基于改进ResNet18和声纹SDP的干式变压器局放故障状态检测

摘要:

干式变压器局放故障产生的声纹信号具有非平稳性和多分量调制特性，在含噪环境中难以有效提取瞬态脉冲成分、共振频率分布和空间方向性等关键特征，降低故障状态检测的准确率。为此，提出基于改进ResNet18和声纹SDP的干式变压器局放故障状态检测方法。根据干式变压器局放故障声纹特性，通过声纹传感器实时采集声纹信号，经过SDP处理（谱分解处理）生成声纹信号的时频图与空间谱图。通过输入层、卷积层、残差层、注意力机制层与全连接层的优化，改进ResNet18模型，引入了两条特征提取路径，一条用于提取声纹信号的时域特征，另一条用于提取频域特征。时域路径采用叠加的残差块，以捕捉长时依赖关系和瞬态脉冲成分，而频域路径则使用可变形卷积层，以灵活地调整卷积核形状，更好地捕捉共振频率分布和空间方向性等频域特征。以干式变压器声纹信号时频图与空间谱图为依据，采用改进ResNet18模型制定干式变压器局放故障状态诊断程序，获取最终的局放故障状态判定结果（轻微，一般，较重与严重），从而实现干式变压器局放故障状态的有效检测。实验结果显示:应用设计方法后，局放故障状态检测准确率最大值达到了99.5%，提取的声纹信号时频图与实际声纹信号时频图趋于一致，局放故障状态判定结果与测试样本标注结果相同。

Abstract:

The voiceprint signal generated by partial discharge faults in dry-type transformers has non-stationary and multi-component modulation characteristics, making it difficult to effectively extract key features such as transient pulse components, resonance frequency distribution, and spatial directionality in noisy environments, which reduces the accuracy of fault state detection. Therefore, a partial discharge fault detection method for dry-type transformers based on improved ResNet18 and voiceprint SDP is proposed. Based on the voiceprint characteristics of partial discharge faults in dry-type transformers, real-time voiceprint signals are collected through voiceprint sensors and processed by SDP (spectral decomposition) to generate time-frequency and spatial spectrograms of the voiceprint signals. By optimizing the input layer, convolution layer, residual layer, attention mechanism layer, and fully connected layer, the ResNet18 model is improved by introducing two feature extraction paths, one for extracting time-domain features of voiceprint signals and the other for extracting frequency-domain features. The time-domain path uses stacked residual blocks to capture long-term dependencies and transient pulse components, while the frequency-domain path uses deformable convolutional layers to flexibly adjust the shape of the convolution kernel, better capturing frequency-domain features such as resonance frequency distribution and spatial directionality. Based on the time-frequency and spatial spectrograms of the voiceprint signals of dry-type transformers, an improved ResNet18 model is used to develop a diagnosis program for partial discharge faults in dry-type transformers. The final partial discharge fault status determination results (mild, moderate, severe, and severe) are obtained, thereby achieving effective detection of partial discharge faults in dry-type transformers. The experimental results showed that after applying the design method, the maximum accuracy of partial discharge fault state detection reached 99.5%, and the extracted voiceprint signal time-frequency map tended to be consistent with the actual voiceprint signal time-frequency map. The partial discharge fault state determination result was the same as the labeled result of the test sample.