基于层级多尺度特征提取网络的开关柜局放模式识别

Pattern Recognition of Switchgear Partial Discharge Based on Hierarchical Multi-Scale Feature Extraction Network

摘要: 空气开关柜中对于局部放电的识别多依赖单一参量，而现有多参量方法虽能提高识别准确率，但在高效融合特征和适应环境变化方面仍面临挑战。为此，本文提出一种基于层级多尺度特征提取网络(Hierarchical Multi-Scale Feature Extraction Network, HMFENet)的多参量融合局部放电识别方法，自研集成温湿度、超声、臭氧、二氧化氮和紫外检测的传感装置，并采集4种典型局放类型数据进行建模。该网络采用三层分层结构，每一层先通过质数组合卷积核实现多尺度特征提取，再结合Transformer编码器实现跨模态的多参量特征融合与复杂信号的捕获。实验结果表明，在各项评价指标上，HMFENet都优于其他各类机器学习与深度学习模型，识别准确率达97.50%，为开关柜在线监测与故障预警提供了可靠的技术方案。

Abstract: In air-insulated switchgear, partial discharge (PD) recognition still relies predominantly on a single diagnostic quantity. Although existing multi-quantity approaches improve accuracy, they remain challenged by efficient feature fusion and robust adaptation to varying ambient conditions. To address these limitations, this study proposes a multi-quantity PD identification method based on a Hierarchical Multi-Scale Feature Extraction Network (HMFENet). A self-developed sensor suite that integrates temperature, humidity, ultrasonic, ozone, nitrogen dioxide, and ultraviolet channels was employed to acquire data for four representative PD defect types. HMFENet adopts a three-tier architecture: within each tier, prime-sized convolution kernels first extract multi-scale features, after which Transformer encoders perform cross-modal fusion and capture long-range dependencies inherent in the complex PD signatures. Experimental results demonstrate that HMFENet surpasses both conventional machine-learning and advanced deep-learning benchmarks under all evaluated metrics, achieving a recognition accuracy of 97.50%. The proposed scheme offers a reliable technical solution for online monitoring and fault early-warning of switchgear installations.