Fault Detection Of High-Speed Train Running Gear System Based On Canonical Correlation Analysis

The running gear systems are important components to ensure the stable operations of the high-speed trains.When the trains are in high-intensity services,some parts might be worn out and unpredictably damaged.If these damages are not detected and repaired promptly,they might cause turbulence,unscheduled stop and other faulty operations.In worst scenarios,serious traffic accidents may occur and bring major safety hazards to the passengers or the goods.In this thesis,signal processing in the high-speed trains were selected as the research topic,and the fault detection(FD)challenges of the running gear systems were studied and resolved through various FD models.Targeting the limitations of traditional canonical correlation algorithms in practical fault detection,the innovative enhancement and applications of these FD technologies were implemented.The main content of this research study includes the following aspects:(1)The basic structure and operation mechanism of the high-speed train running gear systems were introduced in Chapter 1.And the working mechanisms of data acquisition of running gear systems were described in Chapter 2.Mainly the causes,theories and evaluation performance of the fault detection models in previous studies were reviewed and discussed,and the experimental data sets of fault signals were identified and extracted from historical databases for analysis.(2)This study proposed a fault detection model,named CCA-JITL,which integrated canonical correlation analysis(CCA)and just-in-time learning(JITL)to process sensor signal data of a high-speed train running gear system.After data preprocessing and normalization,CCA converts the covariance matrix of the high-dimensional historical data into low-dimensional subspaces and maximizes the correlation between the most important latent dimensions.Then,a local fault detection model was established according to the JITL,which used a subset of test samples with larger Euclidean distance to training data.And the feasibility of this algorithm was proved by a series of experiments.(3)In order to further boost the performance of CCA-JITL,a fast Fourier transform(FFT)algorithm was added to the preprocessing procedures to remove noise,which is called as FFT-CCA-JITL.Thus,the influence of noise on fault detection was significantly reduced,and the comparison with CCA-JITL demonstrated that the model improves the detection accuracy,reduces the false positive rate,and speeds up the running time of CCA-JITL and is suitable for fault detection of running gear systems.(4)In this chapter,a fault detection method,so called DNN-CCA-JITL,was proposed similarly to solve the performance problems of the FD model raised bysignal noise and partial correlations among features and variables.Deep Neural Networks(DNNs)were used for overall noise reduction on input data before applying fault detection.And compared with CCA-JITL,the reliability and advanced performance of the algorithm was verified and confirmed by the experiments.