RU, Rand GOLOSINSKI, T.S. Early detectin f mining truck failure by mdelling its peratin with neural netwrks classificatin algrithms. Applicatin f Cmputers and Operatins Research ill the Minerals Industries, Suth African Institute f Mining and Metallurgy, 2003. Early detectin f mining truck failure by mdelling its peratin with neural netwrks classificatin algrithms H. HU and T.S. GOLOSINSKI Mining Engineering, University f Missuri-Rlla, Missuri, USA The paper reprts n use f neural netwrk algrithms t mdel perfrmance and cnditin f mining trucks. The mdel inputs are selected digital data streams cllected by truck Vital Infrmatin Management System, VIMS, during its peratin in a mine. The mdels built as the result f this wrk allw fr prjectin f truck cnditin and perfrmance int future. As such they facilitate mre efficient management f mine truck fleets. Keywrds: Mdelling, mdeling, neural netwrks, mining equipment, Vital Infrmatin Management System, Intelligent Miner, mining truck Intrductin Mdern mining equipment is equipped with numerus sensrs that mnitr its cnditin and perfrmance. Data cllected by these sensrs are used t alert the peratr t existence f abnrmal perating cnditins and t perfrm emergency shutdwn if the pre-set upper r lwer limits f the mnitring parameters are reached. These data are als used fr pst-failure diagnstics and fr reprting and analysis f equipment perfrmance. A vailabilit f these vluminus data, tgether with availability f sphisticated data prcessing methds and tls, allw fr extractin f additinal infrmatin cntained in the data. One set f methds that may permit this is data mining1,2. In previus research the authrs attempted t identify and extract this infrmatin by mdelling mining truck peratins using Decisin-Tree algrithms. These mdels used VIMS @ cllected data as inputs and were built t predict VIMS events. While these mdels have shwn reasnable average predictin errr rate (21 %), the high standard deviatin (26%) f errr rate fr test data sets indicates instability f the resulting predictins. As an example a typical mdel implies 14% prbability f false alarms related t high engine speed n ne hand, and 50% prbability f missing high engine speed alarms n the ther. Thus, these mdels are nt sufficiently reliable and can nt be used fr predictin f events that may take place 3,4. Research presented in this paper uses neural netwrk based mdelling t imprve predictin accuracy. This apprach has allwed fr significant reductin in errrs f predictins fr bth the truck perfrmance and its cnditin int the future. The IBM Intelligent Miner fr Data was used t build the relevant mdels. Data preparatin The data analysed in this research is f tw types; it includes snapsht (event recrder) and datalgger VIMS recrds, each cntaining values f 70 truck parameters recrded ver a perid f time. These data were cllected at six Caterpillar 789B trucks during their peratin in a surface mine. One snapsht stres a segment f truck histry that cntains values f all 70 mnitred parameters recrded during the perid f six minutes. Value f each parameter is recrded nce per secnd. The snapsht recrding is triggered by ccurrence f an abnrmal situatin r event, such as when a mnitred parameter reaches a pre-set critical value. A snapsht recrd describes truck cnditins during five minutes befre the event and ne minute after the event takes place. In this paper, every snapsht recrd is called 'event' fr simplicity. Unlike snapsht, the data lgger recrds values f all truck parameters that are mnitred by VIMS ver varying perids f time, als at ne-secnd intervalss. The recrding needs t be triggered and stpped manually, with individual recrds cvering perids f up t 30 minutes f truck peratin. Datalgger recrds d nt have t be