Assessing neural neworks for sensor faul deecion Georg Jäger, Sebasian Zug, Tino Brade, André Dierich, Chrisoph Seup Oo-von-Guericke Universiä Magdeburg Deparmen of Disribued Ssems Magdeburg, German Email: gjaeger@s.ovgu.de, {zug, brade, dierich}@ovgu.de Ana-Maria Creu Universie du Quebec en Ouaouais Deparemen of Compuer Science Canada Email: ana-maria.creu@uqo.ca Chrisian Moewes Oo-von-Guericke Universiä Magdeburg Insiu of Knowledge and Language Engineering Magdeburg, German Email: cmoewes@ovgu.de Absrac The idea of smar sensing includes a permanen monioring and evaluaion of sensor daa relaed o possible measuremen fauls. This concep requires a faul deecion chain covering all relevan faul pes of a specific sensor. Addiionall, he faul deecion componens have o provide a high precision in order o generae a reliable quali indicaor. Due o he large specrum of sensor fauls and heir specific characerisics hese goals are difficul o mee and error prone. The developer manuall deermines he specific sensor characerisics, indicaes a se of deecion mehods, adjuss parameers and evaluaes he composiion. In his paper we exploi neural-nework approaches in order o provide a general soluion covering pical sensor fauls and o replace complex ses of individual deecion mehods. For his purpose, we idenif an appropriae se of faul relevan feaures in a firs sep. Secondl, we deermine a generic neural-nework srucure and learning sraeg adapable for deecing muliple faul pes. Aferwards he approach is applied on a common used sensor ssem and evaluaed wih deerminisic faul injecions. I. INTRODUCTION Fuure conceps for sensor acuaor ssems ( Inerne of Things (cf. [1]), Cber-Phsical Ssems (cf. [2]), or Pervasive/Ubiquious Compuing (cf. [3])) are focused on disribued smar devices. The organize hemselves based on a curren ask and aggregae adapivel he needed environmen informaion from available and relevan sensors. The coninuous adapaion and reconfiguraion breaks wih classic design paerns based on a saic configuraion on design-ime. In order o provide he dnamic composiion, we have o appl addiional conceps relaed o loosel coupled communicaion, self-descripion (daa pes, phsical unis, uncerain), snchronizaion and faul-handling [4]. The las poin is paricularl challenging due o he large specrum of sensor faul pes. If an applicaion should be able o weigh or validae a sensor measuremen correcl, i needs addiional validi informaion. This value indicaes he possibili of a faul during he measuremen process. Previous publicaions mapped he resul of he mos relevan faul deecion operaion on simple scalar values [5]. We enhanced he concep and developed a vecorized faul indicaor for pical sensor faul pes summarized in Tab. I. The vecor covers all faul pes relevan for a cerain sensing seup and provides a fine grained absracion of he measuremen validi. Consequenl, he approach requires a specific validaion algorihm for each faul pe ha is relevan for he implemened ransducer. Relaed o he differen characerisics of he sensor fauls (duraion, derivaion, sochasic properies) a huge amoun of mehods were presened for faul deecion (see [6]). Hence, he developer has o choose an appropriae faul deecion mehod and has o deermine he magic consans (hresholds, weighing facors, limis) implemening a complex faul model for a sensor ssem. These manual adjusmens do no correspond wih our idea of an effecive developmen process for smar sensors. In previous work we describe he use of sensor descripion files providing an auomaed sofware-developmen process in Mahworks Simulink. Based on a machine readable sensor characerisics (iming, phsical unis, signal dnamic) and a arge descripion (processor, communicaion inerfaces, ADC properies, ec.) we generae source code for inerfaces and a basic processing chain of a smar sensor [7]. If we wan o embed he sensor faul handling in his approach, we need o define a general and simplified wa o design faul deecors in order o reduce manual inervenions and adjusmens. A he end, we are able o evaluae sensor descripion files conaining possible faul models and generae and configure auomaicall an appropriae se