Time-Frequency Convolutional and Attentional Temporal Network for Prediction-Based Bearing Condition Monitoring

To address the challenges of highly non-stationary vibration signals under complex industrial operating conditions, this study proposes a time–frequency convolutional and attentional temporal network (TFC-ATN) for prediction-based bearing condition monitoring of mining trucks. The proposed framework integrates time–frequency convolution, LSTM-based temporal modeling, and temporal attention to extract spectral–temporal features, capture long-term dependencies, and emphasize informative temporal segments under continuously changing operating conditions. Experimental results based on a field-measured multi-operating-condition vibration dataset from a 300-ton mining truck demonstrate that TFC-ATN outperforms CNN, CNN-LSTM, and FAN-BD in terms of MAE, RMSE, MAPE, and R2, exhibiting superior prediction accuracy, robustness, and adaptability under complex multi-operating-condition environments. Prediction residuals between forecasted and measured signals may provide further useful indications of abnormal operating behaviors for future anomaly-monitoring applications.