Real-time fault detection technology for power transmission and distribution equipment based on intelligent sensing
TL;DR
A deep temporal recognition network consisting of temporal convolutions, residual propagation, and attention aggregation is proposed to model the evolution of short-term disturbances, persistent fluctuations, and abrupt anomalies and maintains lower response delay and more stable real-time fault response capability.
Abstract
In order to improve the real-time fault detection capability of power transmission and distribution stations, this paper presents a new approach for fault identification, which combines smart sensing, deep temporal networks, and edge inference. Addressing challenges such as inconsistent sampling frequency, temporal drift, and the difficulty of separating anomalous coupling characteristics from multi-source monitoring signals (including current, voltage, temperature rise, vibration, and PD). Based on this, we propose a deep temporal recognition network consisting of temporal convolutions, residual propagation, and attention aggregation to model the evolution of short-term disturbances, persistent fluctuations, and abrupt anomalies. At the same time, a lightweight reasoning chain is used to build a dynamic early-warning score based on the combination of fault classification probability, state duration, and feature deviation intensity. The results show that the proposed approach achieves 97.3% accuracy, 96.9% F1-score, and 0.984 AUC. In the edge inference response test, the inference latency remains within the range of 22.3–31.4 ms, and the fault response completion rate fluctuates between 92.6% and 94.3%. Compared with the centralized upload mode, whose latency increases from 48.6 ms to a peak of 72.5 ms during high-frequency abnormal updates, the proposed edge inference mechanism maintains lower response delay and more stable real-time fault response capability.