The rapid development of science and technology promotes the high demand and high standard for machinery and equipment in rail transit.As one of the most common parts in mechanical equipment,the working state of rolling bearing directly determines whether the whole equipment can work normally,and even relates to the normal operation of the whole production line.Therefore,the fault diagnosis of rolling bearing plays a vital role in the operation of rail transit.However,in practice,most of the working state of the rolling bearing is normal,this always causes the lack of fault data,resulting in bearing fault diagnosis data of unbalanced state,and the class of the imbalance conditions inevitably lead to classifier trend to the most of the class of more samples.And the minority class samples cannot be correctly distinguish from the data set.It will affects the accuracy of bearing fault diagnosis.In addition,it is common to see unbalanced data in both industrial and academic research,such as malicious brushing,scalping orders,credit card fraud,equipment malfunctions and so on.The percentage of credit card fraud is generally less than 1 in 1,000.Therefore,strengthening the fault diagnosis of rolling bearing and improving the recognition rate of small samples are conducive to improving the accuracy of bearing fault diagnosis.Therefore,bearing fault diagnosis on unbalanced data is a subject worthy in-depth research.In this paper,the main research contents are listed as follows: Firstly,propose a new stacked under-sampling method(UDTSS).The original unbalanced data is twice stacked to form a multi-channel multi-dimensional vector data set and the under-sampling operation is carried out to balance the number of samples of the majority class and the minority class,so as to improve the overall classification accuracy.Then the first dimension reduction is done through PCA to achieve the overall noise reduction purpose.Then,the second dimension reduction was carried out for different types through LDA.Finally,a fault diagnosis model(UDTSS-CNN)based on the new method was constructed through the convolutional neural network(S-CNN)combined with SVM classifier to diagnose the rolling bearing faults of unbalanced samples.Finally,through experimental comparison,the new stacked undersampling algorithm proposed in this paper achieves a good effect of balancing data,and the diagnostic accuracy of the model for unbalanced data also achieves a significant improvement.Experimental results show that the UDTSS-CNN model combined with the new stackable under-sampling method has high efficiency and accuracy of diagnosis rolling bearing fault problems based on unbalanced data. |