Wireless Sensor Network Simulation for Fault Detection in IndustrialProcesses

Rui Pinto1, Rosaldo J. F. Rossetti1,2 and Gil Goncalves1

1Faculty of Engineering, University of Porto, Rua Dr. Roberto Frias s/n, 4200-465 Porto, Portugal2Artificial Intelligence and Computer Science Lab, University of Porto, Rua Dr. Roberto Frias s/n, 4200-465 Porto,

Portugal{rpinto, rossetti, gil}@fe.up.pt

Keywords: Castalia, Sensor Diagnosis and Validation, Industrial Wireless Sensor Networks.

Abstract: Sensor data is extremely important to monitor machines at the shop-floor level and its environmental surround-ing conditions for condition-based monitoring, machine diagnosis and process adaptation to new requirements.Based on the described scope, self-diagnostics and self-organizing capabilities are core functionalities of anyIndustrial Wireless Sensor Network (IWSN). In the present work, a simulated case study was developed withthe main intent of validating techniques implemented for sensor data diagnosis of error detection and equip-ment failure. The scenarios explored try to mimic some common situations of a manufacturing environmentwhen dealing with WSNs, where a piece of sensor equipment suddenly stops working or an unpredictablechange in the environment leads to faulty data readings. This paper introduces Castalia and describes howit was used to simulate a direct application of an Optical Metrology System on an industrial Resistance SpotWelding process, which is composed of a camera and several luminosity sensors. More specifically, a sensordata validation module was proposed, implemented and used to extend Castalia functionalities.

1 INTRODUCTION

Nowadays, several European initiatives towards smartmanufacturing systems and Industry 4.0 are a cur-rent subject of research, aiming at the improvementof the European Industry competitiveness regardingother leading markets. This can be achieved by turn-ing shop-floor production methods more flexible andefficient facing the constant changes in production,which is driven by consumer demands. Since con-sumers’ preferences are becoming more customizedand competitiveness between companies is extremelyhigh, manufacturers must quickly adapt their productto new trends. Usually, the company pioneer on hav-ing a new product available on the market have theupper hand over the competition. When a mass pro-duction model is used, maintaining a large envelopeof products or enlarging it with more product vari-ants is extremely difficult and entails a great effort andcost due to the inflexibility of mass production man-ufacturing systems. To support the mass customiza-tion paradigm at the shop-floor level, the productionsystems should be capable of rapidly reconfiguring asproduct demand varies, as well as does machine pa-rameter calibration during all production stages. The

time spent to calibrate and to adjust the machine pro-cess parameters when a new product variation shouldbe performed, or after a maintenance phase, can be re-duced if the machines’ parameter adjustment is doneautomatically instead of the usual trial-and-error ap-proach, by performing destructive tests to evaluatethe process quality. Monitoring machine’s executionis extremely important and, to do so, sensors are de-ployed on the shop-floor to infer surrounding environ-mental conditions and adapt the machine executionaccordingly.

An Optical Metrology System (OMS) is a systemwhere optical measurements are used for several in-dustrial applications, e.g. for robot guidance whenoptical metrology sensors are mounted on a robotarm. In this case, a OMS is used on a Resistance SpotWelding (RSW) process, to infer if the metal parts tobe weld are well positioned. In order to insure thecorrect setting up and easy ramping up of the system,in site factory sensor positioning adjustment and im-age setup is a significant procedure, which involvesconfiguration of camera parameters, such as exposureand gain. With help of luminosity sensors by mea-suring the luminosity conditions in site, the cameraparameters can be adjusted on the fly, according to

the context. This work focus on the simulation of anOMS applied to a RSW process using Castalia. Thegoal was to understand the behaviour of the luminos-ity sensors, specially when they are malfunctioning.A sensor data detection module was implemented inCastalia and used on a simulated version of the de-scribed environment.

The paper is organized in four more different sec-tions. Section 2 details the approaches regardingIWSN, the main methods for sensor data validationand categorize WSN simulators, specially Castalia.Section 3 identifies the case study used to develop thesimulation scenario. Section 4 depicts all the resultsobtained, whereas in Section 5 final remarks about theoverall work exposed in the paper and further steps tobe followed are presented.

