An interpretable multivariate remaining useful life prediction method of mechanical equipment based on adaptive threshold aggregation causal discovery

Accurately predicting the remaining useful life (RUL) of mechanical equipment is crucial for ensuring the safety and reliability of mechanical systems. However, existing deep learning-based RUL prediction methods often face challenges related to interpretability and robustness when dealing with complex multivariate time series data. To address this, this paper proposes a multivariate RUL prediction method based on adaptive threshold aggregation causal discovery. Specifically, a Bayesian Information Criterion-based causal discovery method is employed to explore the relationships between variables (i.e., sensor signals) across multiple samples, yielding a corresponding causal graph. An adaptive threshold mechanism is then designed to aggregate these sample-level graphs into a global structure that highlights key dependencies. Based on this, causal effect estimation is performed by combining front-door and back-door adjustment methods to generate a causal effect matrix. This matrix, together with the multivariate time series data, is input into a Temporal Graph Convolutional Network to capture dynamic dependencies and causal associations for RUL prediction. Experimental results on the C-MAPSS dataset show that the proposed method achieves an average RMSE of 13.3, outperforming state-of-the-art benchmarks and providing more reliable and interpretable insights into RUL prediction.