Fault Diagnosis Based On HHT And CNN For Wheelset Bearings Of High-Speed Trains

The wheelset bearing is the core component of bogies in high-speed trains.Its reliability and stability have a vital impact on train safety.With the increase of train speed,the operation condition of wheelset bearings becomes more complicated,which aggravates the occurrence of faults in wheelset bearings.Therefore,the fault diagnosis of wheelset bearings becomes extremely necessary and urgent.At present,signal processing and data-driven methods are widely used in fault diagnosis of rolling bearings.However,the wheelset bearings work for a long time in the harsh conditions of strong impact,high speed and big workload,they are vulnerable to suffering various faults,which can lead to the decline of diagnostic performance of signal processing methods,and the poor anti-noise ability and load domain adaptability of data-driven methods.Therefore,how to extract more discriminant fault features from complex vibration signals is the key to solve the accurate fault diagnosis of wheelset bearings,which is also the research topic of this paper.In this paper,the fault diagnosis of wheelset bearings in high-speed trains is studied.The main research idea is to extract more discriminant fault features.Three improved methods are proposed to improve diagnostic performance of signal processing methods and the anti-noise ability and load domain adaptability of data-driven methods.The main research works of this paper are as follows:(1)In order to improve the accuracy of signal decomposition,signal demodulation,instantaneous frequency and instantaneous amplitude estimation of Hilbert-Huang transform,a soft sifting stopping criterion is proposed.Based on it,an improved Hilbert-Huang transform method is proposed.This method can adaptively determine the sifting iteration number of empirical mode decomposition and normalized Hilbert transform in Hilbert Huang transform method.In addition,a fault diagnosis method based on the improved Hilbert Huang transform and fast kurtogram is proposed,which improves the accuracy of fault diagnosis of wheelset bearings.(2)In order to extract more discriminant high-level features from signals with low signal-to-noise ratio,this paper proposes a one-dimensional residual module to address the problem of training difficulty and performance degradation of deep networks,and then constructs a one-dimensional deep residual convolutional neural network,which obtains high-level discriminant features by using layer-by-layer abstraction of convolution layers.In addition,the introduction of wide convolution kernel and dropout can effectively improve the global feature learning ability and generalization ability of the network.Compared with the state-of-the-art CNN-based fault diagnosis methods,the proposed method has better anti-noise ability and load domain adaptability.(3)In order to learn abundant and complementary features from complex signals and further improve the model's load domain adaptability,this paper combines multiple signal processing methods with convolutional neural network,and proposes a multi-branch multi-scale convolutional neural network based on the idea of multi-scale learning.The network uses multi-branch structures and multi-scale modules to learn features from multiple signal components and time scales,and thus significantly improve the load domain adaptability while having good anti-noise ability.This paper evaluates and validates the effectiveness and superiority of the above three methods on the dataset of wheelset bearings in high-speed trains.The experimental results demonstrate that the proposed three methods are superior to the existing state-of-the-art methods.Finally,the existing problems and future work of this paper are summarized.