average output voltage vs. the size (RA ) of ash particles. Symbols represent real observations, while the continuous curve is the model fitting.

In the case of perfect discrimination (no overlap in the two distributions) the ROC curve passes through the upper left corner (100% sensitivity and 100% specificity). Therefore the closer the ROC curve is to the upper left corner, the higher the overall accuracy of the classifier. The worst case is that in which the two distributions are completely overlapped, thus making the classification task impossible, as represented in Fig. 10(b) by the no discrimination line. With reference to Fig. 8, when the two distributions of small and medium particles are taken in account and a threshold value Th1 is fixed, the four fractions previously introduced will be defined as: TP = fraction of A quantity due to medium ash particles correctly classified as medium. FN = fraction of A quantity due to medium ash particles classified as small. TN = fraction of A quantity due to small ash particles correctly classified as small. FP = fraction of A quantity due to small ash particles classified as medium.

Applying the above theory of ROC curves to the ash granulometry problem, taking in account the distribution of voltage peaks in Fig. 8, curves in Fig. 11 have been obtained. Considering that each point on the ROC curve identifies a threshold value, the following step consists of fixing suitable thresholds. As an example, if the classification problem for the small and medium sized particles is taken in account, to move the threshold to the left (higher TPR) means increasing the number of small particles classified as medium, and to decrease the number of medium particles classified as small. Following the same reasoning it is easy to deduce the consequence of choosing a point on the ROC curve with a lower TPR. As emerges from Fig. 11(b) and consistently with information in Fig. 8, a suitable threshold value, Th2, discriminating large sized grains from medium sized grains with very good performances (close to the best case of separated distributions) can be easily defined. Similarly, a threshold value, Th1, to discriminate small ash particle from medium particles could be also defined with good performance (less good compared to the other case). The results obtained for the case of the study presented, and in general the proposed methodology, can provide strategic input to experts in order to properly fix classification criteria. The latter could be definitively tailored by the position of the monitoring station in the route between craters and airports as well as by the interest of properly detecting the amount of small ash particles rather than medium or large grains. Although the operative context (position of the monitoring station and weather conditions) should be taken into account, some preliminary criteria to properly choose classification thresholds can be drawn.

Fig. 10. (a) Example of binary classification; (b) The ROC curve related to the example in (a).

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

ROC curve for the medium - large case 1

0.9

0.9

0.8

0.8

True Positive Rate (Sensitivity)

True Positive Rate (Sensitivity)

ROC curve for the small - medium case 1

0.7 0.6 0.5 0.4 0.3 0.2

small - medium ash size worst case

0.1 0 0

0.2

0.4 0.6 0.8 False Positive Rate (1-Specificity)

21

0.7 0.6 0.5 0.4 0.3 medium - large ash size worst case

0.2 0.1

1

0 0

0.2

0.4 0.6 0.8 False Positive Rate (1-Specificity)

(a)

1

(b)

Fig. 11. ROC curves for the volcanic ash granulometries classification. (a) Curve for the discrimination between small and medium sized grains; (b) curve for the discrimination between medium and large sized grains.

On the basis of the above considerations, in case the classification problem between small and medium ash particles is addressed, moving the threshold to the right affects the right classification of the medium size ash particles while it improves the correct classification of the small size ash particles. Just the opposite happens in case of moving the threshold to the left. Since small particles can be transported over long distances, the effect of above choice would be the missing of warning alarms with a consequently higher risk for the safety of air transport. For the specific application addressed, a correct classification of small size particles is mandatory also at the expense of false alarms due to medium particles classified as small ash. The same considerations can be made for the classification problem between medium and large ash particles. The same considerations can be applied to the classification problem between medium and large ash particles. Also in this case the right classification of medium particles can be crucial for air transport safety with respect to the correct classification of large ash. 5. Conclusions In this paper a low-cost smart multisensor system for the monitoring of ash fall-out phenomenon developed by researchers at DIEEI of the University of Catania, and involved in the SECESTA project, has been presented. In particular the paper focuses on the methodology adopted for ash granulometry detection. The main idea is to use a piezoelectric transducer to convert ash impacts into electrical signals which convey information about the ash granulometry. Experimental results showing the capability of the proposed approach are presented. Attention has been focused on the peak value of the sensor output signal developed. ROC curve analysis has been applied as a theoretical support to suitably fix thresholds for the sake of ash granulometry classification. Further efforts will be dedicated to investigating the possibility of taking into account the contribution of other parameters of the impulse response (such as the falling time of the voltage signals), in order to improve the system’s classification features. The system’s behaviour in the case of continuous falling flows of volcanic ash will be also addressed. Moreover, considering the hostile working environments in which the system must operate, future efforts will be dedicated to the temperature compensation issue.