assciated with any events. Of the 70 truck parameters mnitred in the field, values f 26 were recrded as categrical and the remaining 44 as numeric values. The example f basic statistical descriptin f the numerical parameter values is presented in Table 1. The actual mdel inputs are statistics f recrded truck parameter values calculated fr ne-t three-minute time intervals f truck peratin. The statistics include: Minimum Maximum Range Average Standard deviatin Variance Regressin slpe Regressin intercept Regressin sum f square. Neural netwrks Research described in this paper used neural netwrk algrithms t mdel truck peratin. Neural netwrks are cmputer implementatins f sphisticated pattern detectin and machine learning algrithms used t build predictive mdels frm large histrical databases. They allw fr cnstructin f highly accurate predictive mdels that serve t slve a large number f different prblems. EARLY DETECTION OF MINING TRUCK FAILURE BY MODELLING ITS OPERATION 191
Table I Example f numerical parameter values Parameter name Minimum value Maximum value Mean value Standard deviatin AFTRCLR_TEMP-110 0 95 41.8 12.8 AMB_AlR_TEMP_791 0 38.5 21.9 7.0 ATMOS_PRES_790 0 93 89.4 9.1 BOOST-PRES_105 0 164 31.0 50.1 Nde Nde Detail Input Hidden Layer Output Figure 1. Back-prpagatin neural netwrk architecture 8 Varius visualizatin techniques were used in cnjunctin with neural mdels t help explain and cntrl the mdel, and t assure its clarity and transparency. IBM Intelligent Miner implementatin f relevant algrithms was used t cnstruct the mdels the tl. The 'back prpagatin' f a specific algrithm was used t train the neural netwrk. Discvery f an effective methd f training a multiplayer neural netwrk led t the re-emergence f neural netwrks as a tl fr slving a wide variety f prblems. This training methd is called back prpagatin (f errrs) r the generalized delta rule. It is simply a gradient descent methd that minimizes the ttal squared errr f the utput cmputed by the netwrk6. A multilayer neural netwrk with ne layer f hidden units (Z units) is shwn in Figure 1. The input layer cntains the statistical parameters used t predict VIMS events. The utput are the predicted VIMS events. The left side f Figure 1 is a basic neurn, utput f which is calculated by the activatin functin fez). One f the mst typical activatin functins is the binary sigmid functin, which has the range f (0, 1) and is defined as Equatins [1], [2]. 1 J;(x) = 1 +e-x with the fllwing prperty J;' (x) = J; ( x )[ 1-1; ( x ) ] The Z is calculated as the sum f the prducts f the inputs and related weights. The utput units (Y units) and the hidden units may have biases, while the input value is [1] [2] always ne. Only the directin f infrmatin flw fr the feed-frward phase f peratin is shwn in Figure 1. During the back-prpagatin phase f learning, signals are sent in the reverse directin. VIMS event mdelling T facilitate mdelling f truck peratins the pattern f changes in VIMS parameter values was analysed, as assciated with varius events. The bjective was t identify such patterns in changes f parameter values that, if repeated, may prvide early indicatin f an impending failure. The ther VIMS data were then screened fr presence f these patterns. Initial analysis and mdelling were dne by building a decisin tree classificatin mdel f the truck. The mdel was t allw fr predictin f an ccurrence f a selected event, based n the pattern f changes in values f ther parameters. Event predictin The 'high engine speed' events were mst numerus in analysed data. Cnsequently these were chsen t be the main targets f mdelling and analysis. Caterpillar defines the 'engine speed' as the actual rtatinal speed f the crankshaft. Fr the mdelled truck, this event is activated when the engine speed reaches 2250 rpm and deactivated when the speed drps t 1900 rpm. Fr cmparative analysis VIMS data cllected during nrmal peratin was selected, gruped in 'ther' class. In this wrk the truck peratin was divided in three-, tw-, and ne-minute intervals. Fr each interval the patterns f interest were sught. Figure 2 illustrates VIMS 192 APCOM2003
