数据不平衡分布下燃气调压器故障识别方法  

Fault Identification Method for Gas Pressure Regulators Under Imbalanced Data Distribution

作　　者：尹孟伟 王勇 王超群 YIN Mengwei;WANG Yong;WANG Chaoqun(School of Computer Science and Technology,Shanghai Electric Power University Shanghai,201306,China;Shanghai Aerospace Energy Co.,Ltd.Shanghai,201201,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海201306 [2]上海航天能源股份有限公司,上海201201

出　　处：《振动.测试与诊断》2025年第2期346-353,415,共9页Journal of Vibration,Measurement & Diagnosis

基　　金：国家工程实验开放课题资助项目(QAX-20180);上海自然科学基金资助项目(20ZR1455900)。

摘　　要：针对燃气调压器故障识别中不平衡数据影响模型识别能力的问题,提出一种一维卷积神经网络(one-dimensional convolutional neural network,简称1D-CNN)与注意力机制(squeeze-and-excitation,简称SE)相结合的改进深度卷积神经网络(SE-1DCNN)方法。首先,使用一维卷积核提取故障特征;其次,在交替的卷积层后添加SE模块用于通道加权,选择性地保留所需的重要信息特征,并抑制弱相关的特征;最后,使用类平衡损失函数代替交叉熵损失函数来抵消不平衡分布给网络造成的影响。实验结果表明,根据真实环境中采集的不平衡故障数据,所提改进模型与其他故障识别模型相比有更好的故障识别能力,准确率高达98.17%。To address the problem that imbalanced data distribution significantly degrades model performance in gas pressure regulator fault identification,an improved deep convolution(SE-1DCNN) integrating a one-dimensional convolutional neural network(1D-CNN) with a squeeze-and-excitation(SE) attention mechanism is proposed.The proposed SE-1DCNN achieves efficient and accurate fault identification through three key steps.First,1D convolutional kernels extract fault features from raw data.Second,after stacking multiple convolutional layers,an SE module assigns channel-wise weights to selectively enhance critical features while suppressing weakly correlated ones.Finally,a class-balanced loss function replaces the traditional cross-entropy loss function to mitigate the impact of data imbalance on network training.Experimental results show that the proposed model,trained on imbalanced fault data collected from real-world environments,outperforms existing fault identification models with an accuracy of 98.17%.

关 键 词：故障识别 燃气调压器 类平衡损失函数 卷积神经网络 注意力机制 

分 类 号：TH138.52[机械工程—机械制造及自动化] TP206.3[自动化与计算机技术—检测技术与自动化装置] TP306.3[自动化与计算机技术—控制科学与工程]