of deecion mehods. To reach his goal, we have o encapsulae he developmen of a suiable deecion mehods and heir parameers in a more absrac wa. Consequenl we looked for an approach ha covers a large specrum of sensor fauls based on a parameer se developed in an auomaable process and can be execued on embedded devices. In his paper we evaluae he capabiliies of Neural Neworks for his purpose. Neural Neworks can be applied on complex classificaion problems. The number of parameers ha have o be adjused are limied o some basic parameers like nework srucure, nework pe, or ransfer funcions. Furhermore his values can be deermined b cross validaion an herefore are also auomaable. Hence, we summarize convenional faul deecion approaches and discuss previous implemenaions based on neural neworks in Sec. II. Saring from his poin we analze he required adjusmens in he neural nework design and configuraion process Sec. III. Sec. IV describes a firs implemenaion and evaluaes he resuls. In Sec. V we provide a final conclusion and an oulook on fuure work. II. STATE OF THE ART Faul deecion represens he firs sep in he faul diagnosis oolchain of Faul Deecion and Isolaion (FDI). According o he auhors of [8], faul deecion onl indicaes ha somehing wen wrong. The correc idenificaion of a cerain faul pe or is origin follows in separae seps. These exensions are planned for fuure work, in his paper we concenrae on A general indicaor for faul measuremens. All faul deecion sraegies require an addiional reference. B comparing i wih curren measuremens or heir feaures (noise, deviaion, correlaion) fauls can be recognized. The reference is generaed based on hardware redundanc (homogeneous or heerogeneous sensors in muli-sensor applicaions),
analical redundanc (mahemaical models of he observed ssem wih predicions) or signal analsis (knowledge on he measuremen characerisic (one or more samples). Comprehensive discussions on sensor faul deecion for all 3 cases are given in [6], [9]. In his paper we address he deecion mehods for individual smar sensors. Hence, hardware redundanc is no considered in he following paragraphs. A. Sensor faul deecion Faul deecion mehods var from simple hreshold checks up o complex signal filer algorihms [10]. The firs variaion uses an implici knowledge of he environmen model and of he inegraed sensor o esimae ranges for a single/muliple se of feaures. If a deecor should be able o recognize ouliers, we will implemen a gradien check for insance. Is upper limis represens he dnamic of he observed ssem and he known normal noise level of he sensor. Consequenl, he developer has o define a suiable feaure in a firs sep and o idenif and o evaluae he hresholds in a second sep [11]. In man cases he hresholds can no be reliabl deermined on design-ime. Adapive hresholds can be an appropriae mehod o cope wih his challenge. Insead of validaing a value b a predefined consan limi, he hresholds are dnamicall arranged. I is common pracice o define a consan and a dnamic par in his case. This mehod guaranees a higher flexibili bu requires more effor for adjusmen and calibraion. The second pe of mehods uses filers o generae a residual signal which can again be hresholded o deec a sensor faul. Tpicall specral filers are used o deec sensor faul signals, oo [12]. Digial filers such as he Savizk-Gola filer [13] can also be used o generae a residual signal. Again exper knowledge is necessar o design such filers. Oher filers are based on he phsical laws of moion, e.g., he Kalman filer ha can even deec fauls in real ime [14]. An alernaive form of redundanc evaluaes mahemaical models. The developer describes he environmen b balancing equaions or rules and ransforms his informaion ino ssem sae models, even/siuaion-calculus, sae machines ec. The faul deecion algorihm analsizes he deviaions and looks for a significan deviaion. A large number of modeling echniques and residual validaion are available ailored for specific scenarios, sensor parameers, or communicaion aspecs [10], [15]. The previousl menioned echniques originae from digial signal processing and/or phsics. All of hese mehods