Nrmal Engine Speed High Engine Speed Snapsht Nrmal Engine Speed VIMS.. Data 5 6 0 High Eng Event Predicted Other Other Label l,- 767_1 767_2 En9_1 En9_2 Other Other Figure 2. VIMS event predictin mdel event predictin mdel f ne VIMS event, 'high engine speed' based n three-minute truck peratin intervals. T predict ccurrence f this event statistics are defined fr each three-minute interval f VIMS recrds. If the analysed three-minute VIMS recrds shw similar statistics as d the first three minutes f the 'high engine speed' snapsht, the prbability exists that the high engine speed will ccur after anther tw minutes f truck peratin. As shwn in Figure 2, the ne-minute interval mdel can predict events that will ccur within the next 4 minutes f truck peratin. Similarly, the tw-minute interval mdel can prvide predictins f events that may ccur within the fllwing 3 minutes. The three-minute mdel can nly predict events that will ccur within the fllwing tw minutes 3. Mdeling The tw main prcedures f data mining are training, called als mdel cnstructin, and testing called als mdel validatin. In training mde, the functin builds a mdel based n the selected input data. This mdel is later used as a classifier. In the test mde, the functin uses a set f data t verify that the mdel created in the training mde prduces results in satisfactry precisin f predictins. In this wrk all available data were split int tw parts. Bulk f the data, 86.4%, were used fr mdel training. There are remainder, 13.6% f available data, were used fr mdel testing. The test data include dataset #1 (randm selectin) and dataset #2 (the whle set f all snapsht and datalgger data). Three mdels were built based n the ne-, tw-, and three minute interval statistics. The errr rate was defined fr each and used fr evaluatin f the training and the testing prcesses. This was dne by analysing cnfusin matlix determined fr each mdel. Results During mdeling, the ne-minute and tw-minute based mdels were unable t cnverge t predefined errr rate after as many as 2000 passes in their training mde. Only three-minute mdel was able t satisfactrily cnverge. T quantify the distributin f the misclassificatins f the neural netwrk mdel runs the cnfusin matrix was used. In every matrix, the number n the diagnal indicates the crrect classificatin; ther numbers indicate misclassificatin. The cnfusin matrix btained fr neminute and tw-minute interval statistics shw that the netwrk has nt learned enugh t be able t differentiate 'high engine speed' frm 'Other'. Only the mdel built frm the three-minute statistical data is able t d s (Figures 3 t 7). Discussin Figure 8 cmpares results f Neural Netwrk Classificatin mdeling and that f Decisin Tree Classificatin fr threeminute truck peratin intervals. The Neural Netwrk Classificatin mdel run has lwer errr rates f training and test fr dataset #1. The errr rates f the individual classes are presented in Table IT. While Decisin Tree mdel has 6% and 1 % lwer average errr n training and test data f dataset #1, the Neural Netwrk has 20% lwer errr rate (standard deviatin) as evident in Figure 9. While the Neural Netwrk mdel has smewhat higher verall errr rate, it shws much higher stability. The perfrmance predictin is quantified by accuracy f mdel runs n tw datasets, 'test #1' and 'test #2'. This cmpares t 14% f alarms missing and 50% false alarm prbability