Data Processing in Condition Monitoring Intelligent Modeling

Artificial neural network (ANN) modeling is an effective method of intelligent modeling. All the information that ANN needs is got from training samples, so the accuracy of ANN modeling is based on quantities and quality of the samples. Such main problems of ANN model as training sample quality, relative definition, and symptom feature extraction algorithm had been solved by using rough set theory. A rough set neural network intelligent hybrid system model was presented, its implementation steps were analyzed, and the validity of above method was proved by application examples. Condition monitoring simulation of roll bearing installed in a band conveyor had been completed using SAS software. The results indicate that the data processing method of this paper can improve the performance of intelligent modeling. Introduction For many systems having complex mechanism, it is difficult to illustrate the inherent running law from the mechanism, and the system has the feature of non-linear, distributing parameters, time variance and time-lagging, etc. Intelligent modeling is often used in condition monitoring [1]. But the ANN condition monitoring method has some shortcomings such as : in order to make the networks convergent and gain stable diagnosis results, it needs more examples to train the network, so it is not suitable for new system or having no diagnosis case base system; for complex diagnosed system, nodes of each layer of ANN network are so many that more computation time is needed, diagnosis results cannot be explained, non-monotony and nonaccuracy resulting from incompleteness domain knowledge is difficult to be disposed; meanwhile, many factors such as ANN architecture, parameter setting, size and quality of learning samples, etc. can affect ANN precision and generalization directly [2-3]. Rough Set (RS) theory that is a data processing tool of depicting incompleteness and uncertainty had been brought forward for developing automation rule generation system based on indiscernibility relation and knowledge reduction method by Pawlak Z [4-5]. It can analysis and process efficiently all kinds of qualitative, quantitative and hybrid imperfect information of imprecise, inconsistent, incomplete and uncertain problems, find concealed knowledge, and discover potential rules. Its main idea is to acquire concept’s classification rules through knowledge reduction maintaining classification ability. Beginning with the RS theory, we systematically study training sample quality, define relative concepts, present a rough set–neural network intelligent hybrid system model, and simulate an application example. Research on Training Sample Quality and Symptom Feature Extraction Based on RS Theory Consistency Checking of Training Sample. RS theory adopts Knowledge Representation System (KRS) to express knowledge. The general form of KRS is: S=(U, A, V, f),where S is a KRS, U is Key Engineering Materials Online: 2004-03-15 ISSN: 1662-9795, Vols. 259-260, pp 783-786 doi:10.4028/www.scientific.net/KEM.259-260.783 © 2004 Trans Tech Publications Ltd, Switzerland All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of Trans Tech Publications Ltd, www.scientific.net. (Semanticscholar.org-13/03/20,19:03:52) 784 Advances in Grinding and Abrasive Processes non-empty finite set of objects called the universe, A is non-empty set of attributes, where A=C∪D, C D= , C is a set of condition attribute, D is a set of decision attribute, V is attribute value field, V= A a Va, a A, u , Va is attribute range D , f:U×A V is information function, defining the attribute “a” value of object u in the U. One of the most important concepts of RS theory is indiscernibility, which is used to define equivalence classes for the objects in the form of ind . On the basis concepts of KRS , ind and other RS, we set du as learning rule, relation function is du: a V, C | du represents the value of sample u in the condition attribute set C , D | du represents the value of sample u in the decision attribute set D . As to ANN, its sample is corresponding to one of the objects, its input sample set is the condition attribute set C, its output set is the decision attribute set D. So, some definitions relational to consistency checking of training sample are presented as follows: Conception 1: Consistent sample: if there are two samples w u , where U u , U w , C | d C | d w u = , D | d D | d w u = , i.e. the same inputs must exist the same outputs, then it is assumed that sample u and w are consistent samples. Conception 2: Inconsistent sample: if there are two samples u w, where u U, w U, C | d C | d w u = , D | d D | d w u , i.e. the same inputs must exist the different outputs, then it is assumed that sample u and w are inconsistent samples. Training Sample Data Set Reduction. Following definitions are relational to training sample reduction. Conception 3: Redundancy attribute: if there is an input sample c C, where ( ind ) C ( ind = C-{c}), then attribute c can be taken out, i.e. is called redundancy attribute. Conception 4: Main attribute: for the dependency of input sample on output sample, some input sample values are necessary, and cannot be changed, i.e. are called main attribute. Conception 5: Omissible attribute: some input sample values have little effect on the dependency on output sample, and can be taken out, i.e. are called omissible attribute. Conception 6: Minimum attribute set: including main attributes set and a part or all of omissible attribute set. Conception 7: Minimum sample set: sample set includes minimum attribute set. Conception 8: Training attribute set: attribute set is used in the training procedure, and includes minimum attribute set. Conception 9: Training sample set: sample set is used in the training procedure, and includes training attribute set. Symptom Feature Extraction. On the basis of attribute reduction principle of RS theory, the feature extraction algorithm of condition monitoring are as follows: 1. constructing decision table; 2. deleting repetitive samples; 3. classifying object set to gain concept set K by decision attribute; 4. computing classification quality ) K ( C of total attribute set C ; 5. computing classification quality ) K ( Q of attribute set C Q ; 6. if } C Q , Q { min C Q = = ,then is the reduction of attribute set C; 7. resulting in Minimum attribute set and Main attribute set. Rough Set – Neural Network (RNN) Intelligent Hybrid System Model RNN Model. Flowchart of RNN system [8] is shown in Fig.1. Implementation Steps. 1. Deleting all omissible condition attribute based on the dependency among condition attributes; 2. Discretization of each sample attribute value using domain knowledge to construct decision table; 3. Checking sample consistency to delete inconsistent sample; 4. 784 Advances in Grinding and Abrasive Processes