深度残差收缩网络的多特征锅炉炉管声波信号故障识别  

Fault identification of multi-feature boiler tube acoustic signalbased on deep residual shrinkage network

作　　者：杨正理[1] 吴馥云 陈海霞[1] YANG Zhengli;WU Fuyun;CHEN Haixia(School of Mechanical and Electrical Engineering,Sanjiang University,Nanjing 210012,China)

机构地区：[1]三江学院机械与电气工程学院,江苏南京210012

出　　处：《智能系统学报》2023年第5期1108-1116,共9页CAAI Transactions on Intelligent Systems

基　　金：江苏省高校自然科学基金面上项目(20KJB470029)。

摘　　要：为了提高锅炉炉管声波信号故障识别的学习效果和识别精度,采用特征向量并行和拼接两种融合方式构成特征层,以及平均得分和最大值得分两种融合方式构建决策层等不同信息融合机制,提出基于深度残差收缩网络的多特征锅炉炉管声波信号故障识别方法。首先,考虑锅炉炉管上各声波传感器的差异性,分别计算声波信号谱特征一阶和二阶差分构建三通道特征集作为二维网络的输入特征向量;然后,在卷积神经网络和双向长短时记忆网络基础上引入注意力机制构建基线模型,并采用深度残差收缩网络对二维网络的通道权重进行优化分配,提高模型的故障识别精度。大量实验结果表明:采用特征向量并行融合方式构成特征层的信息融合机制是一种更有效的策略;本文模型的识别精度得到较大程度提高,与基线模型相比较,未加权平均召回率提高了4.32%。To improve the learning effect and recognition accuracy in fault identification of boiler tube acoustic signals,different information fusion mechanisms were used in this study.Specifically,feature vector parallelism and splicing techniques are used to form the feature layer,while the decision layer is formed using average and maximum score ap-proaches.A fault identification method of multifeature boiler tube acoustic signal based on a deep residual shrinkage network is proposed.First,the first-order and second-order differential characteristics of the acoustic signal spectrum were calculated separately,considering the variations of acoustic wave sensors on the boiler tube.This process con-structs a three-channel feature set,which serves as the input feature vector of the two-dimensional network.Sub-sequently,an attentional mechanism is introduced to construct a baseline model using a combination of a convolutional neural network and a bidirectional short and long memory network.Additionally,a deep residual shrinkage network is used to optimize the allocation of channel weights within the two-dimensional network to improve the fault identifica-tion accuracy of the model.Extensive experimental results show that:it is a more effective strategy to construct the in-formation fusion mechanism of the feature layer by using feature vector parallel fusion;compared with the baseline model,the unweighted average recall rate has increased by 4.32%,highlighting significant improvements in the recogni-tion accuracy of the model presented in this paper.

关 键 词：深度学习 故障识别 深度残差收缩网络 双向长短时记忆网络 注意力机制 卷积神经网络 锅炉炉管 声波信号 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN912.3[自动化与计算机技术—计算机科学与技术] TM621.2[电子电信—通信与信息系统]