An Automated Auto-encoder Correlation-based Health-Monitoring and Prognostic Method for Machine Bearings

An A utomated A uto-encoder Corr elation-based Health-Monitoring and Pr ognostic Method for Machine Bearings Ramin M. Hasani, Guodong W ang and Radu Gr osu Cyber -Physical Systems Group, V ienna Uni versity of T echnology , Austria ramin.hasani, guodong.wang, radu.g ros u@tuwien.ac.at Abstract This paper studies an intelligent ultimate techniqu e for health -monitor ing and p rogno stic of common rotary ma c hine compon ents, particular ly bear ings. During a run-to- failure experiment, rich un super- vised features fr om v ibration sensory d a ta ar e ex- tracted by a trained sparse auto-enco der . Then, the correlation of the extracted attributes of the initial samples (presum ably healthy at the begin- ning o f the test) with the succeeding samples is calculated and passed throu gh a moving-average filter . The normalized output is named auto- encoder corr elation-ba sed (AEC) r a te which stand s for an infor mativ e attribute of the system depict- ing its health status an d precisely identify ing the degradation starting po int. W e show that AEC technique well-generalizes in several run- to -failure tests. AEC collects rich unsupervised features form the vib ration data fu lly auto nomou s. W e d emon- strate th e super iority of the AEC ov er many other state-of-the- art appro aches for the health mo nitor- ing and p rogno stic of machine bea r ings. 1 Intr oduction In almost all industries health m anagemen t of ma chines is notably e ssential. A key subsidiary of h ealth m anagemen t is condition -based monitoring ( CBM) where one pro gnoses ab- normal statu s of a mac hine based on extracted f eatures f rom a gro up of imp lemented sensors an d pa rameters. Th e CBM proced u re the r efore includes two steps; 1) Feature extraction during a run-to - failure e xperiment and tw o 2) Data p rocessing for p redicting th e d egradation starting point and monitoring the defect p ropagatio n during the test. Quite a large n umber of methods have been p roposed for the pr o gnostic o f ke y ma c h ine co m ponen ts. In many cases, hand crafted time- and frequency-do main features a r e derived f rom the sen sors mounted o n th e machine a nd are used directly or p ost-proc e ssed by numer ous meth ods in ord er to predict the c o ndition o f the machin e com po- nent [ Lee et a l. , 2 014 ] . Recently , on th e other hand, artifi- cial intelligence (AI) solutions have be e n vastly utilized in fault classification and con dition mo n itoring [ Jia et a l. , 2016; Sun et a l. , 20 16; Thiru kov alluru et a l. , 2016 ] . AI techniques significantly enh ance the quality of the feature extraction and data processing. Howe ver , they have no t yet realized a fully- automated m ethod without the u se o f prior knowledge for health monitoring tests. The main attributes of an ideal h ealth-con d ition monitorin g and p rogno stic method are descr ibed as follows: The method is able to collect unsupervised u seful features from the a v ail- able sensory data, regardless of its size, auton o mously . Such data-driven procedure sho uld en able the user to pe rform mo n- itoring in both online an d offline tests, while providing an in- telligent trend on the health status of the system and precisely pointing out the degradation starting point. It is high ly de- sirable that the method au tomatically combines the steps of the CBM process and be entirely hum an-labo r indepen dent. Moreover , the ideal approach is univ ersal and can be feasibly applied to th e progno stic of various key ma chine comp onents such as b e arings, gears and spindles. In the present st udy , we pro pose a novel p rogno stic method for m achine bearings, as a key mac hine co mpone n t which sat- isfies the main chara c teristics of an ideal CBM method. The technique is called auto-enc o der-correlation- based (AE C) progn ostic a lgorithm. W e train a spar se auto- encoder for ex- tracting unsuperv ised features from collected sensory data in se veral test-to-failure experiments and co rrespon dingly , com- pute the Pearson correlation of th e extrac ted featu res of the initial sam ples, with the u pcomin g samples. The o utput is then passes thr ough a m oving a verage (MA) filter . AEC al- gorithm then n ormalizes th e output o f the filter an d accurately illustrates the health condition of the system. W e ev aluate the perfor mance o f our alg o rithm over several run-to-failure tests of machine bearing s and prove its superior ity in finding th e degradation starting point compared to the existing metho ds. According ly , the paper is structure d as follows. Section 2, d e scribes the-state of -the-art method s emp loyed fo r the CBM and progn ostic. In Section 3, we delineate the de- sign proced ure of the AEC a lgorithm while recapitulating the sparse auto-encoder architectu re and describing the re st of the method. In Section 4, w e p erform v arious experiments with AEC and repr esent our results cor r espondin gly . W ithin Sec- tion 4, we qu alitativ ely a n d qualitatively com pare our method to the other appro aches. W e conclude our rep ort in Sec tion 5 . 2 Related W orks Comprehe n si ve reviews o n the usefu l p rogno stic methods for industrial mach ines includ ing b e a rings have been pro- posed [ Jardine et al. , 2006; Lee et al. , 2014 ] . T raditionally , time domain statistical features such as r oot mean squared (RMS), Kurtosis, Spectral Kurtosis and so on, have been utilised for mo nitoring th e statu s o f the machine key com- ponen ts [ Qiu et al. , 2 006 ] . Howe ver such fe a tu res fail to pro- vide u seful inf ormation o n the status of the system in so m e cases and therefo re are no t g eneralizab le . In this regar d , se- lection o f the prop er statistical features wh ich con tain use- ful inf ormation , is a challenge which is addre ssed in many health mon itoring method s [ Y u, 2012a; Y u, 2012b ] . Such ap - proach e s lack au to mation as well as generalization. Moreover , unsuperv ised feature extraction techniq ues such as principle component analysis (PCA)-b ased method s [ He e t al. , 2 007 ] followed b y a post-pro c essing stage such as hidden Markov m odel (HMM) for h ealth degradation mon- itoring pr ovided a reasonable predic tio n on the status of the system [ Y u, 2012a ] . Despite their attractive le vel of accuracy , such metho ds ar e also no t automated and th eir genera liza - tion ability has no t yet clearly shown. Mo reover , combination of the f requen cy-domain feature extraction meth o ds such as wa velet p acket d ecompo sition with HMM are also emp loyed in the calculation o f the remaining u seful life (RUL) of th e key mach ine co mpone nts [ T obon-M e jia et al. , 2012 ] . T obon et al. pro posed a useful trend on the R UL and successfully tested it o n few b earing run- to-failure experiments. The ap - proach is however , subjected to modification s to be utilized in the o ther test-c a ses [ T obon-M e jia et al. , 2012 ] . Kalman filter ( KF) is also used in c ondition m onitorin g [ Reuben and Mb a, 20 1 4; W ang et al. , 2 016 ] . Reu ben e t al. propo sed a method based on switching KF which can be simultaneou sly em ployed in the diag nostic and p rogno stic of the data [ Reuben and Mb a, 2014 ] . W ang et al. used an enhanced KF with expectation- maximization algorithm fo r providing accurate estimations on the health con d ition of the machines [ W an g et al. , 2016 ] . Such method s lack the fully automation prope r ty and are subjected to manual pre- processing. Data-driven approaches using AI techniqu es, have brought great advancements to the diag nostic and health co ndition monitorin g of key machine co mpone n ts. V arious architec- tures of neural networks in particular auto-encoder s, are suc- cessfully employed for precise-classification of faults in to different su b group s, by using pre-pro cessed statistical time or frequen cy d omain features and in g e n eral feature s extracted using prior knowledge [ Y ang et al. , 2002; Jia et al. , 2016; Sun et a l. , 20 16; Thirukov alluru et al. , 20 1 6 ] . Although p r o- posed meth ods are n ot fully au tonomo us, th ey ar e less human- labor dep endent. Howe ver , such methods have n ot been yet utilized in th e pro gnostic and con dition monitoring fully autonom ous. Here, for the first time , we train a sparse auto-enc o der d irectly over the v ib ration data and compute the correlation of th e extracted f eatures and th e refore, p r ovide a comp rehensive prognostic and health mo nitoring method which its per forman ce is superior com pared to many o f the remarked app roaches. 