Fault Prediction of Mechanical Equipment Based on Hilbert-Full-Vector Spectrum and TCDAN

To solve the problem of "under-maintenance" and "over-maintenance" in the daily maintenance of equipment, the predictive maintenance method based on the running state of equipment has shown great advantages, and fault prediction is an important part of predictive maintenance. First, the spectrum information of the equipment is extracted by the Hilbert-full-vector spectrum as the input of fault prediction. Compared with the traditional spectrum, this spectrum information fuses the signals of two sensors in the same section of the device, which can reflect the actual operational state of the device more comprehensively. Then, the temporal convolutional network is used to predict the amplitudes of different feature frequencies, and the double-layer attention mechanism is introduced to mine the correlation between the corresponding amplitudes of different feature frequencies and between the data at different historical moments, to highlight the more important influencing factors. In this way, the prediction accuracy of the model for the amplitude corresponding to the feature frequency of concern is improved. Finally, experimental verification is carried out on the XJTU-SY dataset. The results show that the TCDAN model proposed in this paper is significantly superior to TCN, GRU, BiLSTM, and LSTM, which can provide a more effective decision-making basis for the predictive maintenance of equipment.