Investigation on Intelligent Recognition of Failure Axis Orbit Based on HWT and CNN

Abstract

In condition monitoring and fault diagnosis systems of the large machine, such as the shafting of MU train or the rotating system of the compressor. Axis orbit can visually and comprehensively reflect the unit operating status. In this study, an intelligent recognition method of axis orbit based on Harmonic wavelet (HWT) and convolution neural network (CNN) is presented to deal with the problems of lower automation and accuracy. Firstly denoised axis orbit is obtained by reducing the noise of two-channel signals collected in the field using HWT with phase lock function. Then, Gaussian Blur is adopted to extend the training set for the small sample problem in deep learning. Automatic feature extraction and recognition are achieved by CNN network with LeNet-5 architecture. After optimizing the convolutional kernel parameters and training parameters, field data is used to verify the accuracy of the model which is higher than 99.96%, Meanwhile using the analog signal to produce the new training set with more complex and variable shapes, resulting in the accuracy of new model not less than 97.11%. This method has high practical value for the shafting running state detection and the early fault identification.