军事知识服务平台
基于双通道声纹特征融合的变压器故障诊断方法
The Transformer Fault Diagnosis Method Based on Dual-channel Acoustic Fingerprint Feature Fusion
摘要 变压器声纹特征是反映其运行状态的关键特征量之一.传统变压器声纹诊断方法存在特征提取不充分、时序信息利用不足以及泛化能力有限等问题,难以满足多故障类型和复杂工况下的高精度诊断需求.针对上述问题,该文提出一种基于多级声学特征融合的变压器故障诊断方法.首先,采用梅尔倒谱系数(Mel-frequency cepstral coefficients,MFCC)和Mel频谱图提取算法提取声纹信息的差异化特征,以增强其特征表达能力;其次,构建了卷积神经网络(convolutional neural network,CNN)和双向长短期记忆网络(bidirectional long short-term memory,BiLSTM)-Transformer 的并行双通道诊断模型,第一通道采用CNN提取声纹图谱的时频特征,第二通道结合BiLSTM和Transformer编码器,充分挖掘MFCC的时序依赖关系;最后,通过双模态特征融合策略整合特征并进行分类诊断.研究表明,相较于传统单模态声纹诊断方法,在未添加噪声的环境下,该文方法的诊断准确率达99.61%,准确率提升4.8%以上;在高噪声环境下,诊断准确率分别提高29.1%、21.04%、30.75%和21.74%,表现出更强的鲁棒性和泛化能力.
Abstract:
The acoustic fingerprint of a transformer is one of the key characteristic indicators reflecting its operational status.Traditional transformer-based acoustic fingerprint diagnosis methods often suffer from insufficient feature extraction,inadequate use of temporal information,and limited generalization,making it difficult to meet the high-precision diagnostic requirements for multiple fault types and complex operating conditions.To address these issues,the paper proposes a transformer-based fault diagnosis method that leverages multi-level acoustic feature fusion.Firstly,a Mel-frequency cepstral coefficient(MFCC)and Mel-spectrogram extraction algorithm are employed to extract differentiated features of the acoustic fingerprint,thereby enhancing its feature representation.Secondly,a parallel dual-channel diagnostic model based on CNN and BiLSTM-Transformer is constructed.The first channel uses a CNN to extract spatial features from the acoustic fingerprint spectrum,while the second channel combines a BiLSTM and a Transformer encoder to exploit temporal dependencies in the MFCC fully.Finally,a bimodal feature fusion strategy is applied to integrate spatiotemporal features for classification and diagnosis.The efficacy of the proposed method is demonstrated by its ability to surpass the conventional unimodal acoustic pattern diagnosis approach,achieving a diagnostic accuracy of 99.61%across multiple fault types in a noise-free environment.This enhancement exceeds 4.8%when compared to the existing method.Furthermore,in a high-noise environment,the diagnostic accuracy is augmented by 29.1%,21.04%,30.75%,and 21.74%,respectively,thereby indicating greater of robustness and generalization capability.
刊 名 电网技术 ISTICEIPKU
年,卷(期): 2026, 50(10)
分类号: TM721
机标分类号: TP391 TP183 TN912.34
在线出版日期: 2026年11月2日