3 A uto-encoder Correlation-based (AEC) Fault Prognostic Method In this section we d e scribe the mech a nism of the au to - mated AEC gener al fault p rogno stic metho d d uring the run - to-failure test of machine bearings. Figu r e 1 gr aphically il- lustrates the structure of the AEC. An autoenco der network is trained by using th e V ibra tion d a ta-samples. The auto- encoder generates rich no nlinear features from the sensory data f or ea c h samp le. Afterwards, th e cor relation co efficient matrix of the past samples until the curren t sample, is com- puted. The correlation co efficients of the feature s extracted for the samples generated at the beginning of the run-to- failure pro cess, with the other available samples is normalized and correspondin gly filtered by means o f a movin g a verage (MA) filter . This provides an ou tput r ate which can pred icts the status of the system at each sampling time step. I n the followings, we sketch our design procedures, in details. 1 x 1000 1 x 20480 1 x 20480 Correl ati on Coeff ici ent m x m m input samples 1 x m 1 x m Norma li zed Movi ng Aver age Syste m Stat us Encoding De coding Figure 1: The architecture of the AEC fault prog nostic method. 3.1 Sparse A uto -encoder Revisit An auto-en coder [ Boureau et al. , 2008 ] , tr ies to learn an ab- stract of th e identity fu nction, to estimate the same input p at- terns at its output. On e can place con stra ints on the ne twork by limiting the size of the h idden layer and presumab ly dis- cover attractive features form the data. Let us define x ∈ R D x , a D x -dimension input to the auto- encoder ; the encoder initially maps x to a lower dimension vector z ∈ R D , and corre sp onding ly generates an estimation of x , ˆ x ∈ R D : z = f ( W x + b 1 ); ˆ x = f ( W T z + b 2 ) , (1) where f , in our design , is a saturating linea r transf e r function denoted in Equ ation 2 , W ∈ R D x × D stands for the weight matrix, and v ectors b 1 ∈ R D , b 2 ∈ R D x represent the bias values. f ( z ) =    0 , if z ≤ 0 z , if 0 < z < 1 1 , if z ≥ 1 (2) W e define th e f ollowing cost fun c tion to be optimized sub- sequently: C = 1 N N X n =1 I X i =1 ( x in − ˆ x in ) 2 | {z } mean squared error + λR L 2 | {z } L 2 regularization + σ R sparse | {z } sparsity regularization . (3) The first term in C is the mean squared err or . The sec- ond term denotes an L 2 regularization ap plied f or prev enting over -fitting, where λ is the regularization co efficient. L 2 reg- ularization is co mputed as follows: R L 2 = 1 2 n X j k X i ( w j i ) 2 , (4) where n and k are the num ber of samples(o bservations) and number of variables in the training data, respec ti vely . The third term in the cost function determine s an sparsity regularzation with an effectiv eness coefficient, σ . Let the a v- erage output activ ation o f a neu ron i be denoted as: ˆ ρ i = 1 m m X j =1 f ( w T i x j + b i ) , (5) and define ρ a s the desired o utput value of a neur on, One can measure th e ir difference by using th e Kullback-Leibler div ergence function described b elow: R sparse = D X i =1 K L ( ρ k ˆ ρ i ) = D X i =1 ρ log( ρ ˆ ρ i ) + (1 − ρ ) log ( 1 − ρ 1 − ˆ ρ i ) . (6) KL function ou tputs zero when ρ and ˆ ρ i are clo se to each other an d increases when they diverge fr om each othe r . Th is implies sparsity in the network and it is defined in our cost function as the sparsity regularization, R sparse . W e then tra in the network by applyin g a scaled conjugate gradient (SCG) alg orithm [ Møller, 1993 ] using MA TLAB’ s Neural Network T oolbox . 