in Decisin Tree mdel. While runs f the Neural Netwrk mdel shw 17% f alanns missing, the prbability f false alann is nly 29%. Thus Neural Netwrk mdels allws fr making mre accurate predictin f the 'high engine speed' VIMS. The Intelligent Miner incrprates the sensitivity analysis functin, which is represented by a list f input fields ranked accrding t their respective imprtance t the classificatin functin. The results are nrmalized s that they ttal 100%. A parameter that is listed as having 20% scre is twice as imprtant in making the desired classificatin as a parameter with a 10% scre 7. The parameters with highest sensitivity, defined fr the Neural Netwrk mdel and nt lwer than 0.5 are shwn in Table m. This Table als gives parameter descriptins. Cnclusins The mdelling apprach presented in this paper cmpresses the VIMS data int statistical table and facilitates predictin EARLY DETECTION OF MINING TRUCK FAILURE BY MODELLING ITS OPERATION 193
Ttal Errrs = 24 (24%) Pred,cted Class --> I Other I Eng1 I Eng3 I Eng2 I Eng4 I Eng6 I Eng5 I Unknwn Other 1353 1 01 01 01 01 01 L 1 32 Eng1 37 1 11 01 01 01 01 '1 28 Eng3 39 1 01 01 01 01 01 11 28 Eng2 42 1 01 01 01 01 01 0 1 23 Eng4 40 1 01 01 01 01 01 11 28 Eng6 36 I 01 01 01 01 01 11 30 Eng5 15 1 01 01 01 01 01 3' 1 10 Figure 3. Training with ne-minute statistical data Ttal Errrs = 15 (15.31 0 /0) Predicted Class --> 1 OTHER I ENG1 I ENG2 I ENG3 I Unknwn OTHER 643 I 17 I 01 01 15 ENG1 141 44 0 I 0 I 10 ENG2 391 14 I 0 0 15 ENG3 581 21 01 01 5 Figure 4. Training with tw-minute statistical data Ttal Errrs = 72 (12.7 0 /0) Predicted Class --> 1 OTHER I ENGl I ENG2 Unknwn OTHER 373 47 1 31 15 ENGl 3 60 1 1 2 p ENG2 1 0 68 I 1 Figure 5. Training with three-minute statistical data f events with certain accuracy. The time hrizn fr predictins is tw minutes. Perfrmance f Neural Netwrk Classificatin mdeling n the analysed data differs frm that f Decisin Tree Classificatin mdeling. While the Neural Netwrk mdel built n the three-minute statistics f VIMS data sets shws an errr rate clse t that f the Decisin Tree mdel, the errr rate f standard deviatin is nly 6%. It fllws that the first shws mre rbustness. The balanced prbability that an alarm will be missed and that a false alarm will be 194 APCOM2003
Ttal Errrs = 15 (23.8 % ) Predicted Class --> 1 OTHER I ENG1 1 ENG2 Unknwn OTHER ENG1 ENG2 ---------------------------------------------------- 41 31 11 81 41 1 1 1 011 310 Figure 6. Test #1 with three-minute statistical data Ttal Errrs = 9 (14%) Predicted Class --> I OTHER I ENGl 1 ENG2 Unknwn OTHER ENGl ENG2 --------------------------------------------------~- 44 I 1 61 6\ 1 2 I 0 5 0 Figure 7. Test #2 with three-minute statistical data -;!2. ~ a: l...... W 5 D Neural netwrk n three- minute data Decisin three n three-minute data training test #1 test #2 Figure 8. Ttal errr rate cmparisn between Neural Netwrk and Decisin Tree classificatin D Decisin Tree n three-minute data NN three-minute data ~ 0 -~ "- g 0) r "- > «0.25 0.20 0.15 0.10 0.05 0.00 trai'ning test c 0 0.3 1il '> 0.25 "0 "0 0.2 "- r "0 c 0.15 r t5 0.1 -~ 0.05 g 0 w training test Figure 9. Errr rate statistics cmparisn between Neural Netwrk and Decisin Tree classificatin EARLY DETECTION OF MINING TRUCK FAILURE BY MODELLING ITS OPERATION 195
