In recent years, China has made great efforts to develop fault diagnostic systems, and predictive maintenance in advance based on operational data can reduce downtime, operational costs and personnel safety in construction projects. In this paper, the time-frequency analysis method of instantaneous Fourier transform is used to study the time-frequency variation of non-smooth signals, and the appropriate window length is selected according to the vibration signals in order to provide better signal preprocessing results. Combine LSTM and CNN to deal with the system fault classification task, and introduce CBAM attention mechanism on this basis to improve the ability of recognizing and classifying fault signals in vibration data. Construct a predictive maintenance model to perform predictive maintenance on the system. The time-domain signals of severe faults are different from the other two states, and the amplitude is close to 10 m·s-2, which is easy to judge the system state after deep convolutional neural network calculation. Meanwhile, the predictive maintenance model designed in this paper incorporates the temporal convolutional network model with attention mechanism, which improves 7.042% and 45.223% in scores compared with the long memory network and bidirectional LSTM, which is of great value for practical system lifetime prediction.