Avalable onlne www.ocpr.com Journal of Chemcal and Pharmaceutcal Research, 2014, 6(4):834-838 Research Artcle ISSN : 0975-7384 CODEN(USA) : JCPRC5 Applcaton research on rough set -neural network n the fault dagnoss system of ball mll Xuan Wang 1*, Tao Zhou 1, Jngxao Feng 2 and Yupeng Pe 2 1 School of Physcal and Electrcal Informaton, Luoyang Normal Unversty, Luoyang-Chna 2 Luoyang Mnng Machnery Research Insttute, Luoyang-Chna ABSTRACT The fault parameters of ball mll was unnotced among large amount of data, Accordng to the phenomenon, t put forward a fault dagnoss method of rough set to optmze neural network, and by usng wdth algorthm, so that the fault sample set of ball mll had been processed wth the dscrete way. Frstly, t bulded a dagnoss model of rough set-neural network, the dagnoss model had practcablty and superorty compared wth sngle network model. Keywords: Rough Set, Fault Dagnoss, Ball Mll, Neural Network INTRODUCTION For a long tme, due to the restrctons of dagnostc equpment and technology, the techncan can't predct accurately and quckly the occurrence of ball mll fault, so t drectly leads to shutdown, a maor economc loss, and causes casualtes f t broke down n the process of equpment operaton; In theory, some fault can be avoded through regular repar and mantenance of the equpment, but because of blndness s very bg, t easly lead to "Over Mantenance" or "owe mantenance", and t was a waste of resources and unnecessary economc losses. It need many montorng parameters n the process of ball mll work. In addton, the correlaton between the dfferent parameters made the modelng more dffcult. Due to the sudden, randomness, dsorder of fault, mantenance and repar s very dffcult. Therefore, how to effectvely use the recorded hstory data to dagnose the cause of the fault n lne wth the known data, t was a problem to be solved that predct the future runnng status of ball mll. EXPERIMENTAL SECTION Desgn of the rough set -neural network dagnoss model The operaton parameters varaton showed whether there were faults of the ball mll or not but the change of these parameters may not be perceved among a large amount of date, so the paper used the fault dagnoss method of rough set theory combnaton of BP neural network, n other words, rough set had been as the front system, usng the rough set optmzed neural network. Frstly, t selected the fault sample set of ball mll, and set up nformaton system tables. Secondly, t had data preprocessng, and had reduced the nformaton system by usng rough set, elmnated the redundant attrbutes and repeat nformaton, acqured the smplest decson table by the composton of mnmum condton attrbute sets and nuclear propertes, and as the bass of neural network tranng sample, the tranng nput and output sample of neural network had been determned from the smplest decson table. Fnally, accordng to tranng samples, the node number of neural network nput layer, hdden layer and output layer had been determned, based on these samples, t traned and tested the neural network. The flow chart of ball mll fault dagnoss as shown n Fg.1. 834
Fg. 1. The flow chart of ball mll fault dagnoss Fault dagnoss process of ball mll Ball Mll Fault Sample Set The ball mll had very huge ob data, so n some selecton fault attrbute were regarded as the condton attrbute of rough set, t need to apply the observaton data of feld sensors, nstrumentaton, etc, now t selected part of the fault data as fault sample sets, and had applcaton research of the algorthm. Specfc as follows: Selecton of ball mll fault: hgh pressure ppelne of feed end leakage, hgh pressure ppelne of dscharge end leakage, low voltage tube leakage, pnon-gear lubrcaton ppng leakage, hgh pressure pump falure, low pressure pump falure, pnon-gear pump falure, lubrcatng ol drty, coolng system falure, as the decson attrbute of rough set, these were denoted by (1, 2, 3, 4, 5, 6, 7,8,9), respectvely. Selecton of real-tme parameters when the fault happened: man bearng temperature of feed end, man bearng temperature of dscharge end, pnon-gear bearng temperature, hgh ol pressure of feed end, hgh ol pressure of dscharge end, low ol pressure, pnon-gear ol flow, hyperbarc pressure dfferental transmtter, low voltage pressure dfferental transmtter, pnon-gear pressure dfferental transmtter, as the condton attrbute of rough set, these were denoted by(a, b, c, d, e, f, g, h,, ), respectvely. And t created the attrbutes decson table of ball mll parts fault sample sets, as shown n Tab. 1. Table 1. The ball mll parts fault samples nformaton table Fault samples Condton attrbutes a b c d e f g h Decson attrbute U1 63 35 32 1.2 3.5 0.3 100 0 0 0 1 U2 36 61 32 3.9 1.0 0.3 100 0 0 0 2 U3 36 35 32 3.9 3.5 0.06 100 0 0 0 3 U4 36 35 63 3.9 3.5 0.3 50 0 0 0 4 U5 65 65 32 0.8 0.9 0.3 100 0 0 0 5 U6 40 41 38 3.9 3.5 0.07 100 0 0 0 6 U7 36 35 65 3.9 3.5 0.3 30 0 0 0 7 U8 65 66 67 1.6 1.3 0.15 60 1 1 1 8 U9 64 65 66 3.9 3.5 0.2 100 0 0 0 9 Preprocessng of data (1) Dscretzaton of data The saved data of mll falure data acquston system were contnuous analogue, so the data need dscretzaton before adoptng the algorthm of rough set. The dscretzaton can reduce the number of gven contnuous attrbute values, the expresson of dscrete attrbutes were closer than the contnuous attrbutes n knowledge. Dscrete attrbute were easer to understand, use and nterpretaton to users and experts. The paper used the method of data dscretzaton were wdth algorthm: there used nterval number were 6, the range of numercal attrbutes [ X, X mn max ] X mn X, max to ( ) 6 had been dvded 6 nterval, and each nterval had the same wdth, whch were equal. Through analyss of fault sample data of mll, there had dscretzaton to fault sample nformaton table wth same wdth, after arrangement, as shown n Tab.2. 835