3.2 Corr elati on Analysis, Normalization and Filtering For each training sample, D non linear fea tures are con- structed b y the Au to-enco d er . W e calc u late the linear de- penden cies of the abstract rep resentation of each sam p le by utilizing the Pearson correlatio n coef ficient ( CC) which is de- scribed bellow for tw o samples A and B : r ( A, B ) = 1 D − 1 D X i =1 ( A i − µ A σ A )( B i − µ B σ B ) , (7) where µ A and σ A are the mean and the stand ard deviation of A , respectively , and µ B and σ B are the mean and stand ard deviation of B [ Fisher , 1 925 ] . The correlation coefficient ma- trix f or the a vailable samples du ring the run-to - failure test is calculated. The first colum n of the CC matrix d epicts the cor- relation of the first sam p le data which is reco r ded at the be- ginning of the r un-to-failur e process, with the other a vailable samples. W e norm alize this vector b e twe en zero and one, and define it as th e criteria for predicting th e faulty s ample and consequently determ ine the health-status of the system. W e finally , smo othen the shap e of the ou tput by passing it through a moving average filter . Th e filter is designed a s follows: ˆ m = 1 w size ( y ( n ) + y ( n − 1) + · · · + y ( n − ( w size − 1))) . (8 ) For the sam p le data y , the filter slides a sample-window of length w size , over the d a ta, and calculates the average of th e covered data in each window . 4 Experiments with AEC In th is section we ev aluate the p erforma n ce of our fault prog- nostic method by employing it in several ru n-to-failure exper- iments on bearing s. W e initially introduce the dataset con- tents together with the o b jectiv e o f the tests, and illustrate the performance of the AEC meth od in various run -to-failure tests. W e finally benchma rk o u r results with the state-of-the - art methods in fault prognostic of the bearing mac h ines. 4.1 IMS Bearing Data S et fr om P CoE NASA Datasets W e u se th e b earing dataset provided by the the Center for In- telligent Main tenance Systems (I M S), University of Cincin- nati [ Lee et a l. , 2 007 ] , co llected from the Progn ostics Data Repository of NASA [ PCoE, Accessed 201 6 ] . The b e aring test rig is shown in Figu re 2. A shaft is coupled to a AC motor and is r otating at 200 0 RP M while a 60 00 l bs load is installed on it. Four for c e-lubricated bearing s are mou nted on the shaft. Accelerometer s with high sensitivity are placed for each bear- ing for recording the v ibrations (the position o f the sensors are shown in the Figure 2). Th ree test-to-failure experiments Figure 2: Bearing test implementation [ Qiu et al. , 2 006 ] . Four bearing s are set-u p o n a shaf t and vibration data is co l- lected in three run-to -failure tests. are performed ind ependen tly . In such tests, failures usually happen ed at the e nd o f the test [ Qiu et al. , 2 006 ] . In the first exper iment, two accelerometers are utilized for each b earing while in the s econd an d third experiments one accelerometer is used. Datasets contain 1-second record- ings from the accelerometers with a sampling fr equency of 20 K H z , every 10 min , during the run -to-failure tests [ Qiu et al. , 200 6 ] . T able 1 r epresents the prope r ties of the collected data in each experiment. T able 1: I MS Bearing tests specification Experiment # of Samples sample size Faulty Bearing t est-to-fail ure time S1 2156 4 × 20480 B3 and B4 35 days S2 984 4 × 20480 B1 8 days S3 4448 4 × 20480 B3 31 days For o ur simu lations, we only train the AE network on the faulty bear ings therefo re we will h ave four d ifferent simulate d experiment cases as follows: 1) Dataset 1 b earing 3 (S1B3), 2)Dataset 1 Bearing 4 (S1B4), 3) Dataset 2 bear ing 1 (S2B1) and 4) Da ta set 3 Bearing 3 (S3B3). 4.2 Results W e train the auto-enco der an d consequently implement our progn ostic method in MA TL A B due to its comp atibility with most of the industrial production systems and specially online test-to-failure en vironme ntal tools. The training process is perfor med o n a Microsoft Azure NC-Series virtual machine powered by one NVIDIA T e sla K80 GPU. W e demonstrate our method ’ s func tio nality in tw o ge n eral framew orks where in the first one we train the auto-enco d er with all the av ailable data, for each experiment, in or der to m onitor the status of the system. In the second fr a mew ork, we train the auto-enco der with 70% of the data a n d keep 30% of it for testing in or d er to chec k the prediction po wer of the pr oposed metho d . Th e training time varies between 55 min and 80 min fo r each case, depend ing on the number of input sam p les. Since we o nly consider the vibration samples of on e bear- ing in each simu lation, the inp u t samples’ dimension, u nder 20 kHz sampling frequen cy , is a 20480 -length vector . W e choose 1000 hidden units for the AE wh ich enables us to ex- tract 