Research On Fault Image Recognition Algorithm For The Train Fault Detection System

Since2000,Train Fault Detection System (TFDS) has been put into use in major train detection institute gradually. High-speed camera array is placed at the side of railway track to take image of train key components, so train inspector can identify train component fault from the image taken. The system makes the way of train inspection shifted from field operation to room operation. However, the existing train inspection mode means labor intensity is still very high, and is greatly influenced by subjective factors. Therefore, it has great significance to design a train component fault auto-detection system which can improve current fault detection and location method and realize the way shift from artificial detection to automatic recognition.TFDS component fault auto-detection system is divided into five key steps: image preprocessing, feature extraction, feature selection, fault recognition and fault location. First of all, some preprocessing technique, the mean and median filter, Laplacian sharpening, histogram equalization and scale transformation, is applied to solve noise pollution, fuzzy, uneven exposure and image high dimension problems which exist in train component image. This step raises visual contrast of the train part image, reduces the noise and lays a foundation for subsequent fault image identification and positioning. Secondly, train component image texture information is complex, train component has many types of fault and the train part fault can’t be located easily. To solve above problems, the article put forward a kind of texture features extraction method based on gray-level co-occurrence matrix and the method realizes the feature extraction of image local information. And then, an improved Q-ReliefF feature selection algorithm is proposed, the algorithm selects the feature which is closely relevant with image pattern and accomplish the individual character blind selection problem in Relief algorithm. The article adopts k neighborhood, support vector machine and BP neural networks method in Image classification, utilizes image feature selected and identify fault in side-frame, brake beams, bolster, cover plate, lifting rod, bolster spring and other key parts respectively. Finally, the article come up with a kind of SIFT feature match algorithm based on scale invariance, the algorithm solves the problem that component image fault site has considerable difference in geometry and accomplishes the fault site location.In MATLAB platform, the article develops a set of software system based on train fault auto-detection, identification and fault location method. The system has been brought into service in detection system of train detection institute. Experience certificates that the algorithm designed has a certain practical value to train fault identification and location.