2 INDUSTRIAL WIRELESSSENSOR NETWORKS

In the past few years WSNs have become more in-teresting to be explored in and applied to several do-mains. Their use was motivated mainly due to lat-est advances in wireless communications as well asin more reliable, robust and long-lasting hardware.These factors have a great impact on the feasibilityof installation, when it is difficult to use wired so-lutions, either by harsh location or high number ofsensor nodes used and also because of the easy main-tenance and cabling reduced costs. As main advan-tages, Chen et al. (Chen et al., 2015) pinpoints thelarge coverage area, ubiquitous information, fast com-munication via RF and self-organisation throughoutthe direct communication between entities. A SmartSensor Platform (Ramamurthy et al., 2007) was de-veloped, which applies the plug and play concept bymeans of hardware interface, payload, communica-tion between sensors and actuators, and ultimatelyallows for software update using ’over-the-air’ pro-gramming. Cao et al. (Cao et al., 2008) developed adistributed approach to put closer sensors and actua-tors in a collaborative environment using WSNs and,Chen et al. (Chen et al., 2015), push this approachforward considering the same methodology, but tak-ing into account all the industrial domain restrictions.

Despite the huge potential of WSNs, these so-lutions still have several issues and future researchchallenges. Data collected from WSNs is proneto be faulty due to internal and external influences,such as environmental effects, hardware malfunc-tions, software problems, energy constrains, net-work issues, security threats, among others, as shownelsewhere (Tolle et al., 2005), (Barrenetxea et al.,

2008), (Ramanathan et al., 2006) and (Szewczyket al., 2004). IWSNs are used to monitor real-time production equipment in the factory and thesurrounding environmental conditions, acting accord-ingly when changes are observed. According to Neu-mann (Neumann, 2007), there are important restric-tions in industrial applications to be met, such as real-timeliness, functional safety, security, energy effi-ciency, and Quality of Service (QoS). So, IWSNs facesome challenges, such as safety-critical functions, se-curity and privacy of collected data, availability toavoid complete production stop when failures occur,latency/retransmission of messages, support for ac-tuators (by using the same sensor controller), effi-cient integration with existing automation infrastruc-tures, scalability, coexistence and wireless communi-cation interference avoidance and energy harvesting(Akerberg et al., 2011).

2.1 Sensor Data Validation

To maintain QoS, one should be able to detect anddeal with compromised sensor nodes during opera-tion. To determine whether a sensor node is mal-functioning, validation methods are applied to sen-sor data, aiming to find devious sensor readings fromnormal ones. There is no such an ideal validationmethod, because they are very dependent on sensormeasurement conditions and the overall environmentcontext (Freitas et al., 2010), (Branisavljevic et al.,2011), (Bertrand-Krajewski et al., 2003), (Vasconce-los et al., 2012), and (Liu et al., 2013). Usually, sev-eral methods are applied successively to the sensordata, because each method is suitable to detect par-ticular types of faulty data. Sensor data validationmethods are typically divided into online and offlinemethods, where online methods are efficient, simplerto implement and demand less data processing, andoffline techniques, such as Bayesian Networks, Ar-tificial Neural Networks and Regression Techniques,are more robust and complex techniques, which aresuitable to process data not in real-time.

The most common WSNs’ data fault types are:Out-of-range faults - values out of predefined mini-mum and maximum boundaries; Struck-at-fault se-quence of values that have little or no variation for alarge period of time; Outliers misplaced data sam-ples that represent a sensor misbehaviour, which canbe caused by sensor malfunctioning or changes in sur-rounding environment; Stop Communication gaps inthe dataset caused by an absence of sensor informa-tion on the gateway, due to power outage or commu-nication problems; Devious Data when the behaviourof a certain sensor is devious from the majority, and it

persists for a large period of time. A graphical exam-ple of these faults’ behaviour (on a temperature sen-sor) is presented in Figure 1.