1000 featu res from each large inpu t vector . W e then calcu late the co rrelation coefficient matrix of the input samples (dep ending on the framework in which we are working on, fo r the first case we feed in all the a v ailable data while for the second framew ork we de d icate 70% of the d ata for training an d the network has to output a pr ediction on the status of the system based on the previously obser ved sam- ples) and no rmilaze the correlation coefficient of the initial sample ( which is considered to rep resent the healthy status of the sy stem) with the n ext samples. W e then filter out th e output an d pr ovid e a repr esentation of th e statu s of the sys- tem. Framework1 - Health-Co ndition Monitoring Figure 3 illustrates the output of the AE C for the f our simula- tions of the three run-to-failure experiments, where we mon- itor the r ecorded data. A high AEC rate correspo nds to a 5 10 15 20 25 30 35 Time (days) 0 0.5 1 Normalized AEC rate S1-Bearing 3 S1-Bearing 4 S2-Bearing 1 S3-Bearing 3 Figure 3: AE C e v aluation rate for four different d atasets. For ev ery dataset, our AE C platform g enerates a rich tr e nd co rre- sponding to the statu s of the bearing based on the vibration sensory data. One can feasibly d etect abr upt changes in the AEC rate. The v ertical dot lines correspo nds to the sam ple time after which AEC platf orm start decreasing d ramatically . This indicates the degradation startin g point. 500 1000 1500 2000 2500 3000 3500 4000 0 0.5 1 100 200 300 400 500 600 700 800 900 0 0.5 1 200 400 600 800 1000 1200 1400 1600 1800 2000 0 0.5 1 200 400 600 800 1000 1200 1400 1600 1800 2000 0 0.5 1 Sample A B C D 250 450 650 850 1050 1250 1450 1650 1850 2050 250 450 650 850 1050 1250 1450 1650 1850 2050 150 250 350 450 550 650 750 850 950 720 1220 1720 2220 2720 3220 3720 4220 Figure 4: V isualizatio n of the status of fou r different bearings in four run-to- failur e experiments. T h e colo r bar represents the value of the norm a lize d AEC ra te . The h igher is the r a te (the more r ed is th e color bar) the less p robab le it is for the system to be in a faulty condition. The ho rizontal axes shows the samples collected for the run-to-failure test in every ex- periments. A) S1B3, AEC rate record ing. B) AEC for S1B4 C) AEC f or S2B1 . D) AEC fo r S3B3. healthy beh avior of the system wh ile a downward trend de- picts star tin g po int of an ab norma l state. Hig h AEC indicates more co r related samples with the initial healthy status of the system. AEC clear ly displays the beginning of a faulty trend together with its prop a gation effect. It can also de m onstrate a health- characteristic portfolio for the noisy dataset, S3B3 for which many oth er appro aches hav e failed to provide a health-degrad ation characteristics. Figure 4 graphica lly in- dicates the status of the four experiments S1B3 , S1B4, and S2B1 and S3B3 in parts A, B, C and D r e spectiv ely . The Color bar represents th e AEC rate from 0 (blue ) to 1 (red). W e can distinctly o bserve where a significan t ch ange in the rate is o ccurred and ther e fore stop th e p rocess. W e c a ll a sample ab normal, wh en its AEC rate is measu red 90% belo w the reco rded sample at t − 100 . The AEC delicately indica tes the start of a faulty status and represents its gradual propa- gation. T able 2 summarises the detection p erform ance of our AEC method together with some of the e xisting data mon itor- ing approach es, for the IMS dataset. Detection perfo rmance is defin ed as the samp le-time (sample n u mber) at which the algorithm notices initiation of a degrad a tion. Therefo re, an early fault-detection determ ines a better performance. T able 2 : Detection performa n ce. HMM-DPCA: Hidden Markov model with dyn amic PCA, HMM-PCA: Hidden Markov mod el with PCA [ Y u, 2 0 12a ] . MAS-K ortusis: Mov- ing av erage spectral k urtosis [ Kim et al. , 2016 ] . VRCA: V ariable-rep lacing-b a sed contribution analysis [ Y u, 2012a ] . - means that the dataset has no t been analyzed in the corre- sponding e xperimen t Algorithm S1B3 S1B4 S2B1 S3B3 Degradation starting datapoin t AEC 2027 1641 547 2367 HMM-DPCA 2120 1760 539 - HMM-PCA - 1780 538 - RMS 2094 1730 539 No detection MAS-Kurtosis 1910 1650 710 No detection VRCA - 1727 - No detection W e can