Research On Defect Detection Algorithm Of Railway Freight Car Rolling Bearing Based On Deep Learning

At present,in the actual disassembly and maintenance of rolling bearing of railway vehicles,it mostly depends on the inspectors to directly observe the bearing surface with naked eyes or touch the bearing surface with hands,so as to judge whether there is damage.This method is not reliable because it requires the inspector to have strong responsibility and is easily affected by fatigue.Although there are many machine vision algorithms used to identify and detect the rolling bearing damage,most of them are based on the traditional target detection algorithms,which have some shortcomings more or less.In this paper,the object detection algorithm based on deep learning is used to detect the bearing damage.Experiments show that this method has better performance than other machine vision algorithms.The main work of this paper is as follows:1.Aiming at the damage of rolling bearing of railway freight car,Ada Boost detection algorithm based on BP neural network is designed.Before training Ada Boost BP model,PCA weight reduction method is used to reduce the dimension of data.The cumulative contribution rate of PCA weight reduction is 85%,and 22 weak classifiers are selected to form a strong classifier.The detection result of Ada Boost BP algorithm for bearing damage shows that the recognition accuracy is 82.15%,and the average single recognition time is 0.98 seconds.2.Aiming at the damage of rolling bearing of railway freight car,sift+SVM algorithm is designed.Firstly,the bearing damage image is grayed and SIFT feature extraction is performed on the preprocessed image.The feature vector obtained after SIFT feature extraction is the description of each key point.The K-means clustering algorithm with k=200 is selected to divide these key points into k categories,and the vocabulary model(bag of word)composed of these classifications is used to describe the bearing damage image.SVM is used to classify the extracted features,and the kernel function of SVM is selected as RBF kernel function through experiments.Sliding window strategy is selected in the damage location method.Sift+SVM algorithm for bearing damage detection results:recognition accuracy is 85.53%,recall rate is82.13%,the average single recognition time is 3.5 seconds.3.Aiming at the damage of rolling bearing of railway freight car,an improved yolov3 algorithm based on deep separable convolution structure is designed.The accuracy rate of the original yolov3 algorithm is 91.54%,the recall rate is 89.64%,and the single detection time is0.31 seconds.In this paper,the depth separable convolution is introduced to improve the yolov3 algorithm,which can improve the detection speed while ensuring the recognition accuracy.In the network model training,simulated annealing algorithm is used to reduce the learning rate to one tenth of the initial learning rate when the number of iterations reaches 5000,which effectively avoids the situation that the model can not converge.The results show that the recognition accuracy is 92.25%,the recall rate is 90.15%,and the average recognition time is0.13 seconds.Through the above experiments,it can be found that the accuracy of bearing damage detection algorithm based on Ada Boost BP algorithm is low,and the recognition speed is relatively slow;The bearing damage detection algorithm based on SIFT + SVM algorithm has high recognition accuracy,but the detection speed is reduced due to the sliding window strategy in the damage location;The bearing damage detection algorithm based on the improved yolov3 algorithm has good performance in both recognition accuracy and recognition speed,and the recall rate is much higher than that of sift+SVM algorithm,which has great advantages compared with the former two traditional target detection algorithms.