A multi-class support vector machine real-time detection system for surface damage of conveyor belts based on visual saliency

Abstract

The primary goal of mining is safety and efficiency. Owing to the long-term operation of a conveyor belt and the damage caused by sharp objects, the belt is easily torn, and such damage can be difficult to discover in time. These issues can cause major accidents in a very short time. Therefore, a reliable, real-time detection of the surface damage on a conveyor belt is critical. Most existing systems have only been applied to images for identifying one type of damage. However, various types of damage can appear in one image, in the form of scratches, cracks, or tears. In this study, a multi-class support vector machine (SVM) detection system is proposed, based on visual saliency. After adding light sources, the system collects images using a charge-coupled device (CCD) camera, and conveys them to a decision-making subsystem via a data transmission subsystem. Then, in a processing module of the decision-making subsystem, the damage is located coarsely using an adaptive threshold, and its connected components are extracted. The grey values are quickly extracted as salient values used to identify the location and type of damage, using the multi-class SVM model. Finally, the system offers a real-time response to the output. The experiments are divided into two groups. One group concerns the detection of the conveyor belt under ideal conditions, whereas the other concerns the belt under wet conditions. As compared with other algorithms, and when there are several types of damage in an image, the method proposed in this study has an improved accuracy rate for all types of damage. This is especially true for tears, for which the detection accuracy reaches 100%. Even when the environmental conditions are complex, the detection accuracy for tears is as high as 99.11%, and the damage can be reported, stopped, and responded to in time to ensure the safety of the person and equipment.