Research On Fault Diagnosis Algorithm Of Rotating Machinery Of High-speed Train Based On Deep Learning

China’s high-speed railroad has been developing rapidly in recent years.As the key components of EMU,gearboxes and rolling bearings often work under the harsh working environment of high speed and heavy load,and are very prone to various types of failures.Therefore,it is important to study the fault diagnosis algorithm of gearboxes,rolling bearings and other rotating machinery for the safe and reliable operation of EMU as well as the subsequent repair and maintenance.In this thesis,taking the gearbox and rolling bearing of high-speed EMU as the research object,the deep learning technology is used to deeply study the fault diagnosis algorithm of rotating machinery from three aspects: fault feature extraction and health status identification,multi-sensor information fusion and deep transfer learning.The main contents of the thesis are as follows.(1)Aiming at the problems that the gearboxes and rolling bearings of high-speed EMU are complicated in structure and harsh in working environment,resulting in more noise and complex components of the collected vibration signals,an end-to-end multi-scale residual network deep learning model based on dynamic weighting(DWMR-Net)is designed to realize intelligent fault diagnosis of gearbox and rolling bearing.Firstly,the raw data is taken as the input of the network,and a wide convolution layer is designed to preliminarily fuse the information and expand the receptive field of the model;Then three independent parallel branch networks based on residual blocks are constructed,and the depth features are extracted from the parallel branch networks by designing multi-scale convolution kernels;Next,a dynamic weighting layer is designed to model the dynamic nonlinear relationship between feature channels using global information,and recalibrate the feature channels of each scale to improve the sensitivity of the network to fault information.Finally,the features of the three scales are fused,and the fault diagnosis is realized by the classifier.Experiments are carried out on the dataset of rotor gear comprehensive fault simulation experiment rig and public dataset,respectively,and the effectiveness of the algorithm proposed in this thesis is verified.(2)To address the problems of increasing difficulty in early faint fault identification and decreasing diagnostic accuracy,a faint fault diagnosis method based on multi-sensor information fusion is further proposed on the basis of DWMR-Net.The method uses sensor data from multiple channels,extracts fault features separately for fault diagnosis,and performs the final multi-sensor information fusion at the decision level by using the random forest model in ensemble learning,which improves the accuracy of weak fault diagnosis.It is verified that the proposed multi-sensor information fusion method can effectively improve the diagnosis accuracy of weak fault samples in the dataset of rotor gear comprehensive fault simulation experiment platform.(3)For the problem that often the sample working conditions are unknown or the labeled samples are missing in practice,an unsupervised transfer learning fault diagnosis method based on DWMR-Net is proposed.In order to reduce the difference between source and target domains in the common feature space,an adaptive metric is added to quantitatively measure the difference between source and target domain features by the maximum mean difference.And the metric criterion is embedded into the DWMR-Net deep learning model to make it adaptively learn domain invariant features on the one hand and achieve end-to-end fault diagnosis on the other hand.The proposed method is verified to have better transfer effect compared to other comparison methods in the transfer experimental task of different working conditions.Finally,the transfer effect of the model on deep features is shown by visualization experiments.(4)The algorithm proposed in this thesis is tested on the rotating machinery dataset of high-speed EMU.Experiments were carried out on the axle box bearing dataset of high-speed EMU and the traction motor bearing dataset of high-speed EMU,respectively.DWMR-Net can achieve 100% accuracy on both datasets of high-speed EMU,and the unsupervised transfer learning method based on DWMR-Net also achieved certain transfer effect in the transfer experiment task.