Anomaly detection and explainable intelligent diagnosis for multisource measurement and control signals via physical-consistency-aware edge networks

Abstract

Industrial measurement and control systems in pump-valve stations collect heterogeneous edge signals whose fault evidence is delayed and physically coupled. Conventional thresholds are transparent but miss early faults, whereas large fusion models may exceed edge resources or attend to implausible sources. This paper presents PCX-Net, a physicalconsistency- aware edge framework for pump-valve diagnosis in an anonymized steelmaking secondary-cooling circulating-water system. PCX-Net aligns vibration, current, temperature, pressure, acoustic, and PLC state streams, estimates delayed thermal response from vibration/current energy through a physical consistency residual (PCR), and uses the residual to guide lightweight cross-source attention. A predefined, non-causal PCS evaluates attention-to-expert-vector mechanism alignment. On a same-site held-out cross-pump field batch of 145,000 windows, deployed INT8 PCX-Net achieved F1 0.9356, missed rate 0.0483, FAR 0.0092, and PCS 0.9865, compared with 0.6821, 0.2743, 0.0447, and 0.8878 for a Transformer baseline. The INT8 target reached 29.6 ms P95 latency and 420 KB model size. A 72 h replay and 30- day online record further support tested-site deployment feasibility.