| In the field of fault diagnosis,the research objects of current deep learning methods are focused on single faults,but few are involved in compound faults.Gearbox compound fault vibration signals are coupled,strong and weak signals are mutually submerged,and frequencydomain signals cannot effectively provide features,so it is difficult to diagnose accurately.In the face of complex compound faults,the traditional Softmax classifier + cross-entropy function model does not effectively use the correlation information between homogeneous samples and heterogeneous samples,and is unable to complete the fault diagnosis task accurately and efficiently.This paper proposes a two-dimensional deep measurement learning convolutional neural network(DML-CNN)model.Its advantage is that the model uses the convolutional neural network to extract features in the time-frequency diagram of fault signals.Triplet loss as a loss function will further measure the distance between sample features in the classification process,so as to learn the embedded relationship between features,so that the feature distance corresponding to the samples within the class is closer,and the feature distance corresponding to the samples between classes is farther.In order to verify the superiority of the DML-CNN model compared to the traditional Softmax classifier + cross-entropy function model in the face of composite faults,a large number of comparative experiments are performed in this paper.In the data set divided by percentage,the DML-CNN model’s recognition accuracy for composite faults reached 96.82%,and the traditional model reached 94.26%.In order to better fit the phenomenon that different operating conditions(speed,load)will affect the fault signal in actual industrial production,this paper makes separate data sets divided by speed and load to simulate the lack of operating condition information.In the data set segmented by load,the accuracy of the DML-CNN model is still 95.57%,while the traditional model is 91.02%.However,in the data set segmented by speed,the traditional model has a serious overfitting phenomenon,only 67.14%,indicating that the model has poor adaptability to rotational speed,while DML-CNN still has an accuracy of 90.09%.In order to visualize the model’s feature extraction capability and measure the distance between features,this paper uses deep learning visualization technology to show the convolution kernel and classification results of deep metric learning models.As the accuracy of the DML-CNN model decreases to varying degrees in the training experiments with missing data,the traditional model has poor generalization ability and is prone to overfitting.It is also related to whether deep learning is based on big data.Can the model be effective? The features learned in the data depend heavily on the quality and quantity of the data.Therefore,it is particularly important to enhance the diversity and quantity of data.In view of the small sample volume and difficult to collect under specific working conditions in actual industrial production,this paper uses Deep Convolutional Generative Adversarial Neural Network(DCGAN).The adversarial learning mechanism expands the data volume of small samples,and solves the problem of poor generalization ability and low recognition accuracy caused by insufficient data volume or uneven sample distribution. |