According to Sharma et al. (Sharma et al.,2010), Ravichandran and Arulappan (Ravichandranand Arulappan, 2013), the most common techniquesused to perform online detection of the described sen-sor data faults are Min/Max Detection, Flat Line De-tection, Modified Z-Score, No value Detection andSpatial Correlation. The Min/Max Detection findsout-of-range faults using upper and lower limits thatare defined by process limitations. On the other hand,the Modified Z-Score finds outliers faults which, de-spite being inside the process limitations, are valuesthat are unusual when compared to the sensor datahistory. The Flat Line Detection finds struck-at-faults,which are characterized by little or no variation indata and the No value Detection detects situationswhen the data stops being communicated. Finally, theSpatial Correlation is the technique used to comparedatasets of several sensors nodes physically locatednear to one another, and if one of them presents a de-vious dataset in comparison to its neighbours, then thecorresponding sensor node is probably faulty.

2.2 WSN Simulation

A WSN simulation consists in using simulatorsspecifically designed to imitate WSNs behaviour.Traditionally, the three main methodologies foranalysing the performance of WSNs are physicalmeasurements, which consists in setting up all thephysical nodes and collecting sensor data, analyticalmethods, and computer simulation. Since deployinga WSNs requires a huge effort, due to constrains re-garding the cost associated with hardware and the factthat analytical models are not effective, because of thecomplexity of the models for WSNs regarding energylimitation, decentralized collaboration and fault toler-ance, simulation is the only feasible approach to thequantitative analysis of sensor networks, getting ac-cess to fine grained results easier than with real worldexperiments (Dwivedi et al., 2010). Because thereare so many simulators with many different charac-teristics, Eriksson (Eriksson, 2009) proposes to cat-egorize WSNs as Generic Network Simulators - fo-cus more on network aspects, such as radio protocols,network stacks and channel distortions; Code LevelSimulators - focus more on the simulation of the sen-sor nodes, such as deployable code and functional-ity logic; Firmware Level Simulators - focus more onemulating the sensor node, such as microprocessor,radio chip and other peripherals.

Song (Song et al., 2011) implemented a fault de-

tection mechanism to analyse the stability and relia-bility of data transmission in a ZigBee network, whichwas simulated on the NS-2 simulator. Results focuson the data loss rate in the communication betweensensor nodes. Dai (Dai et al., 2011) also proposeda fault detection mechanism, but focused more onthe delays introduced by the wireless communicationnetworks, where the control system was simulated inMATLAB and the network-induced delays were sim-ulated in OMNeT++. Both Song and Dai focusedmore on faults with root cause on network aspects. Onthe other hand, Zhang (Zhang et al., 2009) and Szc-zodrak (Szczodrak et al., 2008) used Castalia to im-plement a fault detection algorithm based on temporaland spatial correlations of sensor data from neighboursensor nodes, classifying each node as good or faulty.

2.2.1 Castalia

Castalia (Boulis, 2007) was the simulator chosenfor this work. Figure 2 presents Castalia’s architec-ture, where each module accepts messages from othermodules or itself and, according to the message, it ex-ecutes a given code. The nodes do not connect toeach other directly, but through the Wireless Chan-nel module, which estimate the average path loss be-tween two nodes using the Lognormal Shadowing.The nodes are also linked through Physical Processmodules that they monitor, which ”feed” the sensorswith data. For every physical process there is onemodule which holds the ”truth” on the quantity of thephysical process that is representing. The nodes sam-ple the physical process in space and time to get theirsensor readings.

The Physical Process modules are based on an ar-bitrary number of point sources whose ’influence’ isdiffused over space and they can change their posi-tion and their value over time. The effect of multiplesources in a certain point is additive. Calculating thevalue of the physical process at a certain location andat a certain time is represented in Equation (1), whereV (p, t) denotes the value of the physical process atpoint p and at time t, Vi(t) denotes the value of the ith

source at time t, di(p, t) denotes the distance of pointp from the ith source at time t, K and a are parame-ters that determine how the value from a source is dif-fused. N(0,θ) is a zero-mean Gaussian random vari-able with standard deviation θ. The ”ground truth”offered by the physical process is distorted by the in-accuracies of the sensing devices, implemented by theSensor Manager.

