Abstract: In the process of fatigue degradation of gears, there are usually steady-type faults and shock-type faults. These two types of fault features contain rich information about gear degradation. Most of the current end-to-end prediction methods on remaining useful life are based on the semantic features extracted by neural network. Basically, the establishment of the prediction model has a low degree of correlation with the gear, and the physical interpretability is poor. To this end, combined with the steady-type and shock-type fault response mechanism and the convolutional neural network, a convolution kernel in time domain and angle domain fused with fault prior is designed, and the feature extraction ability of the neural network is used to mine the steady-type and shock-type faults in the vibration signal. At the same time, an attention mechanism is introduced to suppress noise, dynamically capture different fault components and degraded information in signals at different times. Finally, the weighted optimization of the prediction error is realized by the noise covariance matrix of the model. The simulation and experimental results show that the method can effectively mine the fault information of the original signal, has better physical interpretability, and has higher prediction accuracy and better robustness on remaining useful life than the traditional intelligent prediction method, which verifies the effectiveness of the proposed method.

Key words: gear, steady-type faults, shock-type faults, neural network, remaining useful life