Research On Non-negative Tensor Factorization Based Rail Defect Detection Algorithms

The development of high-speed railway is of great strategic significance to theeconomic growth in China. However, the development of high-speed railway musttake safety as a fundamental premise. Without safety, the value of high-speedrailway is worth nothing. As the rail is a lynchpin of the railway, it should be paidmore attention to. Currently, the rail faults detection is deficient in China. Themethod of detecting the rail flaws with detecting vechiles and hand-held devices cannot meet the growing demand of rail defect detection. In order to improve thissituation, this dissertation proposes a method based on nonnegative tensorfactorization algorithm for real-time monitoring of rail defect.In the first place, the physical and mathematical model of vertical system ofhigh-speed railway are confirmed. The predicted-correction method is applied insovling the vertical vibration model of the vechile-track coupling system. Fourkinds of excitation models for rail defect are dicussed here. The vibration signalsare obtained by inputting the excitation models for rail defect to the vechile-trackcoupling system.Then a method based on nonnegative matrix factorization and support vectormachine for telling whether the defect occurs or not is proposed. Sparseness and ainitialization strategy based on singular value decomposition are introduced forimproving the traditional nonnegative matrix factorization. The improved method isused to extract the feature of vibration signals. After the feature extraction, supportvector machine is employed to tell whether the defect occurs or not.Eventually, a method based on nonnegative tensor factorization and extremelearning machine is proposed to classify different kinds of rail defects. Ainitialization strategy based on singular value decomposition is introduced forimproving the nonnegative tensor factorization. The wavelet analysis method isapplied to extract local characteristics in time and frequency domain of the signalsobtained from the rail with defect. Combining the time-frequency feature withdifferent carriages, a three-dimensional tensor is constructed. Then the improvednonnegative tensor factorization is employed to mine the hidden information.Meanwhile, extreme learning machine is introduced to the classification of raildefects.The two method mentioned above will be applied to different phases of thereal-time monitoring of railway. The method based on nonnegative matrixfactorization and support vector machine for telling whether the defect occurs or not is carried out by the microprocessor of the sensor node. It can tell whether thedefects occur or not right after the colletion of vibration signals. If the senor nodesconfirm the rail defect occurs, the signals will be sent to the information centre forthe detail analysis. Then the method based on nonnegative tensor factorization andextreme learning machine will confirm which kind of defect occurs. The twomethods supplement each other and both can achieve great performance accordingto their own characteristics.