The possibility of using the piezoelectric system to implement a trigger mechanism for the automatic operation of the multisensor node in the case of volcanic explosions or seismic events will also be better investigated and assessed. Acknowledgments This work is developed under the SECESTA project of the POR FESR Sicilia 2007–2013 (CUP: G53F11000040004). In particular, the authors wish to thank Dr. M. Coltelli of INGV, Catania. References [1] W.H. Mayer, The Mitigation of Ash Fall Damage to Public Facilities: Lessons Learned from the 1980 Eruption of Mt. St. Helens, Washington Federal Emergency Management Agency, Region X, 1984. [2] J. Neild, P. O’Flaherty, P. Hedley, R. Underwood, D. Johnston, B. Christenson, P. Brown, Agriculture recovery from a volcanic eruption, MAF Technical Paper 99/2, 1998. [3] C. Horwell, P. Baxter, The health hazards of volcanic ash – a guide for the public, Available at www.ivhhn.org/pamphlets.html [4] Flight Operations Briefing Notes – Operating Environment – Volcanic Ash Awareness, Rev. 01, September 2006, Available at www.airbus.com [5] ICAO, Management of flight operations with known or forecast volcanic cloud contamination, Guidance Material, ICAO Document, vers. 3.1, December 2010. [6] ENAC, APT 15 – Operazioni volo su aeroporti in presenza di nube di cenere vulcanica and annex Doc. APT-ETNA ed. 1, ENAC Circular (in Italian), 23 July 2003. [7] S. Scollo, M. Prestifilippo, G. Spata, M. D’Agostino, M. Coltelli, Monitoring and forecasting Etna volcanic plumes, Natural Hazards and Earth System Sciences 9 (2009) 1573–1585, http://dx.doi.org/10.5194/nhess-9-1573-2009. [8] F. Marchese, R. Corrado, N. Genzano, G. Mazzeo, R. Paciello, N. Pergola, V. Tramutoli, Assessment of the robust satellite technique (RST) for volcanic ash plume identification and tracking, in: Second Workshop on the Use of Remote Sensing Techniques for Monitoring Volcanoes and Seismogenic Areas (USEReST 2008), 2008, pp. 1–5. [9] F.S. Marzano, E. Picciotti, G. Vulpiani, M. Montopoli, Synthetic signatures of volcanic ash cloud particles from X-band dual-polarization radar, IEEE Transactions on Geoscience and Remote Sensing 50 (1) (2012) 193–211. [10] A.J. Prata, C. Bernardo, Retrieval of volcanic ash particle size, mass and optical depth from a ground-based thermal infrared camera, Journal of Volcanology and Geothermal Research 186 (2009) 91–107. [11] B. Andò, S. Baglio, V. Marletta, A smart multisensor system for the ash fall-out monitoring, in: Proceedings of the 26th European Conference on Solid-State Transducers (Eurosensors 2012), Krakow, Poland, September 9–12, 2012. [12] B. Ando, S. Baglio, N. Pitrone, C. Trigona, A.R. Bulsara, V. In, M. Coltelli, S. Scollo, A novel measurement strategy for volcanic ash fallout estimation based on RTD Fluxgate magnetometers, in: IEEE Instrumentation and Measurement Technology Conference Proceedings (I2MTC 2008), 2008, pp. 1904–1907.

