Abstract: In order to solve the problem that the traditional voiceprint diagnosis method is highly dependent on artificially constructed feature engineering,which limits the improvement of diagnostic accuracy,this study proposes a voiceprint diagnosis method with multimodal residual connection.Firstly,a two-modal feature extraction architecture is constructed,and a two-dimensional convolutional neural network is used to capture the time-frequency spatial features of voiceprint signals,and the time-series dynamic characteristics of the signals are extracted by combining with the gated recurrent neural network.Then,the cross-modal feature fusion is realized by weighted residual connection.Finally,the feature dimension of global average pooling compression is adopted,and the generalization ability is optimized by- DOI:10.3969/j.issn.1000-1565.2025.06.004融合多模态残差连接的变压器故障声纹诊断方法庞永林¹,王文超¹,刘天奇¹,粱佳宇¹,张欣伟¹,王坤涵¹,董乐聪²,张雄²(1.国网内蒙古东部电力有限公司 电力科学研究院,内蒙古 呼和浩特 010010;2.华北电力大学 机械工程系,河北 保定 071003)摘 要:基于传统声纹诊断方法因高度依赖人工构造特征工程而难以进一步提升诊断准确率,提出一种融合多模态残差连接的声纹诊断方法.首先,利用二维卷积神经网络捕捉声纹信号的时频空间特征,同时结合门控循环神经网络提取信号时序动态特性构建双模态特征提取架构;其次,通过加权残差连接实现跨模态特征融合;最后,采用全局平均池化压缩特征维度,结合批量归一化及分层Dropout策略优化泛化能力.基于110 kV与500 kV变压器实测数据的实验结果表明,模型平均准确率较传统卷积神经网络与长短期记忆网络模型提升10.26%,在低信噪比(-10 dB)仍保持90.2%的准确率,且在噪声环境下的性能显著优于同类模型,充分验证了本文方法的高效性与优越性.关键词:卷积神经网络;门控循环神经网络;残差连接;特征融合中图分类号:TM41 文献标志码:A 文章编号:1000-1565(2025)06-0590-11Transformer fault voiceprint diagnosis method based on multimodal residual connectionPANG Yonglin¹, WANG Wenchao¹, LIU Tianqi¹, LIANG Jiayu¹, ZHANG Xinwei¹, WANG Kunhan¹, DONG Lecong², ZHANG Xiong²(1.Electric Power Research Institute, State Grid Inner Mongolia Eastern Electric Power Co., Ltd., Hohhot 010010, China;2.Department of Mechanical Engineering, North China Electric Power University, Baoding 071003, China)Abstract: In order to solve the problem that the traditional voiceprint diagnosis method is highly dependent on artificially constructed feature engineering,which limits the improvement of diagnostic accuracy,this study proposes a voiceprint diagnosis method with multimodal residual connection.Firstly,a two-modal feature extraction architecture is constructed,and a two-dimensional convolutional neural network is used to capture the time-frequency spatial features of voiceprint signals,and the time-series dynamic characteristics of the signals are extracted by combining with the gated recurrent neural network.Then,the cross-modal feature fusion is realized by weighted residual connection.Finally,the feature dimension of global average pooling compression is adopted,and the generalization ability is optimized by- 收稿日期:2025-07-07;修回日期:2025-09-09 基金项目:国家自然科学基金项目(52105098);河北省自然科学基金项目(E2024502052;E2021502038);中央高校基本科研业务费专项资金项目(2025MS137);国网内蒙古东部电力有限公司科技项目(SGMDDK00GGJS2400134) 第一作者:庞永林(1995—),男,国网内蒙古东部电力有限公司工程师,主要从事高压绝缘技术及在线监测技术方向研究. E-mail:17303223392@163.com 通信作者:张雄(1990—),男,华北电力大学副教授,博士,主要从事旋转机械设备状态监测与故障诊断方向研究.E-mail:hdjxzx@ncepu.edu.cn 第6期庞永林等:融合多模态残差连接的变压器故障声纹诊断方法河北大学学报(自然科学版) 第45卷combining batch normalization and hierarchical dropout strategies.The experimental results based on the measured data of 110 kV and 500 kV transformers show that the average accuracy of the model is 10.26% higher than that of the traditional CNN-LSTM model,the accuracy of 90.2% is still maintained under the low signal-to-noise ratio(-10 dB),and the performance in the noise environment is significantly better than that of similar models.The result fully verifies the efficiency and superiority of the proposed method.

Key words: convolutional neural network, gated recurrent neural network, residual connections, feature fusion