Multi-Dimensional Industrial Data Fusion-Based Causal Transformer for Fault Propagation Prediction

Abstract

Synchronous virtual models of equipment situations, process parameters, control actions, and maintenance records have been produced by production-line digital twins; nevertheless, fault propagation analysis in coupled-machine scenarios with delayed responses is still an unresolved issue. A Causal Transformer model for fault propagation reasoning in production-line digital twins is presented in this research. In order to differentiate directed fault influence from regular statistical correlation, the method creates a multi-source state sequence from sensors, PLC alarms, equipment topology, operation logs, and maintenance records. It then applies causal masking, topology-aware temporal encoding, and propagation-path decoding. The experimental platform is a simulated discrete production line with 26 equipment nodes, 148 process variables, and 18 typical failure types. The suggested model outperformed the LSTM, TCN, conventional Transformer, and graph neural network baselines, achieving 94.6% fault-source accuracy, 92.8% propagation-chain F1, and an average warning lead time of 11.7 minutes. Therefore, causal attention may enhance the robustness and interpretability of predictive maintenance based on digital twins.