基于联合定位密集网络的变载齿轮箱故障特征提取

Conjoint localization dense networks for fault feature extraction of variable load gearbox

为解决变载齿轮箱故障诊断中因冗余特征而导致的脉冲信号提取难题,提出一种基于注意力模块(CAM)的脉冲特征提取方法。首先,设计联合定位CAM,包括2个阶段:第1阶段使用多层感知机建模通道依赖关系,增强与故障相关的关键通道特征;第2阶段通过卷积层学习与故障相关的信号段,结合2个阶段重新校准特征,聚焦关键脉冲特征;然后,基于CAM构建联合定位密集网络(CLDN)的变载齿轮箱故障特征提取方法,CLDN通过自适应地重新校准每一层的特征,进一步提高对脉冲信号的学习和表征能力;最后,将提取到的特征输入Softmax分类器,验证所提方法的特征提取效果。结果表明:相比于Self-Attention等4种注意力机制,CAM的准确率平均提升3.8%,可实现脉冲特征的准确定位;相比于ResNet34等7种诊断方法,CLDN的准确率提升3.7%~14.6%,显著增强故障特征的提取效果。

To address the challenge of extracting pulse signals in fault diagnosis of variable-load gearbox caused by redundant features,a pulse feature extraction method based on CAM was proposed.First,a CAM was designed,which consisted of two stages.In the first stage,a multilayer perceptron was used to simulate the channel dependencies and enhanced the important channel features related to faults.In the second stage,the convolutional layers were employed to learn signal segments related to faults.By recalibrating the features in two stages,the module focused on the critical pulse features.Next,based on CAM,this study proposed a CLDN method for extracting fault features in variable-load gearboxes.CLDN further improved the learning and representation of impulse signals by adaptively recalibrating the features at each layer.Finally,the extracted features were fed into a Softmax classifier to validate the feature extraction effect of the proposed method.The results show that CAM's accuracy is on average 3.8%higher than 4 attention mechanisms like Self-Attention,achieving accurate localization of impulse features.Compared with 7 diagnostic methods such as ResNet34,the accuracy of CLDN is 3.7%to 14.6%higher,which significantly enhances the extraction of fault features.