Research On Approaches Of Data Mining And Information Fusion Based Fault Diagnosis

Rotating machinery such as turbines and compressors are the key equipments in oil refineries, power plants and chemical engineering plants. Defects and malfunctions (simply called faults) of these machines will result in significant economic loss. Therefore, it is necessary to monitor these key equipments and diagnose their faults in order to improve safety, allow predictive maintenance and shorten significantly the associated out of service time. The development of data mining and information fusion technology leads to a new solution to mechanical fault diagnosis problem. After summarizing the present status and developing trend of fault diagnosis and analyzing the importance of applying data mining and information fusion technologies to mechanical fault diagnosis, some studies are made as below:Experiment on Bently Rotor Kit: Bently Rotor Kit is a general test bed for rotor fault simulating. Five running conditions are of a rotor simulated on the test bed. They are normal (without any fault), unbalance, radial rub, oil whirl and a simultaneous state of unbalance and radial rub. The first condition is normal, and the last four conditions are fault conditions.Information fusion can be divided into three levels, such as data level, feature level and decision level. The information fusion in data level must deal with large amounts of data and has low computing efficiency. Therefore, multi-feature fusion method is firstly used to improve the accuracy of fault diagnosis. There are three type of features such as statistic indices in time domain (6 dimensions), amplitude spectrum in frequency domain (13 dimensions) and wavelet energy spectrum in time-frequency domain (25 dimensions). And the principal component analysis (PCA) method is used to make feature fusion. The fault diagnosis based on test data from Bently Rotor Kit proves that the feature fusion method improves the diagnosis accuracy efficiently.In order to make good use of the advanced technology of data mining, two kinds of classifiers (C4.5 decision tree and support vector machine, SVM) from data mining are improved. Based on the boundary point theorem given by Fayyad, an improvement method of selecting the optimal threshold is proposed to overcome the time-consuming disadvantage of original C4.5. According to the disadvantages of conventional SVM, a SVM decision tree (SVMDT) is proposed. The SVMDT solves the decline problem of SVM based on"one-against-one"or"one-against-other"strategy. Also the SVMDT makes probability assignment of outputs and is beneficial to decision fusion. Furthermore, the improved C4.5, SVMDT, and back-propagation neural networks (BPNN) are used as the basis classifiers and they make diagnosis decisions respectively. At last the improved D-S evidence theory is used to make decision fusion. The diagnosis analysis of experiment data testify that the diagnosis method based on multi-feature and multi-classifier fusion can improve the diagnosis accuracy greatly.In order to interpret why multi-decision fusion of multi-classifier can improve the accuracy of diagnosis, the diversity between multiple classifiers is studied. At the same time a principle of multi-classifier fusion named efficient diversity (ED) is proposed. The larger ED, the more efficient multi-classifier fusion and the higher diagnosis accuracy of multi-classifier fusion.Based on the process model of data mining and the method of information fusion, a new data acquisition and fault diagnosis system framework based on data mining and information fusion is designed. The system includes seven models such as data acquisition, data transform, feature selection, feature fusion, fault diagnosis of single classifier, knowledge presentation (i.e. rules or models), and decision fusion. The experiment in lab and the application in fan monitoring and diagnosis field prove the validation of the system design.