have o be uned manuall and hus generall depend on human exper knowledge. Especiall in complex sensor-based ssems, he definiion of analical faul deecion mehods is eiher inconsisen or ime-consuming. Shor producion ccles require mehods ha learn o deec fauls solel based on observaions. Such daa-driven approaches demand represenaive raining ses ha include normal saes and fauls, boh labeled as such [16]. Then an saisical learning mehod can be used o learn such a model. Rule-based mehods, e.g., decision rees [17], reurn hreshold-based if-hen rules bu do no perform so well. Black-box models such as arificial neural neworks generalize ver well and herefore more appropriae. B. Sensor faul deecion wih neural neworks A hisorical example of neural neworks for paern recogniion are Probabilisic Neural Neworks (PNN), firs inroduced 1990 b Donald F. Spech[18]. The learning process of his kind of neural neworks is replaced b he generaion of he PNN, because ever sample of he raining daa is ransferred o a neuron and herefore saved inside he resuling neural nework. New samples are classified b calculaing a similari o ever class. The sample is assigned o he class wih he highes similari. The are used for sensor faul deecion in a modified wa b A. Jabbari e al.[19]. His PNN s were applied o monior emperaure sensors in a cold chain for food. Combining curren emperaure measuremens and addiional environmen informaion a, PNN was able o disinguish beween normal saes (emperaure varies due o air exchanges while he door opening) and faul ransducers. Time-Dela Neural Neworks (TDNN) are anoher pe which were applied in common faul deecion b Chrisensen e al. [20]. The paper describes he deecion of hardware fauls for auonomous robo ssems. The consider ha hardware fauls change he flow of sensor daa and also he reacion of he conrol program. These changes are deecable b a TDNN. The deecion mehod is evaluaed in hree differen asks performed b real robos. The srucure of a TDNN is basicall he same as a Muli-Laer-Percepron (MLP). This means here are onl forward connecions. Therefore sandard backpropagaion can be used o learn his kind of neural nework. TDNN are able o deec paerns in ime series b analzing sliding windows of a signal. The lengh of he window mus be saic as a radeoff o he feed-forward characerisic. To overcome he disadvanage of a saic window lengh, Locall Recurren Neural Neworks (LRNN) were applied o faul deecion in [21]. The auhors esed heir ssem on a model consising of hree waeranks. Informaions on differen sensors were available o deec fauls inside he model. Locall Recurren Neural Neworks don have recurren connecions beween neurons bu inside of special neurons. These neurons are called dnamic neurons and have an addiional linear dnamic ssem (LDS) which ransmis he oupu back o he inpu. Wih his recurren connecion, LRNN are able o deal dnamicall wih ime series. Anoher approach wih recurren connecions were inroduced b Hochreier and Schmidhuber in [22]. The did no define a pe of neural neworks, bu a module, called memor cell which can be applied o all neural nework pes. The used differen neurons and recurren connecions in a wa ha memor cells are able o decide which informaion o sore, a which ime and how long o sore his informaion. Furhermore he can decide when o show he sored informaion o he res of he neural nework. These modules could be a par of a neural nework for faul deecion, as i enables he nework o analze ime series dnamicall. An implemenaion of he memor cell concep on sensor faul deecion asks is no known e. Afer inensive lieraure research, we could no find works on faul deecion for single sensor ssems based on neural neworks. Mos works on single sensor seups use addiional informaion abou heir environmen. III. OUR APPROACH The implemenaion of a neural nework requires a number of basic seps. The lef side of Tab. II summarizes he procedure from inpu daa analsis up o raining and evaluaing of a neural nework. We recognize he need for a subdivision while appling he concep on faul deecion asks. The hird column (Index) of Tab. II assigns an index number ha references he following subsecions, where we discuss he inended seps:
Table I CATEGORIES OF FAULTS IN SENSING APPLICATIONS [23]. THE DASHED LINE ILLUSTRATES THE PROGRESS OF A PHYSICAL VALUE. IN CONTRAST, THE SOLID GRAPH DEPICTS THE CORRESPONDING FAULTY MEASUREMENTS. Dela Offse sporadic permanen sochasic Suck-a consan oulier consan consan a zero 1 2 3 4 expeced real 5 variable spike value correlaed value correlaed a X 6 7 8 9 expeced real X 10 omissions / broken link ime correlaed ime correlaed sauraion omissions 11 12 13 expeced real 14 Table II ADAPTATION OF THE COMMON DEVELOPMENT PROCESS FOR NEURAL NETWORKS RELATED TO SENSOR FAULT DETECTION Common developmen Adapaion for sensor faul Index seps deecion Inpu daa analsis Acquisiion of a sample daa base Selecion of a suiable nework pe A. Faul selecion Selecion of relevan faul pes Definiion of an environmen model Derivaion of appropriae feaures Generaion of measuremen samples superimposed b seleced fauls Weighing of neural approaches considering he specific seup Train he nework Evaluaion For each faul deecion applicaion we need an appropriae faul model considering all possible deviaions beween real values and sensor measuremens. A comprehensive classificaion of sensor daa cenric faul models is given in [23]. Fig. I illusraes he major pes organized in relaion o he correlaion, duraion or ampliude characerisic. We disinguish 14 differen faul pes such as ouliers, offses or addiional noise. The menioned sensor descripion involves an idenificaion of he relevan sensor pe and heir mahemaical parameer. A he momen, we jus consider 4 faul pes : ouliers (2), consan offse (3), noise (4) and Suck-a-Zero (5). The selecion covers he faul characerisic of man sensing devices. Hence, we wan o evaluae our ideas based on his subse and inegrae oher faul pes laer. A B C D E F G B. Environmen Model Addiional o he faul model, we have o consider an appropriae environmen model describing he applicaion conex of he sensor. The environmen model characerizes he non-faul sae and defines he ranges of he measuremen value as well as he dnamics of he moniored ssem. In order o use neural neworks, he environmen model has o be described implicil b samples in he daabase/raining daa. Hence, creaing a mahemaical environmen model is replaced b generaing/collecing daa for raining a neural nework. C. Relevan feaures Preprocessing is one par of inpu daa analsis which can increase performance of neural neworks dramaicall. In faul deecion preprocessing is also called feaure exracion. Feaures [10] are addiional informaion calculaed from raw measuremen signals. Therefore during his sep he developer has o choose he relevan feaures as inpu of he neural nework. One main challenge is o idenif he bes composiion of available feaure relaed o he fauls we wan o deec. To decide which feaure o use, some requiremens are given in [24]. The should be:. compuable efficienl. uncorrelaed wih oher feaures. independen of exernal influences. characerized b high differences beween feaures and small inernal differences. Along wih hese condiions, we look for a minimal number of feaures so ha all fauls considered in he faul model are covered. The feaures have o decouple sensor measuremens and faul deecion mehods. These absracion guaranee he applicabili of he approach on a wide range of sensor pes. We seleced he following feaures ha will be used as an inpu for our neural nework. The variable x represens measuremen samples, x he mean value and T he sliding window lengh. Mean Calculaing mean enables our ssem o recognize
