Research On Entropy Based Sensor Fault Diagnosis Methods

With the development of automation technology and intelligent technology, sensors are widely used in all kinds of systems. Both in the application of multi-sensor network system and single sensor system, the single sensor is the basic unit of each application for data collecting. Sensor as the core unit of the information acquisition, its working state directly affects the operation of the system. In practical applications, due to external environmental and sensor internal factors, such as environment overrun,connect failure of the hardware, calibration, aging, sensor is prone to failure or damage and will cause accuracy reduction of the measured information.Based on the low quality measured information, subsequent identification,judgments, decision-making and control processing will lost the true meaning.So sensor fault diagnosis and recognition are the basis to ensure the normal operation of the system, and has important practical significance.In this paper, for three kinds of common sensor faults, multiple signal analysis technologies are utilized to take intensive study of single sensor fault feature extraction and identification methods. Two new feature extraction and fault identification methods are proposed to improve reliability of the information acquisition and effective use of the basic sensor unit.Firstly, based on the crossing subband distribution characteristics of sensor fault, this thesis proposes a new method of the total multi-level wavelet time entropy (WTEtotal) for the sensor feature extraction. Combining extracted features, a back-propagation neural network classifier is designed. The sensor fault identification method based on total multi-level wavelet time entropy feature extraction and BP neural network is put forward. Experiments result demonstratesthat the recognition accuracy of the proposed identification method can reach 96%, and is higher than that of the wavelet time entropy based feature recognition method about 20%. It verifies the effectiveness of the proposed identification method, also shows that the total multi-levelwavelet time entropy can effectively represent sensor fault feature.Secondly, because multiscale weighted permutation entropy can represent the signal properties of the local structure and amplitude characteristics on multi-scales, this thesis combines multiscale weighted permutation entropy with the multi-resolution analysis of wavelet transform, and proposes a new wavelet transform based multi-scale weighted permutation entropy method(WMWPE) for feature extraction of sensor fault. Considering the large dimensions of the extracted feature, multi-cluster feature selection (MCFS)method are examined and employed as the feature selection method. A complete feature extraction and identification method are brought forward by combination of WMWPE , MCFS,BP neural network classifier. By verification, the recognition accuracy of this method is maintained at 99%,and signal feature vector dimension is reduced from 20 to 5. The proposed recognition method is effective and feasible, and the wavelet transform based multi-scale weighted permutation entropy can effectively represent the sensor fault feature.Thirdly, two proposed recognition methods based on WTEtotal and WMWPE are compared with each other from time consuming, recognition accuracy and anti-noise performance respectively. The method based on WMWPE can get higher recognition accuracy than the method based on WTEtotal about 3%, and the anti-noise performance is also better than that of WTEtotal. But the computation time of WMWPE based method is much higher than that of WTEtotal based method. In order to verify the universality of the proposed methods, large variation and small variation of senor data amplitude are used in experiments. Experimental results show that the WTEtotal based method is more effective for sensor data with small variation ampltitude than senor data with large variation amplitude. In contrast, the WMWPE based method can achieve good recognition accuracy for the two kinds of sensor data amplitude.Finally, based on the facts that the information entropy is a quantitative measure of the information generation rate and it reflects the information changes with time elapse, two sensor fault location methods are proposed.The experimental results show the location method based on mulit-level wavelet time entropy can accurately locate the bias and stuck-at faults using impulse peak values of MWTE and MWTEE, but cannot locate drift fault.The location method based on wavelet transform and weighted permutation entropy can accurately locate the bias and stuck-at fault using the pulse rising edge of DWT_WPE, at the same time drift fault can be positioned through the transition zone of DWT_WPE.