V (p, t) = ∑i

Vi(t)(Kdi(p, t)+1)a +N(0,θ) (1)

Figure 1: Sensor Data Faults Characterization.

Figure 2: Model Structure of Castalia.

Regarding the internal structure of a node, theApplication module is the one that specifies the al-gorithm that receives sensor data and acts accord-ingly. The MAC and Routing modules define severalcommunication protocols, which can be used by thenodes. The Mobility module defines the nodes posi-tion and mobility pattern. The Sensor Manager de-fines the sensing devices present in each node, by in-troducing a distortion of the ground truth offered bythe physical process. The Resource Manager definesthe energy consumed by the different components ofthe node.

3 METHODOLOGY

The simulation scenario tried to mimic a OMS,which is composed by an IWSN, where all nodes areequipped with luminosity sensors and a central nodeis equipped with a camera. This OMS is used to anal-yse the quality of a RSW process, namely if both partsto be weld are correctly placed and, in the end of thewelding cycle, evaluates the quality of the weld. Thisanalysis consists in capturing and processing a cameraimage captured, which must have a minimum quality,

otherwise, image analysis will be impossible. In orderto avoid excessive darkness/clarity in the image, theexposure index of the camera must be calibrated ac-cording to the luminosity conditions in the surround-ing environment. This process consists in controllingthe lightness and darkness of the image, which can bedone by the camera light meter. Since the camera lightmeter may have a lower accuracy and, it measuresonly one single point of luminosity per sample, sev-eral luminosity sensors are used to infer the luminos-ity conditions around the welding area and improvethe control of the camera exposure index.

In this case, a camera is placed right on top of thetwo electrodes that will weld the parts together. In or-der to evaluate the luminosity conditions, nine lumi-nosity sensors were placed around the electrodes. Allthe sensor data is processed on a gateway node, result-ing on an average of the nine luminosity points. Thesimulation of this scenario will allow to implementand test a sensor fault detection module before beingdeployed in a real environment. Figure 3 representsa real application of this scenario and the schema ofthe virtual world to be simulated. There are eight sen-sor nodes (represented by the green circles), each oneequipped with a luminosity sensor that outputs val-ues between 0 and 100%. At the center of the worldare placed the luminosity source and sensor node 0.Even numbered sensor nodes are placed at 20 cm dis-tance around the center, so their output should be sim-ilar. The same is true for odd numbered sensor nodes,which are placed at 50 cm distance and, in conse-quence, their output should be smaller than the evennumbered sensor nodes.

Regarding the simulation scenario, only one lumi-nosity source was considered (placed at the center ofthe world), which had default parametrization values,

Figure 3: OMS used on a RSW process: a) real prototype;b) simulation.

namely K equals 0.25, a equals 1.0 and θ equals 0.2.The output of the source was defined at a theoreticalvalue of 65%. According to the simulated world im-plemented, sensor node 0 outputs higher luminosityvalues (around 65%), because it is the one closest tothe luminosity source. Then, even numbered sensornodes output values around 55% and odd numberedsensor nodes output values around 30%. For the com-munication module, the radio’s parameters used arethe ones proposed in the IEEE 802.15 Task Group 6documents, the routing protocol is the Multipath RingRouting and the MAC protocol is the IEEE 802.15.4.

4 TESTS AND PRELIMINARYRESULTS

Some tests were performed in order to simulate devi-ations in sensor readings. The sensor manager’s de-vice noise property was changed accordingly, whichis by default as low as possible (but larger than 0).When this value increases, the difference of sensoroutput values compared to normal outputs also in-creases. Bigger deviations resulted on out-of-rangeand outlier faults. On the other hand, a device noiseequals to 0 corresponds to a null variation, resultingon struck-at-fault samples. Figure 4 plots the percent-age of simulated errors in the dataset of sensor node0 and the efficiency of the techniques for error detec-tion, calculated by determining the false positives andfalse negatives.