ime-correlaed fauls. Furhermore a mean value idenifies on usual values of a signal. If he curren value differs srongl from mean, a faul measuremen can be assumed. We compue he mean value over a sliding window wih lengh T. Hence, mean is defined b E(x ) = x = 1 T 1 T τ=0 x τ Sandard-Deviaion The Sandard Deviaion quanifies he widh of a probabili disribuion and defines he expeced deviaion of a measuremen relaed o he mean. For parameric disribuion funcions we can calculae he probabili of he curren difference from mean. Sandard Deviaion is defined for a sliding window wih lengh T b: Deviaion s(x ) = 1 T T (x τ x ) 2 τ=0 The firs deviaion reflecs he dnamic of he observed ssem. The value allows a neural nework o recognize ouliers, spikes ec. The deviaion can be calculaed b: dx d = x x 1 Signal-o-Noise Raio The Signal-To-Noise-Raio (SNR) allows o esimae he noise level of a signal. In lieraure i is ofen defined as signal power divided b noise power [25]. However, o compue an running SNR we appl he following definiion SNR(x ) = E(x) s(x ) Correlaion-Coefficien The Correlaion-Coefficien describes he similari of wo signals. Therefore he correlaion-coefficien is defined as [25]: r x = E(x ) E(x) E() s(x ) s( ) Furhermore he Correlaion-Coefficien allows o derive a funcional relaion beween o signals. As he coefficien is in he [-1,1] inerval, i can be inerpreed as: r x > 0: high values in x ield high values in. r x < 0: high values in x ield low values in. r x = 0: x and are no correlaed. r x = 1: x, are linear correlaed: = a x + b; a > 0. r x = 1: x, are linear correlaed: = a x + b; a < 0. Addiional ransformaions of signal like Fourier-Transformaion and power-densi specrum are possible feaures ha need o be invesigaed. Tab. III maps our faul models of Tab. I on he menioned feaures. The abular evaluaes he deecion capabiliies of individual Table III HYPOTHETICAL RELEVANT FEATURES FOR SENSOR FAULT CATEGORIES ID Faul caegor Mean Variance Deviaion Corr.-Coefficien Sig.-o-N. Raio 1 Consan Dela 2 Oulier 3 Consan Offse 4 Consan Noise 5 Suck-A-Zero 6 Variable Dela 7 Spike 8 Value-Correlaed Offse 9 Value-Correlaed Noise 10 Suck-A-X 11 Omission 12 Time-Correlaed Offse 13 Time-Correlaed Noise 14 Sauraion feaures regarding ouliers, offses, ec. The assignmen represens a firs hpohesis ha has o be proven in fuure work. For his paper we chose 4 faul pes which are deecable b our feaure se. In furher invesigaions we will exend his number. D. Generaion of daa samples As he neural nework will implicil generae an environmen model during learning, he colleced samples need o represen he daa produced b he final wokring ssem. One possibili o generae hese samples is o se up a real sensor ssem and collec he measuremens. In a pos processing sep all measuremens have o be (manuall) classified as faul or correc. However his approach has wo major drawbacks. Firsl, some fauls occur ver seldom or depend on specific environmen condiions. Hence, daa acquisiion needs o run for a long ime o capure hem. Secondl, he manual classificaion is an exensive work especiall for large daa ses. Alernaivel, a faul injecion framework handling all relevan faul pes is more effecive. In his case a schedule of fauls o happen in he simulaion is used as inpu. The he faul injecion ool creaes he daa based on a simulaed ssem behavior and a defined faul characerisic for each faul in he schedule. Anoher possibili is o emplo he faul injecion on real sensor measuremens. However possible real measuremens fauls compicae he classificaion in his approach. E. Neural nework pe Feaure exracion is a preprocessing sep of faul deecion. Hence, we have o implemen an appropriae deecion mehod, ha is able o cover all relevan faul pes b one approach, namel neural neworks. The pe of neural nework is a basic parameer. As shown in Sec. II here are man pes of neural neworks which could be suiable for sensor faul deecion. As i is a known problem of recurren neural neworks o learn hese, for his work we concenrae
on feedforward neural neworks. Time-Dela Neural Neworks are one possible choice. The are eas o rain bu can even analze ime-series in a bordered range. Besides he pe of neural nework we have o define oher parameers, e.g., he number of hidden laers and he number of neurons in ever hidden laer. This is picall done b crossvalidaing differen neural nework srucures. Choosing he bes number of laers, special aenion should be paid o he curse of dimensionali [16]. This phenomenon describes differen aspecs of high dimensional daa. For insance, wih an increasing number of dimensions, he disances of neighboring poins increases oo. Thus, wih ever addiional hidden laer ha maps inpu daa in a higherdimensional feaure space, his