Yan Peifeng, Chang Chenzhang, Hua Dong, Huang Haomin, Liu Suisheng, Cui Peiyi
School of Electric Power Engineering, South China University of Technology, 381 Wushan Road, Tianhe District, Guangzhou 510641, China.
School of Science and Engineering, Chinese University of Hong Kong, 2001 Longxiang Avenue, Longgang District, Shenzhen 518172, China.
Sensors (Basel). 2025 Mar 10;25(6):1701. doi: 10.3390/s25061701.
Due to long-term outdoor working, High-Voltage Disconnectors (HVDs) are prone to potential faults. Currently, most studies on HVD state diagnosis methods have tested only one type of HVD, and the generalization capability of these methods for other HVDs has not been verified. In this paper, we propose an HVD state diagnosis method featuring adaptive recognition capabilities based on Fault Difference Signals, Adjusted Relative Position Matrix and Convolutional Neural Networks (FDS-ARPM-CNN). First, we align the measured operational power signal of the HVD drive motor with the recorded normal operational power signal, deriving the FDS through subtraction. Next, to address the issue of traditional Relative Position Matrix (RPM) conversion processes that lose sample amplitude information, we introduce a targeted improvement to the relative position matrix calculation method, converting the one-dimensional FDS into a two-dimensional image. Finally, we achieve high-accuracy diagnosis and classification of HVD states using a CNN that incorporates Batch Normalization (BN) and GELU activation functions. Experimental validation demonstrates that the neural network model, trained on one model of HVD, maintains strong generalization capabilities on data from other HVD models. This method effectively alleviates the challenges of acquiring fault samples in data-driven approaches for HVD state diagnosis, showcasing significant practical value.
由于长期户外工作，高压隔离开关（HVD）容易出现潜在故障。目前，大多数关于HVD状态诊断方法的研究仅对一种类型的HVD进行了测试，这些方法对其他HVD的泛化能力尚未得到验证。在本文中，我们提出了一种基于故障差异信号、调整后的相对位置矩阵和卷积神经网络（FDS-ARPM-CNN）的具有自适应识别能力的HVD状态诊断方法。首先，我们将HVD驱动电机的实测运行功率信号与记录的正常运行功率信号对齐，通过减法得出故障差异信号（FDS）。接下来，为了解决传统相对位置矩阵（RPM）转换过程中丢失样本幅度信息的问题，我们对相对位置矩阵计算方法进行了有针对性的改进，将一维FDS转换为二维图像。最后，我们使用结合了批量归一化（BN）和高斯误差线性单元（GELU）激活函数的卷积神经网络实现了HVD状态的高精度诊断和分类。实验验证表明，在一种HVD模型上训练的神经网络模型对来自其他HVD模型的数据保持强大的泛化能力。该方法有效缓解了数据驱动的HVD状态诊断方法中获取故障样本的挑战，具有显著的实用价值。
