Stady Of Detection Technology On Railway Freight Car Rolling Bearing Surface Defects Based On Machine Vision

Bearing, an important role of railway wagon, has a potential safety hazard in the operation of the truck when there's surface defects. But in China, part of railway freight car rolling bearing detection still by human now, which weakness in low working efficiency, high cost and difficult to ensure accuracy. In this thesis, the railway wagon bearing as the research object, bearing image earmark was took out and dealt by using the image processing technology, then classify the defects by its feature.Firstly, this thesis introduces the common type, morphology and formation principle of surface defects, designing the structure of the detection identification by using the way of gray transform to enhance the image contrast. By comparing the methods of mean filter, median filter and wiener filter to de-noise bearing image, the result is that the way of wiener filter is more efficient, also use the traditional method of wavelet transform for image processing, when under uneven illumination to de-noise bearing image, the traditional wavelet threshold method's effect is actually modest. Aimed at this situation, a method of wavelet de-noise and enhancement was put forward to improved threshold function, which firstly using wavelet transform to multiscale decomposition the bearing image, then de-noise and enhancement the first level of wavelet coefficients based on the modified algorithm to improve the image quality, second and third level of wavelet coefficients following.The initial segmentation processing of image is based on the variable threshold processing of the local image characteristics. In image edge detection, this thesis analyses edge detecting computing speed via Roberts, Prewitt, Sobel, Canny and SUSAN algorithms, it turns out that the computing speed of Sobel is more faster than others. Under the appropriate threshold, the way of SUSAN is more efficient. Often the noise is considered to be the edge problem in the edge detection of noisy image. So, this dissertation improves the wavelet edge detection algorithm to enhance the accuracy of edge detection.Recognition of defects by using BP neural network. First of all the input and output nodes of the BP neural network model are established by the empirical formula. Secondly, the established BP neural network is trained, then using it to identify and classify five types of defect image segmentation edge detection. In the process of bearing defect identification, the average area, the solid degree, the circle ratio and the Euler number of defects are selected as the recognition features.The 5 types of defect image that was collected are identified by the experiment methods, and the accuracy rate of identification even can reach to 90%.