Wavelet Analysis And Deep Learning Based Fault Classification For Industrial Processes

A platform of high-level informatization in fault management plays an important role in the equipment maintenance and troubleshooting.Traditional data-driven techniques such as statistical analysis and signal analysis are normally used to model the fault with the mathematical statistics or the features in time,frequency and spatial domain.These methods possess strong interpretability and reliability.While the newlydeveloped deep learning algorithms have revealed strong performance,some of them lack in explicability and applicability for the industrial data.They can be integrated with traditional methods to improve the fault classification accuracy.In the traditional fault classification techniques,wavelet transform is regularly used to denoise the signals,whereas some useful frequency components are swept away.The end-to-end discrete wavelet transform is implemented to extract the frequency domain features adaptively,and then deep residual network(DRN)is followed to describe the deep-level pattern of faults.The results show that the algorithm achieves a good result of fault classification.Regarding the blindness in the feature extraction process of the auto-encoder,wavelet auto-encoder(WAE)is proposed to focally describe the data in the space domain,making it more beneficial for the classification task.A classification model is established by incorporating the WAE with DRN,and the results show that the algorithm surpasses the performance of all conventional algorithms.In the existing implements of convolutional neural networks(CNN),a 2-dimensional receptive field is set to extract the correlated features in the time-space domain.However,such receptive field is sensitive to the order of the feature variables,and the feature extraction is not robust at all.To deconstruct the temporal and spatial correlation,the variable-wise 1-dimensional convolution and pooling operation is proposed to extract the signals' temporal characteristics,and the fixed-width convolution is used to cover the spatial characteristics.And then a 1-dimensional convolutional neural network is established to do fault classification.The simulation results suggest that the algorithm greatly improves the accuracy of fault classification and achieves better performance than the existing algorithms.