observe that on ly AEC provides the degradation starting poin t for all the exp e r iments. It is essential to m en- tion that mon itoring of the test-to-failure pr ocess o f the exper- iment S3B3 is considerd as a hard challenge to be performed . Only fe w app roaches can monitor its state while non p r ovided the degradation starting point [ Kim et al. , 2016 ] . AEC pro - vides a r easonable pr ediction on such dataset as well. Framework2 - Online Prognostic In th e seco nd fram ew ork, as discussed, we train the auto- encoder over the first 70% of the av ailable data in order to ev aluate the prediction performan ce of th e model in an online monitorin g phase. Here, we study six cases inclu ding S1B3- sensor1, S1B3- sensor2, S1B4-sensor1 , S1B4-sensor2, S2B1 and S3B3 wher e Figure 5A to 5F represent the predicted AEC rate in ea ch experiment respectively . A sample is defined as the initiation of an abnorm al state where its AEC r ate reaches 90% of the rate of the 1 00 steps earlier sample. The pred iction process starts from the samples collected from day 5 on, since before that the status of the system has been considered to be normal [ Qiu et al. , 2 006 ] . Using this definition, the degra da- tion starting po int is calculated with high level of accuracy and the p ropag a tion of fault is captured during the simulated run-to - failure test. T able 3 summar ises the p r edicted degra- dation starting p o int together with the prediction accuracy in each exper imnet. The pred iction error is comp u ted as the ratio of the d eference between th e pre d icted sample and the monitore d fault star tin g point, to the total number of samples in each experim ent. W e also comp are AE C with most of the existing meth o ds for fault mo nitoring and pro gnostic systems over se veral at- tributes. Such q ualifications a r e determined as follows: • Gener a lization: Ab ility of th e method to provide clever results for v arious bearing status mon itoring e xperi- ments. • Status Monito ring: Ability of the method to provide a useful health-status trend during the test-to -failure ex- periment of th e bearing s. T able 3: Prediction accuracy of the AEC method in different bearing run-to-failure tests Experiment Fault starting point Prediction Accuracy S1B3-sensor1 2120 95.68% S1B3-sensor2 2122 95.59% S1B4-sensor1 1681 98.14% S1B4-sensor2 1673 98.51% S2B1 6 10 93.60% S3B3 2435 98.47% 200 400 600 800 1000 2000 3000 4000 500 1000 1500 2000 500 1000 1500 2000 1980 2000 2020 2040 2060 2080 2100 1950 2000 2050 2100 A B C D Sample E F 2030 20 50 207 0 2090 211 0 2130 2150 2000 2050 2100 2150 550 1050 1550 2 050 550 1050 1550 2050 250 450 650 850 1220 2220 3220 4220 Figure 5 : Online monitorin g of the status of the system in different experiments: A) S1 B3 first sensor B) S1B3 second sensor C) S1B4 first sensor D) S1B4 seco nd sen sor . E ) S2B1 F) S3B3. The c o lor-bar represents the AEC rate. T he degra- dation starting point of the system in each exper iment, with a high le vel of accuracy is predicted and AEC r ate c ompreh en- si vely elucidates ho w the fault propagates d uring each run-to- failure test. • Autom ated: a f u lly autonomo us fault prog nostic method • Unsup ervised: Capability of the method to extract infor- mation from th e sensory without any prio r kn owledge. • Detection Sensitivity: Ab ility of the method to provide a fast-enough alert on the starting po int of the degradation in the test. • Fault-type Detection (Diagnostics): Ability of the method to detect a certain type of defect in the system and classify them into different fault classes. T able 4 co mprehe n si vely illustrates a qu alitativ e compar i- son among various fault prog nostic meth ods utilized for bear - ings, based on the mentio ned attrib utes. The assessment on the performan ce of e a ch method is car efully performed based on the provid ed re su lts and the detailed specifications of th e metho ds in their co rrespon ding report. Under such ev aluations, resu lts suggest the sup eriority of th e AEC algo- rithm over existing m ethods wher e it can precisely cap tu re the degradation initial point and provides a usefu l tre n d for the fault spread while being auton omous. T able 4 : Qualitativ e c omparision of the perform ance of the existing ap proache