The considered dataset had 18% of out-of-rangefaults, 36% of outliers and spikes and 14% of struck-at-faults. The remaining 32% of the dataset is normaldata. The Min/Max Detection pinpointed all out-of-range faults, presenting zero false positive and nega-tive. The Flat Line Detection technique missed oneout of 14 stuck-at-faults present in the data set, pre-senting 1 false positive and zero false negatives. TheModified Z-Score detected only 12 out of 18 outlierspresent in the sensor data set, having 6 false positivesand zero false negatives. The Modified Z-Score tech-

Figure 4: Efficiency of techniques for sensor data error de-tection.

niques efficiency depends greatly on the data set his-tory, namely the rate between the number of normalsamples and the number of outliers, and the differ-ence between the normal and outlier sample values.In order to improve the efficiency of this method, thenumber of normal samples must be much larger thanthe number of outliers.

5 CONCLUSIONS AND FUTUREWORK

The present work simulates an OMS applied to an in-dustrial RSW process, which analyses the position ofmaterials to be welded by means of a camera and sev-eral luminosity sensors. The simulation implementsan IWSN whose nodes have a luminosity sensor. InCastalia, sensors are ”fed” by a luminosity physicalprocess and run a sensor data validation module todetect failing sensors. Such a sensor data validationmodule implements Min/Max Detection, Flat LineDetection and Modified Z-Score, which are used todetect out-of-range faults, struck-at-faults, and out-liers and spikes, respectively. All techniques per-formed as expected, even though the Modified Z-Score accuracy demonstrated to depend greatly on thehistory of dataset samples, as well as on the mean dif-ference between normal and outlier samples as dis-cussed in the result analysis.

Regarding future steps, the sensor data validationmodule can be improved by implementing the SpatialCorrelation method, considering complex scenarioswith different location luminosity sources and detect-ing the presence and location of a sensor fault. Thiscomplex feature implies the use of online sensor fu-sion techniques of several heterogeneous sensor dataor offline complex methods, such as machine learn-ing techniques. Such an approach to using multipledata types and sources is the essence of a multi-agentmethodology (Passos et al., 2011) and architecture(Rossetti et al., 2007) currently under development,

also with applications to the industrial domain (Bragaet al., 2008).

REFERENCESAkerberg, J., Gidlund, M., and Bjorkman, M. (2011). Fu-

ture research challenges in wireless sensor and actu-ator networks targeting industrial automation. In In-dustrial Informatics (INDIN), pages 410–415. IEEE.

Barrenetxea, G., Ingelrest, F., Schaefer, G., Vetterli, M.,Couach, O., and Parlange, M. (2008). Sensorscope:Out-of-the-box environmental monitoring. In Infor-mation Processing in Sensor Networks (IPSN’08),pages 332–343. IEEE.

Bertrand-Krajewski, J.-L., Bardin, J.-P., Mourad, M., andBeranger, Y. (2003). Accounting for sensor calibra-tion, concentration heterogeneity, measurement andsampling uncertainties in monitoring urban drainagesystems. Water Science & Technology, 47:95–102.

Boulis, A. (2007). Castalia: revealing pitfalls in design-ing distributed algorithms in wsn. In 5th InternationalConference on Embedded networked sensor systems,pages 407–408. ACM.

Braga, R., Rossetti, R., Reis, L. P., and Oliveira, E. C.(2008). Applying multi-agent systems to simulate dy-namic control in flexible manufacturing scenarios. InCybernetics and Systems 2008, pages 488–493, Vi-enna. OSGK.

Branisavljevic, N., Kapelan, Z., and Prodanovic, D. (2011).Improved real-time data anomaly detection using con-text classification. Journal of Hydroinformatics,13(3):307–323.

Cao, X., Chen, J., Xiao, Y., and Sun, Y. (2008). Distributedcollaborative control using wireless sensor and actu-ator networks. In Future Generation Communicationand Networking (FGCN’08), pages 3–6. IEEE.

Chen, J., Li, H., Sheng, D., and Li, W. (2015). A hy-brid data-driven modeling method on sensor condi-tion monitoring and fault diagnosis for power plants.International Journal of Electrical Power & EnergySystems, 71:274–284.