effec makes i harder o esablish suiable separaing hperplanes. Anoher aspec of his phenomenon, and a burden for an machine learning mehod, is he fac ha we need an exponeniall growing number of daa poins o esimae he parameers of addiional hidden laers properl [16]. Thus a higher number of hidden laers poeniall decreases he generalizaion performance of he nework. F. Training of he neural nework The las imporan parameer is he raining funcion and is parameers. ha can deermined b differen approaches. A number of algorihms applies he gradien descen, as he Broden-Flecher- Goldfarb-Shanno(BFGS) algorihm, a quasi-newon backpropagaion algorihm, which is an derivaion of sandard backpropagaion [26]. This algorihm is similar o he Newon Mehod, bu compues he derivaions onl approximael which increases he speed of raining. Unforunael i has o save he Hessian marix which is n n where n is he number of weighs and biases used in he neural nework. For large neworks his marix will increase quadraicall. One benefiing parameer of his raining funcion is he opional weigh deca parameer. This parameer is used while learning o force he weighs of unused neurons o zero. The number of neurons can be defined in a wider specrum and he decision on how o choose he number of neurons per laer is easier. The value of his parameer can be examined wih cross-validaion again. Anoher imporan parameer of ever raining funcion is he goal of performance. The raining of neural nework will sop, when his value has been reached. Zero isn a good goal for raining a neural nework, because of generalizaion problems and overfiing. To avoid overfiing, an appropriae number of epochs has o be chosen. G. Evaluaion and validaing he neural nework Evaluaing a neural nework requires an error funcion. Known varians are mean squared error, sum squared error or mean absolue error. As we wan o perform an classificaion-ask, a more appropriae error funcion is needed. Precision and recall are commonl used in his conex: Precision : p p + f p and Recall : p p + f n where p denoes all faul samples ha were classified correcl (rue posiives). f p references false posiives, whose classificaion is wrong b assuming a faul. The counerpar is marked b f n, faul samples ha were classified as faul free. Therefore precision is an indicaion of how much of he reurned fauls are correc classificaions. Recall onl indicaes how man of he occured fauls are deeced. Boh values can be combined o he so-called F-Score. There are differen approaches for his combinaion, bu commonl he F1-Score is used. Volume l 8 6 4 2 0 0 200 400 600 800 1000 Time Figure 1. Example of a faul free sensor measuremen; The heoreical model (red) and he real ssem behavior wih a nois random ouflow (blue) F 1 = 2 precision recall precision + recall In he end, he assessmen of a neural nework can be quanified b a single value. This concenraed represenaion suppors our goal of an auomaed configuraion process of a faul deecor. IV. EVALUATION We evaluaed our approach based on he commonl used waer ank example. I provides an inake and a drain. The inflow is consan over ime, while he ouflow depends on he curren heigh of he liquid inside he waer ank. Addiional we inegrae a random ouflow wih a specific mean, variance and lengh. The second ouflow is riggered periodicall and represens he uncerain of our environmen model. Fig. 1 shows an example of a faul free sensor run. Our waer ank model and a level sensor are simulaed. I is assumed ha jus one sensor faul ma occur a one poin in ime. As described in secion III-A we consider onl 4 faul pes. The faul sae is calculaed in a faul injecion framework and applied on he simulaed sensor measuremens from he waer ank model. This guaranees, ha onl one faul can occur a a ime. B simulaing his seup wih differen faul parameers we generae samples and arges for learning and validaion. Beside he raw sensor measuremens also he feaures (mean, sandard-derivaion, gradien, Signal-o-Noise- Raio and Correlaion-Coefficien) were saved. If a window lengh was needed o calculae a feaure, we defined T = 64. We chose a Time-Dela Neural Neworks and applied he BFGS