Real-time sleeper defect detection method based on pruning algorithm

Abstract

The invention discloses a real-time sleeper defect detection method based on a pruning algorithm. The method specifically comprises the steps that a high-speed camera obtains a sleeper image; image data are enhanced and expanded, and a training set is made according to the number of normal sleepers, sleeper falling blocks and sleeper cracks being 1: 1: 1; marking the training set; carrying out model basic training on the training set by adopting a deep learning YOLOv3 target detection algorithm; carrying out sparse training on the model, and judging the importance of each channel and layer in the sleeper defect detection network by utilizing a gamma coefficient of a BN layer; importance sorting is carried out, and a pruning proportion is set; pruning unimportant channels and layers of the network; and finely adjusting the pruned model. On the system level, the sleeper defect detection speed is effectively increased, and deployment in embedded equipment and online detection can be achieved; in the model level, the importance of each channel of the network is judged through the gamma coefficient, and the method has the advantages of simple principle and simplicity and convenience in operation.

Description

technical field [0001] The invention belongs to the technical field of sleeper defect detection, in particular to a real-time defect detection method for sleepers based on a pruning algorithm. Background technique [0002] The sleeper is one of the important parts of the track, which has the function of supporting the rail, maintaining the gauge and direction, and transmitting the pressure of the rail to it in all directions to the track bed. Therefore, the sleeper must be strong, elastic and durable, and the missing block of the sleeper (see image 3 ) and cracks (see figure 2 ) will greatly reduce the effect of sleepers. Defects such as missing sleeper blocks and sleeper cracks are subtle and difficult to detect but have huge potential hazards. If they are not repaired for a long time, they may lead to problems such as deformation, displacement and collapse of the rails, and in severe cases, it will lead to major traffic accidents such as train derailment. [0003] At p...

AI Technical Summary

Problems solved by technology

However, these detection algorithms have the problems of slow detection speed and difficult model deployment. These detection methods have a large amount of computation and require expensive computing equipment, and the resulting orbit detection system is bulky and heavy, making it difficult to achieve large-scale applications. deploy

Images