Research Of Fault Diagnosis Based On Wavelet Neural Network

In recent years, with the development of computer technology, signal processing, artificial intelligence, mode identification and etc, the fault diagnosis technology has been continuously promoted. Especially fault diagnosis method based on intelligence is widely studied. With the improvement of neural network technology, fault diagnosis method based on neural network has been paid so much attention. Since one of the main steps of fault diagnosis is signal processing, while wavelet analysis is an effective tool to process signals and wavelet function has many good characteristics, so the combination of wavelet and neural network, so called wavelet neural network, has become a focus in fault diagnosis field recently. In this paper, an adaptive multi-dimensional diagonal wavelet neural network based on single dilation compact wavelet frame is constructed after comparing many wavelet neural network structures and decomposing data using wavelet pack to get characteristics. This network includes the initial network and sub-networks adaptively incorporated during the training according to the accuracy. The hide layers of every network are composed of single dilation compact wavelet frame. The expected output of every sub-network is the error of the last network and the training of the incorporated sub-network has no influence on the parameters that have been trained. Thus the number of nodes in the hide layer can be adaptively set according to the accuracy. Compared with the traditional structure of multi-dimensional wavelet neural network, this network can further solve the "dimension disaster" problem. At the same time, diagonal structure is adopted in the hide layer, which can reflect the dynamic characteristic through the self-feedback without increasing the number of the input nodes and what's more the memory is unlimited. For this network, a dynamic recurrent least square algorithm is derived to train the network parameters. And the method to initiate the dilation and translation parameters together with the method to initiate the network weights is given. Use this wavelet neural network and the data characteristics that are obtained by wavelet pack decomposing to diagnose the faults of the pump jacks. The results show that the fault diagnosis method based on wavelet neural network is much more effective and accurate than the method based on BP neural network, which can be better used in practical application.