Tablell Neural netwrk mdel perfrmance calculatin Decisin Tree-three minute Training Test #1 Test #2 Ttal test Class Crrect Ttal Errr rate % Crrect Ttal Errr rate % Crrect Ttal Errr rate % Crrect Ttal Errr rate % Other 411 438 0.06 42 51 0.18 47 52 0.10 89 103 0.14 Engl 65 66 0.02 5 8 0.38 2 6 0.67 7 14 0.50 Eng2 70 70 0.00 4 4 0.00 6 6 0.00 10 10 0.00 Average % 0.03 0.18 0.25 0.21 (errr rate) Standard 0.03 0.19 0.36 0.26 deviatin (errrrate) N-tlu'ee minute Training Test #1 Test #2 Ttal test Class Crrect Ttal Errr rate % COlTect Ttal Enr rate % Cnect Ttal Errr rate % Crrect Ttal Enrrate % Other 373 438 0.15 41 51 0.20 44 52 0.15 85 103 0.17 Engl 60 66 0.09 4 8 0.50 6 6 0.00 10 14 0.29 Eng2 68 70 0.03 3 4 0.25 5 6 0.17 8 10 0.20 Average % 0.09 0.32 0.11 0.22 (errr rate) Standard 0.06 0.16 0.09 0.06 deviatin (errr rate) Table III VIMS parameters with high sensitivity Parameter Descriptin Sensitivity DIFF FLTR SW AVG The status f the differential axle il filter (plugged r OK) 0.5 ENG_LOAD_MlN This is calculated by the engne ECM (Electrical Cntrl Mdule) after cnsidering the engine 0.5 speed, thrttle switch psitin, thrttle psitin, bst pressure, and atmspheric pressure and is shwn as a per cent f a full lad ENG_SPD_REGR SLOPE The actual rtatinal speed f the crankshaft 0.5 GROUND_SPD_MlN The speed f the machine relative t the grund 0.5 MACHlNE_RACK_MAX This is calculated frm the fur machine suspensin cylinder pressures. 05 VIMS takes the sum f the tw diagnal suspensin cylinder pressure minus the sum f the tw ther diagnal suspensin cylinders pressures PAYLOAD MlN The paylad is calculated by VIMS based n pressure f the fur suspensin cylinders RETARDER AVERAGE The status f the retarder system (On r Off) 0.5 RT_R_BRK_TEMP MlN The cling il temperature frm the right rear brake 0.5 TC_OUT3EMP _MAX The il temperature n the utlet f the trque cnverter 0.5 recrded is 17% and 29% respectively, a significant imprvement ver the results f Decisin Tree Classificatin mdeling. Neural Netwrk Classificatin needs mre time fr mdel training. Furthermre, the results f mdel runs are nt as easy t interpret as that f Decisin Tree Classificatin based mdels. There is als a risk that that Neural Netwrk Classificatin mdel training runs may nt cnverge in the absence f large databases. All cnclusins drawn here are valid fr the investigated dataset f VIMS. The mdel predictin accuracy, while imprved, remains relatively lw. T imprve the accuracy f the mdel, mre VIMS data are needed, including bth mre values fr each parameter and values f mre VIMS parameters. The mdel predictin accuracy, while imprved, remains relatively lw.! Acknwledgements Financial supprt fr the research presented in this paper was prvided by Caterpillar, Inc. f Peria, Illinis. The IBM (Internatinal Business Machines) prvided a cpy f the IBM Intelligent Miner at n charge t the researchers under the IBM Schlars prgram. References 1. GOLOSINSKI, T.S. Data mining uses in mining. Prceedings, Cmputer Applicatins in the Minerals Industries (APCOM), Beijing, China, 2001, pp. 763-766. 2. GOLOSINSKI, T.S. and HV, H. Data mining f mine 196 APCOM2003
equipment databases. Prceedings f the Artificial Neural Netwrks in Engineering Cnference (ANNIE 2001), St. Luis, Missuri, U.S.A. 200l. pp. 387-396. 3. GOLOSINSKI, T.S. and RU, R. Mdeling cnditin and perfrmance f mining equipment. Prceedings, Mine Planning and Equipment Selectin 2002, Czech (in print). 2002. 4. RU, H. and GOLOSINSKI, T.S. Failure pattern recgnitin f a mining truck with a decisin tree algrithm. Mineral Resurces Engineering (in print). 2002. 5. Caterpillar, Inc. Vital infrmatin management system (VIMS): system peratin testing and adjusting. Cmpany publicatin. 1999. pp. 10-13. 6. FAUSETT, V.L. Fundamentals f neural netwrksarchitectures, algrithms, and applicatins' Flrida Institute f Technlgy. 1994. 289 pp. 7. IBM (Internatinal Business Machines Crpratin). Manual: using the intelligent miner fr data. Cmpany publicatin. 2000. 298 pp. S. CLAIR, ST. D.C. CS404: Data mining and knwledge discvery. Lecture Nte. Cmputer Science Department f University, Missuri-Rlla. 200l. EARLY DETECTION OF MINING TRUCK FAILURE BY MODELLING ITS OPERATION 197
198 APCOM2003