Keywords

1 Introduction

At present, train related fault diagnosis is carried out by specially trained inspectors after the train enters the warehouse. The test results are greatly affected by the quality, emotion and environment of the inspectors, and have great human subjectivity. With the increase of train speed and freight volume, the task of train detection and maintenance is becoming more and more heavy, the traditional manual detection method is time-consuming and laborious, and can’t be detected in real time on line. Obviously, manual detection can’t adapt to train fault diagnosis under the new situation. Along with the rapid development of visual technology, fault diagnosis based on visual detection [1, 2] has been paid more and more attention. The visual technology is applicated in the TFDS system [3] now. How to realize the fast dynamics on-line detection of the corresponding fault problems in the operation of the EMU in the TEDS system is a major challenge for the current scientific and technological workers. During the TEDS system, the wear limit detection of brake shoe is an important fault detection point. Brake shoe is an important part of EMU braking system and one of the key components to ensure the normal operation of EMU. The use of visual detection technology makes automatic real-time on-line fault judgment a possibility, thus making manual maintenance into automatic machine detection, saving labor costs and improving detection efficiency. This paper mainly considers the use of a visual detection technology to achieve the EMU train brake block wear limit fault diagnosis.

TEDS system and TFDS system are similar. They are composed of image acquisition system and image processing system. TEDS (EMU running fault dynamic image detection system), which is installed high-speed acquisition equipment and electromagnetic sensors near the rail, can collect EMU pictures in high-speed operation and transmit them to the terminal server. The fault detection is judged by image processing and pattern recognition technology. A schematic diagram of the TEDS system structure is shown in Fig. 1. In front of the camera, an electromagnetic trigger detector is installed in the direction of the EMU to capture the train arrival information, and then the trigger signal is transmitted to the camera and the flash. It makes the image acquisition equipment into the acquisition state. The three cameras installed at the middle bottom of the rail are used to collect the full coverage image of the bottom of the EMU, and the two high frequency flash lights at the bottom are used to supplement the image machine. On both sides of the track, two cameras are installed to photograph the bogie and the outer compartment.

Fig. 1.
figure 1

Schematic diagram of brake shoe wear detection process

Full size image

After obtaining the train picture of EMU, the distance between the brake pad and the brake pad in the brake pad area is judged by the quick positioning of the brake pad position. When the distance between the two is less than a certain value, the wear limit of the brake pad is judged, otherwise, the fault location realization flow chart is shown in Fig. 1.

The structure of this paper is as follows. The second part a fast positioning method is proposed for the position of the brake shoe of the EMU. The thirth part gives the method of judging the wear degree between the upper and lower brake pads of the brake shoe. Finally, the conclusion is drawn in part 4.

2 Fast Positioning of Brake Shoe

Target location is a prerequisite for fault detection. Because of the use of high-speed, high-rate cameras, the content of an image contains a lot of information. Because the wear of the brake shoe changes little relative to the whole image, an improved template matching algorithm is proposed to locate the target of the brake shoe area.

Template matching [4] algorithm referred to the use of template image slides the window in the whole image, the target position is detected through the maximum response. Supposed an image \( I(x,\,y) \) of size \( M \times N \), in which,\( 0\, < \,x\, < \,M,\,0\, < \,y\, < \,N \), A template image \( w(x,\,y) \) of size \( J\, \times \,K \), in which,\( 0 < J < M,0 < K < N \). Using the relevant matching principle, the response of the template in the image is defined as

$$ c(x,y) = \frac{{\sum\limits_{s} {\sum\limits_{t} {[I(s,t) - I^{'} (s,t))][w(x + s,y + t) - w^{'} )]} } }}{{\left\{ {\sum\limits_{s} {\sum\limits_{t} {[I(s,t) - I^{'} (s,t))]^{2} \sum\limits_{s} {\sum\limits_{t} {[w(x + s,y + t) - w^{'} )]^{2} } } } } } \right\}^{{\frac{1}{2}}} }} $$
(1)

in which \( x = 0,1, \cdots ,M - 1 \), \( y = 0,1, \cdots ,N - 1 \), \( w^{'} \) is the average value of pixels in window \( w \), \( I^{'} (s,\,t) \) is the regional average value in \( I(s,\,t) \), which coincides with the current position in \( w \), C is sliding window in the whole image. The algorithm is simple in principle, but when the image size is very large, the calculation will be very large and the operation will be very slow. Therefore, combined with image pyramid processing, an improved fast template matching algorithm is proposed. that is the target position is roughly located by implementing image template matching from low resolution. Then the second template matching is carried out in the rough target location area of the next pyramid image. The location result diagram is shown in Fig. 2, and the two methods match the time result pairs such as Table 1.

Fig. 2.
figure 2

Location of the brake pad area

Full size image
Table 1. Time-consuming comparison of matching algorithms
Full size table

3 Judgment on the Wear of Brake Watt

3.1 Linear Segmentation of Image Edges

The image edge contains the most basic feature information in the image. And it is the fastest pixel in the gray value of the image.

After accurately calculating the position of the brake shoe, the wear limit of the brake shoe is judged by the distance between the brake pads. An accurate, fast linear segmentation detector LSD [3] is used to extract the edge information of an image. LSD is a fast edge extraction algorithm with linear time. It generates horizontal line interval by calculating the horizontal line angle of pixels, and then fuses pixels with the same horizontal line in the interval to form the supporting interval of edge line. Finally, the accurate extraction of image edge is realized by segmentation correction. The pseudo code of the algorithm is as follows:

By LSD algorithm, the edge segmentation image of the brake shoe is shown in Fig. 3. Because of the rigid connection between the gate and the fastening bolt, the distance between the center point of the fastening bolt at both ends of the gate can be easily judged by using the Hough transformation to realize the detection of the circle in the image of the brake shoe. When the center distance between the upper and lower two gates is less than the normal threshold between the two, the wear limit of the brake shoe is judged. Otherwise, it is judged as no fault. The detection effect of the fixed bolt is shown in Fig. 4.

Fig. 3.
figure 3

LSD Linear segmentation result

Full size image
Fig. 4.
figure 4

Detection of fixed bolts at both ends of the gate

Full size image

4 Conclusion

In this paper, an on-line detection method based on machine vision is proposed for fault detection of brake shoe wear limit of EMU train. By using the fast template matching algorithm and the LSD edge extraction method, the wear limit of the brake shoe is realized by identifying the center pixel distance of the fixed bolt at both ends of the brake shoe. The method has high measurement accuracy and real-time performance, and it can be applied to the on-line detection of excessive wear limit of EMU train brake shoe.