algorihm for learning. Furhermore TDNN are able o analze ime series b presening ime slos. In his case, he nework has o analze six sliding windows in parallel, ever window represening one feaure. As he inpu of he TDNN will be a sliding window of a sensor measuremen, is oupu should be a sliding window oo. Inside his window ever ime sep is marked as faul(1) or no(0). We rained TDNN wih 2 hidden Laers. In order o find he bes srucure 3 neural nework wih differen numbers of neurons per laer [25, 50, 75] were creaed. All neworks were rained wih he BFGS algorihm. We used he mean squared error as an error funcion. As cross validaion give he bes regularizaion values beween 0.04 and 0.06 all nework srucures were rained wih a regularizaion parameer of 0.04, 0.05 and 0.06. The ransfer funcion of ever neuron in a hidden laer is he hperbolic angen. To compare he resuling neural nework wih common mehods, we implemened a Limi Checking (LC) faul deecor for ever faul. Ouliers were deeced b checking he gradien. If he curren sensor value is equal o zero, an Suck-a-Zero faul is deeced. To deec a Consan Offse, he mean was checked and Noise was recognized b evaluaing he curren Signal-o-Noise raio. Wih precision and recall we were able o compare boh mehods. We used 25% of he daabase o esimae his values, 75% were used o rain he TDNN. In Table IV
he resuls of comparison are shown. The firs value is he precision of he specific faul deecor assigned o he deeced faul. The second value is he corresponding recall. As LC is implemened onl for one faul pe a a ime, he - -sign indicaes ha no measuremens of precision or recall for he oher faul pes were produced. As he reader can see, TDNNs were no able o deec specific fauls more reliable, bu are coequal in precision and recall. Fauls like Sucka-Zero appeared o be recognizable eas, agains fauls like Oulier. Oulier were onl deecable wih a precision of 0.1 and recall of 0.02. This could be caused b he random ouflow of he waer ank. The blue line in Fig. 1 shows an exemplar sensor measuremen wih a nois random ouflow(e.g. from = 100 o = 180). Timeseps of his random oulow seems o be equal o Ouliers. Therefore he disance beween faul free imeseps and Oulier is ver small(he euclidean disance is someimes less han 0.001), so ha he neural nework canno disinguish beween faul and faul free daa poins. LC of he gradien produces onl a precision of 0.34 and a recall of 0.39 oo. Anoher problemaic faul is he Consan Noise. The TDNN reached onl a precision of 0.47 and a recall of 0.4. As he LC obains a recall of 1 bu also a precision of 0.48 ma he Signal-o- Noise raio is no a appropriae feaure o deec Noise. Neverheless Suck-a-Zero and Consan Offse could be deeced reliabl. Hence, we were able o creae one neural nework o deec differen fauls. Table IV COMPARISON OF DIFFERENT COMMON FAULT DETECTION METHODS WITH A TDNN, (PRECISION/RECALL) TDNN Faul Deecion Mehod Gradien Limi Checking on Mean Signalo-Noise Raio raw signal Oulier 0.1/0.02 0.34/0.39 - - - Consan Offse 1/1-1/0.97 - - Consan Noise 0.47/0.40 - - 0.48/1 - Suck-A-Zero 0.98/0.99 - - - 0.99/1 V. CONCLUSION As one milesone owards an auomaic configuraion of a faul deecor our goal was o show ha neural neworks provide he capabiliies for muli faul deecion on single sensor ssem. Therefore we rained a Time-Dela Neural Nework o deec four differen faul pes. Two of hem were reliabl deecable, Noise was deeced sporadic and Oulier could no be recognized. Neverheless one neural nework was able o deec more han one faul pe. We will coninue our work b esing oher neural nework pes, such as neural neworks wih recurren connecions in order o improve resuls. One promising approach are he Long-Shor Term Memor Cells [22], which are no applied o faul deecion unil now, bu seems o provide promising capabiliies. Furhermore work on how o choose feaures and which feaure can be used o deec specific fauls could increase success rae of faul deecion oo. Wih greaer knowledge abou feaures, selecion of hese will become more auomaable. ACKNOWLEDGMENT This work was pariall suppored b he EU under he FP7-ICT programme, hrough projec 288195 