The Remaining Useful Life Probability Prediction Based on Multi-Sensor Signals Using TCN-Transformer

Deep learning methods have made significant progress in remaining useful life (RUL) prediction, which can provide decision support for preventive maintenance of industrial systems, thereby improving reliability and reducing downtime risks. However, most of the existing depth models are difficult to quantify the prediction uncertainty, and the correlation between sensors is not fully explored. In order to solve these problems, this paper proposes a TCN-Transformer model for RUL probability prediction of multi-sensor signals. Firstly, the encoder uses the temporal convolutional network (TCN) to extract sequence features, and combines the dual attention module to learn weight features from the sensor and time dimensions. Secondly, the regression network introduces approximate Bayesian inference to realize feature fusion and output RUL probability density estimation. Experiments show that the performance of this method is better than the existing models on the C-MAPSS dataset, and the importance of the confidence interval to the prediction results is verified.