基于MSIF-ECACNN的液压系统故障诊断  

Hydraulic System Fault Diagnosis Based on MSIF-ECACNN

作　　者：李仲兴[1] 陈丽丽 LI Zhongxing;CHEN Lili(School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang Jiangsu 212013,China)

机构地区：[1]江苏大学汽车与交通工程学院,江苏镇江212013

出　　处：《机床与液压》2024年第23期199-206,共8页Machine Tool & Hydraulics

基　　金：智能拖拉机远程故障诊断系统设计与开发校企合作项目(HX00097)。

摘　　要：针对液压信号复杂且难以准确识别的特点,提出一种基于多传感器信息融合的有效通道注意力卷积神经网络模型,分别对液压系统中的液压泵和蓄能器进行故障诊断。该模型采用并行网络结构,针对流量和压力传感器在数量、采样频率上的差异,以及流量和压力信号故障时表现出的不同特点,将多个压力和流量传感器信号分别输入卷积核大小不同的一维多通道卷积神经网络,并利用有效通道注意力调整特征通道权重,在全连接层进行特征融合,最终经Softmax层实现分类。结果表明:有效通道注意力能有效提高故障识别准确率,该方法与目前该领域先进的研究方法相比有更好的故障诊断性能;蓄能器故障诊断精度可达99.52%,液压泵故障诊断精度可达99.88%。同时,该方法解决了因非同源传感器数量和采样频率差异而带来的故障难以准确识别的问题。To address the complex and difficulty to identify hydraulic signals accurately,an effective channel attention convolutional neural network model based on multi-sensor information fusion was proposed to diagnose the pump and accumulator in hydraulic system respectively.In this model,a parallel network structure was adopted.According to the differences in quantity and sampling frequency between flow sensors and pressure sensors,as well as the different characteristics of flow and pressure signal failures,multiple pressure and flow sensor signals were input into 1D multi-channel convolutional neural networks with different convolution kernel sizes,and the effective channel attention was used to adjust the weight of feature channels.Finally,feature fusion was carried out in the full connection layer,and classification was realized by Softmax layer.The results show that effective channel attention can effectively improve the accuracy of fault recognition.Compared with advanced research methods in this field,this method has better fault diagnosis performance,with an accuracy of 99.52%for accumulator fault diagnosis and 99.88%for hydraulic pump fault diagnosis.At the same time,it solves the problem of difficulty in accurately identifying faults caused by differences in the number and sampling frequency of non homologous sensors.

关 键 词：多传感器信息融合 卷积神经网络 有效通道注意力机制 液压系统 故障诊断 

分 类 号：TH137[机械工程—机械制造及自动化]