Kernel-based ARchiecure for safey-criical conrol (KARYON). REFERENCES [1] L. Azori, A. Iera, and G. Morabio, The Inerne of Things: A surve, Compuer Neworks, vol. 54, no. 15, pp. 2787 2805, 2010. [2] J. Shi, J. Wan, H. Yan, and H. Suo, A Surve of Cber-Phsical Ssems, in Wireless Communicaions and Signal Processing (WCSP), 2011 Inernaional Conference on, pp. 1 6, IEEE, 2011. [3] P. Bellavisa, A. Corradi, M. Fanelli, and L. Foschini, A Surve of Conex Daa Disribuion for Mobile Ubiquious Ssems, ACM Compuing Surves, vol. 45, no. 1, pp. 1 49, 2013. [4] S. Zug, M. Schulze, A. Dierich, and J. Kaiser, Programming absracions and middleware for building conrol ssems as neworks of smar sensors and acuaors, in Proceedings of Emerging Technologies in Facor Auomaion (ETFA 10), (Bilbao, Spain), 9 2010. [5] H. Pionek, Self-descripion mechanisms for embedded componens in cooperaive ssems. PhD hesis, Universi of Ulm, 2007. [6] M. van der Meulen, On he use of smar sensors, common cause failure and he need for diversi, in 6h Inernaional Smposium of Programmable Elecronic Ssems in Safe Relaed Applicaions, (Cologne, German), 5 2004. [7] T. Brade, M. Schulze, S. Zug, and J. Kaiser, Model-Driven Developmen of Embedded Ssems, in 12h Brazilian Workshop on Real-Time and Embedded Ssems (WTR), (Gramado, Brazil), Brazilian Compuer Socie, 5 2010. [8] S. Ding, Model-based faul diagnosis echniques: design schemes, algorihms, and ools. Springer Verlag, 2008. [9] G. Vachsevanos, F. Lewis, M. Roemer, A. Hess, and B. Wu, Inelligen faul diagnosis and prognosis for engineering ssems. Wile, 2006. [10] R. Isermann, Model-based faul-deecion and diagnosis saus and applicaions, Annual Reviews in conrol, vol. 29, no. 1, pp. 71 85, 2005. [11] B. Hardekopf, K. Kwia, and S. Upadhaa, Secure and faul-oleran voing in disribued ssems, in Proceedings of IEEE Aerospace Conference, vol. 3, 2001. [12] J. E. Whie and J. L. Speer, Deecion filer design: Specral heor and algorihms, IEEE Transacions on Auomaic Conrol, vol. 32, no. 7, pp. 593 603, 1987. [13] A. Savizk and M. J. E. Gola, Smoohing and differeniaion of daa b simplified leas squares procedures., Analical Chemisr, vol. 36, pp. 1627 1639, Jul 1964. [14] A. Zolghadri, An algorihm for real-ime failure deecion in kalman filers, IEEE Transacions on Auomaic Conrol, vol. 41, no. 10, pp. 1537 1539, 1996. [15] R. Paon, Faul-oleran conrol ssems: The 1997 siuaion, in IFAC smposium on faul deecion supervision and safe for echnical processes, vol. 3, pp. 1033 1054, 1997. [16] T. Hasie, R. Tibshirani, and J. Friedman, The Elemens of Saisical Learning. Springer, 2003. [17] Y. Sheng and S. Rovnak, Decision ree-based mehodolog for high impedance faul deecion, IEEE Transacions on Power Deliver, vol. 19, no. 2, pp. 533 536, 2004. [18] D. F. Spech, Probabilisic neural neworks, neural neworks, vol. 3 Issue 1, pp. 109 118, 1990. [19] A. Jabbari, R. Jedermann, and W. Lang, Applicaion of compuaional inelligence for sensor faul deecion and isolaion, Proceedings Of World Academ Of Science, Engineering And Technolog, vol. 22, pp. 503 508, 2007. [20] A. L. Chrisense, R. OGrad, M. Biraari, and M. Dorigo, Faul deecion in auonomous robos based on faul injecion and learning, Auonomous Robos, vol. 24, pp. 49 67, 2008. [21] P. Przsalka, Model-based faul deecion and isolaion using locall recurren neural neworks, Lecure Noes in Compuer Science, vol. 5097, pp. 123 134, 2008. [22] S. Hochreier and J. Schmidhuber, Long shor-erm memor, Neural Compuaions, vol. 9, pp. 1735 1780, 1997. [23] S. Zug, A. Dierich, and J. Kaiser, Faul Diagnosis in Roboic and Indusrial Ssems, ch. Faul-Handling in Neworked Sensor Ssems. S. Franklin, AUS: Concep Press Ld., 2012. [24] J. C. da Silva, A. Saxena, E. Balaban, and K. Goebel, A knowledgebased ssem approach for sensor faul modeling, deecion and miigaion, Exper Ssems wih Applicaions, vol. 39 Issue 12, pp. 10977 10989, 2012. [25] M. Meer, Signalverarbeiung: Analoge und digiale Signale, Sseme und Filer. Springer Vieweg, 2011. [26] P. Werbos, Beond Regression: New Tools for Predicion and Analsis in he Behavioral Sciences. PhD hesis, Harvard Universi, 1974.