Research On Intelligent Fault Diagnosis Method Based On Deep Network Compression For Rolling Bearings

Rolling bearing is the core machine component of many rotating equipment and one of the most frequently used parts in the mechanical equipment.Because it always works under high speed and full load for a long time,the bearing is one of the most easily damaged mechanical parts.Its health states directly affect the safe and reliable operation of the whole equipment.Therefore,accurate bearing fault diagnosis plays a vital role in ensuring the safe and reliable operation of mechanical equipment.In recent years,deep learning(DL)has attracted more and more attention due to its powerful learning ability.DL-based intelligent fault diagnosis methods are effective tools for analyzing industrial big data analysis.Convolutional neural networks(CNNs)are the most widely studied and applied in the field of DL.However,their applications are limited in industry due to their slow training speed and high computation cost caused by their complex structure and a large amount of parameters.Based on the theory of network compression,this thesis focuses on studying the optimization of network structure and its balance with the classification performance,and aims to improve performance and efficiency of the CNN and provides an applicable method for intelligent fault diagnosis of rolling bearings.Main contents and contributions in this thesis are summarized as follows:(1)Soft filter pruning(SFP)algorithm uses the norm as the measure of the network pruning,but it is not accurate for evaluating important convolution kernels when compressing the network.To solve this problem,the SFP based on the first-order Taylor expansion is combined with the CNN in this thesis.First,the pruning rate is preset for CNN,and then the Taylor expansion value of each convolution kernel is calculated.Ranking their importance according to these expansion values,the insignificant convolutional kernels are removed by using the preset pruning rate,and the classification accuracy of the compressed CNN would not obviously decrease after using the network fine-tuning.Experimental results indicate that the proposed method can remove redundant parameters in deep CNN,and the obtained model has higher diagnosis accuracy and fewer network parameters.(2)The prerequisites for pruning are that the norm variance is small enough and the minimum norm is not too large,while engineering datasets do not match with these requirements.In order to solve this problem,the knowledge distillation and geometric median algorithm are introduced in the network pruning algorithm,in which their combination can extract diagnosis knowledge and further reduce the network size.Experimental results indicate that this method can compress the network and ensure high classification accuracy without considering the requirements of the convolution kernels with small values to be pruned,simultaneously descending the network performance as little as possible.(3)Because the noise in raw signals would influence the accuracy of intelligent diagnosis model,the complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)method is introduced in this thesis for noise reduction.After adaptively processing the collected signal,the noise in the complicated vibration signal is filtered out and some related signal components are selected for signal reconstruction.After that,the reconstructed signals are used as the input of the intelligent diagnosis model and processed for network training and compression.Experimental results indicate that a good signal preprocessing method is helpful to avoid the adverse effects of noise,and a higher accuracy is available for bearing intelligent fault diagnosis model even after network compression.