s on the bearing dataset prog- nostic. HMM-DPCA: Hidden Markov model with dy- namic PCA, HMM-PCA: Hidden Markov mod el with PCA [ Y u, 2012a ] . MAS-Kortusis: Moving average spectral kurtosis [ Kim et al. , 2016 ] . VRCA: V ariable- replacing- based contribution analy sis [ Y u, 2012a ] . EET : Energy Entropy tr end [ Kim et al. , 2016 ] . WPSE-EMD: W avelet packet sample E n tropy [ W an g et al. , 2011 ] - Em- pirical m ode decomposition [ Lei et al. , 2007 ] . Spectral- ANN: Thir d-ord e r sp ectral + ar tificial neur al n etworks [ Y ang et al. , 2002 ] . Fuzzy-BP: Fuzz y log ic with back- propag ation [ Satish and Sarma, 2005 ] . SVM: Su pport V ecto r Machine [ Y ang et al. , 20 07 ] . GA-SVR: Gen e tics algorithm- Suppor t vector regression [ Feng et al. , 2009 ] . GLR-ARMA: Generalized likelihood ratio - Autoregressi ve m oving average [ Galati e t al. , 2008 ] . ++: Highly satisfi es. +: Satisfies. -: Th e attribute is not covered -+: The attrib ute is fairly covered. Algorithm Generalization status monitoring Automated unsupervised Detection sensitivity fault-type detection AEC + + + + ++ - HMM-DPCA -+ + - + + - HMM-PCA - + - + + - RMS - + - - -+ - Kurtosis - + - - -+ - MAS-Kurtosis - + - - -+ - Spectral-ANN -+ - - - + + VRCA -+ + - - ++ - EET + + - + - - WPSE+EMD - + - - + - Fuzzy-BP -+ - + - + + SVM - - - - + ++ GA-SVR - + - - + - GLR-ARMA - + - - + -+ 5 Conclusions W e introduced a ne w autonomou s technique for fault prog- nostic in machin e bearings. T he method was based on an unsuper v ised feature extractions fro m sensors b y means of an auto -encod er . Correlation analysis on the extracted fea- tures was performed an d correspondin gly useful trend on the status of the bearin g during the test-to-failure was p rovided. W e showed that AEC successfully m o nitors the status o f the bearings in various experiments which con firms its general- ization ability . AEC accurately predicts the degradatio n start- ing poin t while provid ing an informative trend on its prop a- gation. Furtherm ore, AEC algorithm generates rich unsuper- vised features f rom the vibr a tion data and is auto mated. For the future work we intend to imp rove the qu ality of our method and ap ply it to th e pro gnostic of the other key machine comp o nents such as degrad ation o f gears, cutting tools an d spin dles. One can also think of a mo r e gen eral so- lution where a d eep con v olutional auto -encod er is designed and trained and accordingly e v aluate th e health-status of the system, independently , without any other fo llow-up steps. Refer ences [ Boureau et al. , 2008 ] Y -lan Bou reau, Y ann L Cun, et al. Sparse featu re learn ing for d eep b elief network s. p ages 1185– 1192 , 2008 . [ Feng et al. , 2009 ] Fu Zhou Feng, Do ng Dong Z hu, Peng Cheng Jiang , and Hao Jiang. Ga-svr based b e a ring condition degradation pred iction. I n K e y eng ineering materials , volume 413, p ages 431– 4 37. Trans T ech Publ, 2009. [ Fisher , 192 5 ] Ronald A ylmer Fisher . Sta tistical methods for r esear c h work ers . Genesis Publishing Pvt Ltd, 1925. [ Galati e t al. , 2008 ] F Antonio Galati, David Forrester , and Subhrak anti Dey . Application of the gener alised likeli- hood ratio algo rithm to the d etection of a b earing fault in a helicopter transmission. A ustralian Journal of Mechanical Engineering , 5(2):1 69–1 7 6, 2 008. [ He e t al. , 2 007 ] Qingbo He, Fanrang K o ng, and Ruqiang Y an. Sub sp ace-based g e a rbox conditio n monito r ing by kernel principal com ponen t analysis. Mechanica l Systems and Signal Pr ocessing , 21(4):175 5–17 72, 200 7. [ Jardine et al. , 2006 ] Andrew KS J ardine, Daming Lin, and Dragan Banje vic. A re view on machin ery d iagnostics and progn ostics implemen ting condition-b a sed m aintenance . Mechanical systems an d signal pr ocessing , 2 0 (7):14 83– 1510, 2 006. [ Jia et al. , 2016 ] Feng Jia, Y aguo Lei, Jing Lin, Xin Zhou , and Na L u. Deep neural n etworks: A p r omising tool for fault characteristic min ing and intelligent diagno sis of ro- tating m achinery with ma ssi ve data. Mechanical Systems and Signal Pr ocessing , 72:303–3 15, 2016. [ Kim et al. , 2016 ] Seokgoo Kim, Su ngho Park, Ju-W o n Kim, Jun ghwa Han, Dawn An, Nam Ho Kim, a n d Joo - Ho Choi. A ne w progno stics ap proach for bearing