Table 2. The ball mll parts fault samples nformaton table after dscretzaton Fault samples Condton attrbutes a b c d e f g h Decson attrbute U1 5 0 0 0 5 5 5 0 0 0 1 U2 0 4 0 5 1 5 5 0 0 0 2 U3 0 0 0 5 5 0 5 0 0 0 3 U4 0 0 5 5 5 5 1 0 0 0 4 U5 5 5 0 0 0 5 5 0 0 0 5 U6 0 1 1 5 5 0 5 0 0 0 6 U7 0 0 5 5 5 5 0 0 0 0 7 U8 5 5 5 1 0 1 2 1 1 1 8 U9 5 5 5 5 5 3 5 0 0 0 9 (2) Attrbute reducton of rough set Not all knowledge n the knowledge base were ndspensable, at the same tme, some knowledge was useless for knowledge decson, even t was redundant, t had counterproductve to the fault dagnoss of knowledge reasonng. There used reducton algorthm based on dentfablty matrx. If S ( U, R) U = x1, x2, L, x n ; R was attrbutes combnaton, and R = A D, A = { a, a2,, a n } 1 L was condton attrbutes; D was decson attrbute. a ( x ) was the values of attrbute a n the sample x. The defnton of c was the element of lne column n the dentfablty matrx. C defned as follows: = was a nformaton decson system, where U was doman, and { } c = 0 1 { a A: a( x ) a( x )} a ( ) D( x ) ( ) = x ) ( x ) a( x ) andd( x ) D( x ) D x D x where, [ 1, n] In order to ntroduce convenently the reducton method, f T was the orgnal decson table, M was the dentfablty matrx of T, A was the set of all condton attrbutes n the T, S was the set of all condton attrbutes B. The tem of the condton attrbute combnatons was 1 n the matrx shown: In addton to the attrbutes, other attrbutes wthout the dstngushng ablty between the two records, that s to say, the attrbute was nuclear attrbute. The constructon of rough set and neural network dagnoss model The model used 3 layers BP neural network structure (that s: nput layer, hdden layer and output layer), the goal was dagnoss for the ball mll falure. It bulded BP neural network model wth the fault samples of before and after rough set attrbute reducton. Neural network tranng functon was traned,ts learnng rate was varable n the process of tranng; learnng functon was the functon based on gradent descent method: learned;the transfer functon was log-sgmod; the performance functon was mean-varance functon mse; the tranng number was 2000; the tranng error was 0.001; the learnng rate was 0.08; the nput layer were 15; the output layer were 3; the hdden layer had been selected 31 through the tranng. The tranng results of BP neural network before attrbute reducton was shown n Fg.2. The fault samples (such as {b, d, f,g, } ) after rough set attrbute reducton bult BP neural network model, the tranng results had been shown n Fg.3. Fg.2. The BP neural network before the attrbute reducton 836
Fg.3. The BP neural network after the attrbute reducton As the tranng results of two pctures ndcated, the results of classfcaton were basc consstent, but accordng to the rough set algorthm reducton, the neural network of falure sample set tranng had change tranng parameters, that was: the tranng tme decreased from 9.062s before the attrbute reducton to 3.542s after the attrbute reducton, the tranng number decreased from 1148 before the attrbute reducton to 268 after the attrbute reducton, the mean-square error reduced from 0.000622756 before the attrbute reducton to 0.000452321 after the attrbute reducton, on the whole, accordng to the rough set algorthm reducton, the neural network of falure sample set tranng had shorter tranng tme, less tranng steps and hgher tranng accuracy. RESULTS AND DISCUSSION The network tranng nterface and fault dagnoss results of ball mll The weght and thresholds of the nput layer, hdden layer and output layer wth BP neural network before and after attrbute reducton had been saved, t had the BP neural network knowledge base of fault dagnoss. Fg4 was the network tranng nterface and fault dagnoss tables of ball mll, the falure data were nput to the dagnostc table, set the fault threshold was 0.7, f the fault threshold greater than 0.7, the fault had been consdered. From Fg4, t had fault that BP neural network fault dagnoss results before and after attrbute reducton, and the fault were both pnon-gear lubrcaton ppng leakage, but the fault dagnoss result of rough set combned wth BP neural network were more accurate than the fault dagnoss result of sngle BP neural network. Fg.4. The network tranng nterface and fault dagnoss table of ball mll CONCLUSION Through the above analyss, by comparng the fault dagnoss result of rough set combned wth BP neural network wth the fault dagnoss result of sngle BP neural network, t had reducton to nformaton system usng rough set, elmnated the redundant attrbutes and repeat nformaton, and obtaned the smplest decson table about consstng of mnmum condton attrbute set and nucleus attrbute, as the bass for the neural network tranng samples, the neural network had shorter tranng tme, less tranng steps and hgher tranng accuracy, the fault dagnoss result were more accurate. 837
REFERENCES [1] Donge Xng; Guozhong Chen, Constructon Mechanzaton, March 2013, pp.71-73. [2] Dongxue L, Hefe Unversty of Technology, Aprl 2012, pp.33-51. [3] Kune L, X an Shyou Unversty, Aprl 2012, pp.30-56. [4] Jang Han; Han Huang; Lan Xa; Hua Zha, Chnese Hydraulcs & Pneumatcs, June 2011, pp. 92-95. [5] Xuemng Wang, Unversty of Electronc Scence and Technology of Chna, July 2013, pp. 25-50. 838