| With the rapid development of economy and technological progress in the field of railway transportation,railway freight trains begin to change into heavy-load and high-density operation mode.The high efficiency and accuracy of train fault detection are very important for the safe operation of trains.The trouble of moving freight car detection system(TFDS)produces a large number of freight train fault images,which makes the fault detection of freight train images a difficult problem.In the process of fault detection,the freight train image is difficult to locate,the average recognition accuracy is low,and the detection speed is slow.Based on the above background,thesis mainly studies the automatic fault detection of freight train images,with the goal of improving the detection accuracy and speed of the model detection process.The main research contents and contributions of thesis are as follows:(1)An object detection model Bi-directional You Only Look Once(BD-YOLO)is proposed to solve the problem of low average recognition accuracy during detection.The process of BD-YOLO detection is divided into four steps.In the first step,feature extraction network extracts and separates image features.The second step is multi-scale feature fusion to aggregate features of different scales.The third step,prediction across scale is used to predict the characteristic layers obtained after aggregation.The fourth step is to obtain the final prediction result by decoding the prediction feature layer.At the same time,mosaic data enhancement method and Kmeans clustering algorithm were used to improve detection accuracy and speed during training.Experimental results show that compared with the classical object detection model,the m AP of BD-YOLO model improves17.57% on average in freight train test data set.The BD-YOLO model can quickly detect three typical faults,such as upper lever pumps out,deviation of locking plate and closing of handle door plug handle,with high recognition rate,low false detection rate and good robustness.(2)In view of the slow detection speed in the detection process,thesis proposes a lightweight target detection model Fast Channel Attention Network(FCANet)embedded with Fast Channel Attention(FCA)mechanism.FCANet’s feature extraction network is CSPDarknet53 tiny.FCANet embedded two FCA modules to improve detection accuracy.FCANet also includes simplified multi-scale fusion module,cross-scale prediction module and decoding module.Experimental results show that FCANet model has fast detection speed and can accurately identify the above three kinds of typical freight train faults. |
