Conjoint localization dense networks for fault feature extraction of variable load gearbox
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.
万方数据
维普
