A Spike-Driven Transformer Network for Edge Computing-Based Partial Discharge Monitoring Using Ultrahigh Frequency Pulse Signals

The ultrahigh frequency (UHF) partial discharge (PD) detection method is widely adopted for early diagnosis of insulation breakdowns. However, the existing UHF method for acquiring low-noise signals and precisely differentiating PD types, are usually with high hardware/software cost and time-consuming. While deep learning (DL) methods have demonstrated strong capabilities in classification of UHF PD signals, their practical application on edge devices is constrained by high computational and power demands, latency issues, and sensitivity to noisy data. To address these issues, this article develops a novel UHF monitoring system by combining a low-noise UHF sensing frontend and a spike-driven Transformer network toward edge computing, enabling a robust, energy-efficient and cost-effective PD recognition. Specifically, a multiscale pulse feature extraction strategy driven by weight integration is proposed, which aggregates multilevel feature representations layer by layer to enhance the model’s ability to capture critical information from noisy datasets. Additionally, a projection optimization-based training framework is designed to improve feature learning and activation efficiency within a limited number of time steps, ensuring high detection accuracy and performance despite constrained computational resources. The proposed method is validated in accordance with IEC 62478, benchmarking against strong baselines and state-of-the-art (SOTA) techniques. Experimental results demonstrate the highest recognition accuracy, and reductions in computational complexity and energy consumption by 56.25% and 52.0%, respectively.