Dai, X., Long, Z., Yang, B., and Tan, X. (2011). Robustfault detection observer in wireless networked controlsystems. In Chinese Control and Decision Conference(CCDC), pages 2147–2152. IEEE.

Dwivedi, A., Patle, V., and Vyas, O. (2010). Investigationon effectiveness of simulation results for wireless sen-sor networks. Information Processing and Manage-ment, pages 202–208.

Eriksson, J. (2009). Detailed simulation of heterogeneouswireless sensor networks. Licentiate thesis, UppsalaUniversity, Sweden.

Freitas, T. R., Coelho, A., and Rossetti, R. (2010). Correct-ing routing information through GPS data processing.In Intelligent Transportation Systems (ITSC), pages706–711. IEEE.

Liu, S., Araujo, M., Brunskill, E., Rossetti, R., Barros,J., and Krishnan, R. (2013). Understanding sequen-tial decisions via inverse reinforcement learning. In

Mobile Data Management (MDM), pages 177–186.IEEE.

Neumann, P. (2007). Communication in industrial automa-tionwhat is going on? Control Engineering Practice,15(11):1332–1347.

Passos, L. S., Rossetti, R., and Gabriel, J. (2011). An agentmethodology for processes, the environment, and ser-vices. In Intelligent Transportation Systems (ITSC),pages 2124–2129. IEEE.

Ramamurthy, H., Prabhu, B., Gadh, R., and Madni, A. M.(2007). Wireless industrial monitoring and control us-ing a smart sensor platform. Sensors Journal, IEEE,7(5):611–618.

Ramanathan, N., Balzano, L., Burt, M., Estrin, D., Harmon,T., Harvey, C., Jay, J., Kohler, E., Rothenberg, S., andSrivastava, M. (2006). Rapid deployment with confi-dence: Calibration and fault detection in environmen-tal sensor networks. Center for Embedded NetworkSensing.

Ravichandran, J. and Arulappan, A. I. (2013). Data valida-tion algorithm for wireless sensor networks. Interna-tional Journal of Distributed Sensor Networks, 2013.

Rossetti, R., Oliveira, E. C., and Bazzan, A. (2007). To-wards a specification of a framework for sustainabletransportation analysis. In 13th Portuguese Confer-ence on Artificial Intelligence, EPIA, Guimaraes, Por-tugal, pages 179–190. APPIA.

Sharma, A. B., Golubchik, L., and Govindan, R. (2010).Sensor faults: Detection methods and prevalence inreal-world datasets. ACM Transactions on Sensor Net-works (TOSN), 6(3):23.

Song, C.-C., Feng, C.-F., Wang, C.-H., and Liaw, D.-C.(2011). Simulation and experimental analysis of azigbee sensor network with fault detection and recon-figuration mechanism. In Asian Control Conference(ASCC), pages 659–664. IEEE.

Szczodrak, M., Zahedi, S., Ji, P., Mylaraswamy, D., Sri-vastava, M., and Young, R. (2008). Simulation frame-work for qoi characterization of sensor networks in thepresence of faults.

Szewczyk, R., Mainwaring, A., Polastre, J., Anderson, J.,and Culler, D. (2004). An analysis of a large scalehabitat monitoring application. In 2nd InternationalConference on Embedded networked sensor systems,pages 214–226. ACM.

Tolle, G., Polastre, J., Szewczyk, R., Culler, D., Turner,N., Tu, K., Burgess, S., Dawson, T., Buonadonna,P., Gay, D., et al. (2005). A macroscope in the red-woods. In 3rd International Conference on Embeddednetworked sensor systems, pages 51–63. ACM.

Vasconcelos, G., Petry, M., Almeida, J. E., Rossetti, R., andCoelho, A. L. (2012). Using UWB for human tra-jectory extraction. In 24th European Modeling andSimulation Symposium, EMSS, pages 428–433.

Zhang, C., Ren, J., Gao, C., Yan, Z., and Li, L. (2009). Sen-sor fault detection in wireless sensor networks. In IETInternational Communication Conference on WirelessMobile and Computing, pages 66–69.