Intelligent Prediction and Maintenance Decision-making for Hidden Faults of Power Equipment Based on Deep Learning + Gate-Controllable Recurrent Network

Abstract: Long-term operation under strong electromagnetic interference and complex working conditions buries faint early faults of main transformers and GIS, hindering early identification via traditional maintenance. Conventional data-driven methods ignore irreversible degradation physics, leading to unreliable decisions. This paper proposes a physics-prior embedded CNN-GRU: fault-sensitive physical channels capture incipient anomalies from degradation mechanisms; loss function adds monotonic health index constraint to conform to irreversible decay; probabilistic risk modeling and expected cost optimization enable predictive maintenance. Tests show its RMSE reaches 0.062 (transformer) and 0.071 (GIS), outperforming existing algorithms.