【目的】为了实现对隔离开关机械故障的高精度识别,提出了一种融合注意力机制的并行混合网络,结合时序与图像特征进行智能诊断的方法。【方法】为充分利用双通道数据的特征信息,时序通道采用双向长短时记忆网络提取振动信号的时域特征,捕捉信号的时间动态变化,充分反映机械故障的时变特征;图像通道通过格拉姆角场将振动信号转换为二维图像,利用极坐标映射保留信号的时间动态特性,然后利用卷积神经网络提取关键图像特征。此外,两通道分别引入自注意力机制和通道注意力机制,能够自适应地调整每个通道的权重,从而突出关键信息,并有效减少特征冗余。【结果】针对GW4-126型隔离开关进行故障模拟试验,采集4种状态下的隔离开关振动信号,将本文所提故障诊断方法与其他5种深度学习模型相比。试验结果表明,本文所提方法的故障识别准确率超过97%,能够精确识别隔离开关的机构卡涩、松动及三相不同期等典型机械故障。【结论】本文提出的并行混合模型通过融合两种不同特征信息,克服了单一通道方法的局限性;通过引入注意力机制,模型能够更好地动态调整权值,突出关键特征,提高故障识别的精度和准确性。该方法为隔离开关的状态识别提供了可靠的理论依据和技术参考,对于未来的故障诊断和设备维护具有重要的应用价值,并为智能电网技术的发展提供了新的思路。

Abstract: [Objective] To achieve high-precision identification of mechanical faults in disconnectors, a parallel hybrid network incorporating attention mechanisms is proposed, which combines temporal and image features for intelligent diagnosis. [Methods] To fully exploit the feature information of dual-channel data, a bidirectional long short-term memory network was employed in the temporal channel to extract time-domain features from vibration signals, capturing the dynamic temporal variations of the signal and effectively reflecting the time-varying characteristics of mechanical faults. In the image channel, vibration signals were converted into two-dimensional images using Gramian angular fields, where polar coordinate mapping was utilized to preserve the temporal dynamics. A convolutional neural network was then used to extract key image features. Furthermore, a self-attention mechanism was introduced in the temporal channel and a channel attention mechanism in the image channel, enabling the model to adaptively adjust the weight of each channel, thereby emphasizing critical information and effectively reducing feature redundancy. [Results] Fault simulation experiments were conducted on GW4-126 type disconnectors, and vibration signals under four operating conditions were collected. The proposed method was compared with five other deep learning models. Experimental results demonstrated that the proposed method achieves a fault recognition accuracy exceeding 97%, effectively identifying typical mechanical faults such as mechanism jamming, looseness, and phase asynchrony. [Conclusion] The proposed parallel hybrid model overcomes the limitations of single-channel approaches by integrating two distinct types of feature information. The introduction of attention mechanisms enables the model to dynamically adjust weights, highlight salient features, and enhance the accuracy of fault identification. This method provides a reliable theoretical foundation and technical reference for the condition monitoring of disconnectors, holds significant potential for future fault diagnosis and equipment maintenance, and offers new insights for the development of smart grid technologies.

国网江苏省电力有限公司重点科技项目资助(J2024047)