22

B. Andò et al. / Sensors and Actuators A 202 (2013) 13–22

[13] W.J. Stronge, Impacts Mechanics, Cambridge University Press, Cambridge, 2000. [14] L. Gao, Y. Yan, G. Lu, On-line measurement of particle size distribution using piezoelectric sensors, in: IEEE International Instrumentation and Measurement Technology Conference (I2MTC), May 2012, pp. 1154–1158. [15] B. Ando, M. Coltelli, M. Prestifilippo, S. Scollo, A lab-scale experiment to measure terminal velocity of volcanic ash, IEEE Transactions on Instrumentation and Measurement 60 (4) (2011) 1340–1347. [16] Inc. Measurement Specialties, Piezo Film Sensors Technical Manual, Measurement Specialties Inc., 1999, Available at www.msiusa.com [17] J.C. Lagarias, J.A. Reeds, M.H. Wright, P.E. Wright, Convergence properties of the Nelder–Mead simplex method in low dimensions, SIAM Journal on Optimization 9 (1) (1998) 112–147. [18] T. Fawcett, ROC graphs: Notes and practical considerations for data mining researchers, HP Technical Report HPL-2003-4, HP Laboratories, 2003. [19] A. Slaby, ROC analysis with Matlab, in: 29th International Conference on Information Technology Interfaces (ITI 2007), 2007, pp. 191–196. [20] D. Hand, R.J. Till, A simple generalisation of the area under the ROC curve for multiple class classification problems, Machine Learning 45 (2001) 171–186. [21] Z.-C. Qin, ROC analysis for predictions made by probabilistic classifiers, in: Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, August 2005, pp. 18–21.

Biographies B. Andò is an associate professor in measurement science and he received his M.S. and Ph.D. in EE at the University of Catania, Italy, in 1994 and 1999, respectively. From 1999–2001, he worked as a researcher with the Department of Electrical and Electronic Measurement (DIEES) of the University of Catania and in 2002, he became an assistant professor. His main research interests are MEMS and NANO systems, inkjet printed sensors, new materials for sensors, smart multi-sensor architectures for environmental monitoring and AAL and nonlinear techniques for signal processing. Dr. Andò holds several national and international scientific

collaborations and during his activity he has co-authored several scientific papers, presented at international conferences, and published in international journals and books. S. Baglio received the “Laurea” and Ph.D. degrees from the Università degli Studi di Catania, Catania, Italy, in 1990 and 1994, respectively. Since 1996, he has been with the Dipartimento di Ingegneria Elettrica Elettronica e dei Sistemi, Università degli Studi di Catania, where he is currently an associate professor. He teaches courses in “measurement theory,” “electronic instrumentations,” and “integrated microsensors.” He is the coauthor of more than 250 scientific publications, including books, chapters in books, and papers in international journals and proceedings of international conferences. He is the holder of several U.S. and European patents. His research interests are mainly focused on micro- and nanotechnologies, micro- and nanosensors, hysteretic materials for sensors, and nonlinear dynamics for transducers. Dr. Baglio has served as an associate editor for the IEEE Transactions on Circuits and Systems and as a distinguished lecturer for the IEEE Circuits and Systems Society. He is currently an associate editor of the IEEE Transactions on Instrumentation and Measurement and the Chair of the Italian IEEE Instrumentation and Measurement Italy Chapter. V. Marletta received the M.S. and the Ph.D. degrees from the University of Catania, Italy, in 2007 and 2011, respectively. He is currently working as temporary research fellow with the Dipartimento di Ingegneria Elettrica, Elettronica e Informatica (DIEEI) of the University of Catania. His research interests include sensor design and characterisation, new materials for sensors, inkjet printed sensors, energy harvesting, smart multi-sensor architectures for environmental monitoring and AAL, soft-computing methodologies for instrumentation and measuring systems, smart sensors and exploitation of nonlinear dynamics in sensors, microsensors, and microsystems in standard and dedicated technologies. He is the coauthor of more than 40 scientific publications, which include chapters in books and papers in international journals and proceedings of international conferences. S. Medico received the B.S. from the University of Palermo, Italy, in 2008 and the M. S. from the University of Catania, Italy, in 2011. He is currently working with the Dipartimento di Ingegneria Elettrica, Elettronica e Informatica (DIEEI) of the University of Catania. His research interests include sensor design and characterisation of inkjet printed sensors, energy harvesting, soft-computing methodologies for instrumentation and measuring systems.