Research On Intelligent Recognition Technology Of Fault Image Of High-speed EMU Components

With the rapid development of high-speed trains,the fault detection of trains has become particularly important.The existing high-speed train fault detection system adopts the detection method of "human inspection as the mainstay",which has the problem of low efficiency and low detection accuracy,and how to realize the automatic detection of train faults is an urgent problem to be solved.In recent years,deep learning has developed rapidly.Compared with the traditional image recognition algorithm,deep learning-based image recognition technology has higher speed and detection accuracy,and it has been rapidly applied and developed in many fields.In view of the above,this paper applies the image recognition technology based on deep learning to the fault detection of high-speed trains,and proposes two new improved algorithms to detect the typical faults of bolts and oil pollution.The contents are as follows.(1)An improved YOLOv4 algorithm is proposed for detecting bolt fault images.Through the analysis of bolt-images’ feature,this paper selects a target detection model based on deep learning for fault image detection.And YOLOv4 was selected as the basic detection model by analyzing and comparing the mainstream object detection models.Aiming at the problem of small size bolt detection,an improved YOLOv4 algorithm is proposed.By introducing the attention mechanism module,the network deepens the feature extraction ability of small bolts.The K-means++ clustering algorithm is used to re-determine the size of the bounding box,and the detection effect of bolts of different sizes is improved.The data enhancement method and the training method based on transfer learning are used to improve the generalization ability of the model and speed up the network training.After the training was completed,through experimental comparison with faster RCNN model,YOLOv3 model,and original YOLOv4 model,the improved YOLOv4 algorithm showed superior performance in terms of detection accuracy and speed.The average accuracy reached 95.3% on this dataset,and the missed detection rate of lost bolts was reduced from 4.38% to 0.73%.(2)An improved DeepLabv3+ model is proposed for oil image detection.Through the characteristic analysis of the oil image,the image segmentation model based on deep learning is selected for the detection of the oil spills image.Through the preliminary test of the mainstream image segmentation model,the superior DeepLabv3+ model was selected as the basic segmentation model.Based on this,the improved DeepLabv3+ model is proposed: modify the loss function,replace the original cross-entropy loss function with the focus loss function to improve the segmentation accuracy of the network;for the light sensitivity of the oil pollution image at the bottom of the train,a software module for light-contrast adjustment is designed to amplify the oil pollution data under different lighting conditions and improve the generalization ability of the model;the original Xception backbone network is replaced by a more lightweight Mobiletv2 network to reduce the network parameters and accelerate inferencing.Through experimental comparison with the U-Net network and the original DeepLabv3+ network,the optimized DeepLabv3+ model has a slight improvement in the segmentation accuracy,and its detection speed reaches 4 times that of the original network,which greatly accelerates the model inference speed when ensuring the accuracy rate.