An identification method of catenary key components based on deep learning

Abstract

The invention discloses a method for identifying key components of catenary based on deep learning, which includes the following main steps: (1) target image acquisition; (2) image data enhancement; (3) image layered recognition; (4) key components and Component feature recognition; the step (3) includes two steps of (3a) image region recognition and (3e) key component recognition network model training. Aiming at the characteristics of the catenary image set, the present invention adopts a variety of methods such as data enhancement, layered recognition and error set fine-tuning, which mainly have the following advantages: improve the recognition rate of small parts; reduce the loss in calculation; reduce the amount of labeling , improve the labeling efficiency; improve the generalization ability of the network; significantly improve the network recognition rate of the catenary image network, and the established model can be easily converted between different lines. When the model is applied to a new line, only targeted fine-tuning can be carried out. It saves a lot of design, labeling, and training work, and facilitates manual and automatic detection work.

Description

technical field [0001] The invention relates to the field of catenary defect detection, in particular to a method for identifying key components of catenary based on deep learning. Background technique [0002] The railway transportation industry has been in the process of continuous development. In recent years, in order to meet the needs of the rapid development of Chinese society, the railway system has been increasing in speed. At the same time, it has brought about more and more important safety issues: catenary system distribution Extensive, and usually located in the open-air system, due to the harsh working environment, the support and suspension components are prone to various defects. Once some small problems occur, chain reactions are likely to occur, causing serious safety accidents and threatening the safety and security of the people. The development of social production. [0003] At present, the railway bureaus mainly use the method of manual inspection to an...

AI Technical Summary

Problems solved by technology

[0002] The railway transportation industry has been in the process of continuous development. In recent years, in order to meet the needs of the rapid development of Chinese society, the railway system has been increasing in speed. At the same time, it has brought about more and more important safety issues: catenary system distribution Extensive, and usually located in the open-air system, due to the harsh working environment, the support and suspension components are prone to various defects. Once some small problems occur, chain reactions are likely to occur, causing serious safety accidents and threatening the safety and security of the people. development of social production

[0003] At present, the railway bureaus mainly use manual viewing to analyze the massive images collected by the "catenary suspension state detection and monitoring device" (4C), and manually identify those places where there are potential safety hazards on the line

Due to some human factors, plus the problems of overexposure and low contrast in catenary pictures, it is difficult to achieve high-efficiency and high-reliability detection for massive offline pictures manually, and it is also difficult to obtain an effective detection frequency. promote

In addition, the traditional automatic detection algorithm currently used often designs features manually, and then sends the features into a more complex classifier for classification. Due to the large amount of problems, it is difficult to manually extract features, the reliability of the recognition results is low, and it is not suitable for such a complex environment as catenary.

Therefore, the traditional algorithm has high false detection rate and missed detection rate, and the generalization ability of the algorithm is poor.

[0004] Deep learning has developed rapidly in the field of image recognition and achieved good results. However, the existing technology only builds a single-layer Faster R-CNN network model without performing hierarchical recognition. The samples obtained by image data enhancement are diverse. Insufficient performance, the recognition rate of the key components of the complex catenary can not meet the requirements, the generalization ability is poor, and because there is no image scaling processing, the training recognition efficiency is relatively low

Images