based on entropy decre ase an d com parison with existing meth- ods. Annu al Conference of the Pr o gnostics and Health management Society , 2016. [ Lee et a l. , 200 7 ] J Lee, H Qiu, G Y u , J Lin, et al. Bear- ing data s et. IMS, University of Cincin nati, NASA Ames Pr o gnostics Da ta Repository , Re xnor d T echnica l Services , 2007. [ Lee et a l. , 201 4 ] Jay Lee, Fangji W u, W enyu Zhao, M a so ud Ghaffari, Linxia Liao , an d David Siegel. Prog nostics and health management d esign for r otary machinery sy stem s- revie ws, methodo logy and application s. Mechanica l sys- tems and signa l p r ocessing , 42(1) :314–3 34, 2014. [ Lei et al. , 2007 ] Y agu o Lei, Zhengjia He, Y anyang Zi, an d Qiao Hu. Fault diag n osis of rotating machinery based on multiple anfis combin ation with gas. Mechanica l systems and signal pr ocessing , 21(5):2280 –229 4 , 2007 . [ Møller, 1993 ] Martin F odslette Mø ller . A scaled conjugate gradient alg orithm for fast sup e rvised learning. Neur al networks , 6(4):525– 5 33, 1993. [ PCoE, Accessed 201 6 ] PCoE. PCoE, Progn ostics Cen- ter of Excellence o f NASA: The prognostics data repository . https: //ti.arc.nasa.gov/tech/ dash/pcoe/prognostic- data- rep o sitory/ , Accessed 2016. [ Qiu et al. , 200 6 ] Hai Qiu, Jay Lee, Jing Lin , and Ga n g Y u. W a velet filter -based weak signature detec tio n method and its application on rolling element bearing prognostics. Journal of sound a n d vib ration , 289 (4):10 66–10 90, 20 06. [ Reuben and Mb a, 20 1 4 ] Lim Chi K eong Reuben and Da vid Mba. Diagn ostics and prognostics using switching kalman filters. Structu ral Health Mon itoring , 13(3):29 6–30 6, 2014. [ Satish an d Sarma, 2 005 ] B Satish and NDR Sarma. A fuz z y bp app roach for d iagnosis and prognosis of bea ring faults in inductio n mo tors. In P ower Engineering S ociety Gen- eral Meeting, 2005. IE E E , pages 2291 –229 4. IEEE, 2 005. [ Sun et al. , 2016 ] W en jun Sun, Siyu Shao, Rui Zha o, Ruqiang Y an , Xingwu Zhang , an d Xuefen g Chen . A sparse auto -encod er-based deep neur al network appr oach for induction motor faults classification. Mea sur ement , 89:171 –178, 2016 . [ Thirukovalluru e t al. , 201 6 ] Raghuveer Thirukovalluru, Sonal Dix it, Rahul K Sev akula, Nishcha l K V erma, and Al Salour . Generating featur e sets fo r fault diagn osis using denoising stacked auto-encoder . In Pr ognostics and Health Management (ICPHM), 2016 IEEE Internation a l Confer ence on , pages 1– 7. IE EE, 2 0 16. [ T obon-M e jia et al. , 2012 ] Diego Alejand ro T obon -Mejia, Kamal Medjah e r, Noureddine Z erhoun i, and Gerard T ripot. A data-driven failure progn ostics method ba sed on mixture of gaussians hid d en m a rkov models. I EEE T r ans- actions o n Reliability , 6 1(2):4 91–50 3, 2 012. [ W an g et al. , 2011 ] Fengtao W ang, Y ang yang Zh ang, Bin Zhang, and W ensh eng Su. Applica tio n o f wa velet packet sample entropy in th e forecast of rolling element bearing fault trend. In Multimedia a nd Signal Pr ocessing (CMSP), 2011 In ternationa l Con fer ence on , v olume 2 , pages 1 2–16. IEEE, 2011. [ W an g et al. , 2016 ] Y u W ang, Y izhen Peng, Y anyang Zi, Xi- aohang Jin, and Kwok-Leun g Tsui. A two-stage data- driven-based pro gnostic approach for bearing degradation problem . IEEE T ransaction s on Indu strial Informatics , 12(3) :924–9 32, 2 016. [ Y ang et al. , 2002 ] D-M Y ang, AF Stronach , P MacConne ll, and J Penman. Th ird-ord er spectral techn iques fo r the diagnosis of moto r bear ing condition using artificial neu - ral networks. Mechan ic a l systems and signal pr ocessing , 16(2- 3):391 – 411, 2 002. [ Y ang et al. , 2007 ] Junyan Y an g, Y ouyun Zhang , and Y ong- sheng Zhu. In telligent fault diagnosis of ro lling eleme nt bearing b ased on svms and fractal dimen sion. Mechanical Systems a nd Signal P r o cessing , 2 1(5):2 012–2 024, 2 0 07. [ Y u, 2 0 12a ] Jianbo Y u. Health condition mo nitoring of ma- chines based on hidden markov model an d con tribution analysis. IEEE T ransactio n s on Instrumentation and Mea- sur ement , 6 1(8):2 200–2 211, 2 012. [ Y u, 2 0 12b ] Jianbo Y u. Lo cal and nonloc a l p reserving pro- jection for b earing defect classification an d perfor mance assessment. IEEE T r ansactions on In dustrial Electr onics , 59(5) :2363– 2376, 2 012.