Research On Fault Diagnosis Method Of High-speed Train Running Gear System Based On Broad Learning

The running gear system not only has the inherent characteristics of complex structure,multiple units and modules,but also has the external characteristics of the observation information,such as uncertainty,ignorance and incomplete.In addition,running gear is the core system of high-speed train,so the modeling accuracy of its evaluation model is required to be high.The traditional method will be affected by internal and external factors of the system,resulting in poor evaluation effect.In the running gear system,the fault probability and fault severity are higher than those in the simple system.The occurrence of faults has the properties of randomness,secondary,concurrent,explosive,concealment,etc.Different fault components have different ways and degrees of influence on the fault characteristics,and the sensitive characteristic parameters of various faults may be irrelevant.The corresponding relationship between fault characteristics and fault mode categories is unclear,and the mechanism model is difficult to be established.Therefore,a data-driven method based on broad learning system is proposed in this thesis.The advantages of this method are that it does not need to establish complex high-speed train data model,and it can process non-Gaussian data with high sampling rate.Compared with the previous method based on neural network,it can effectively improve the computational efficiency and ensure good fault detection ability.The research in this thesis is mainly divided into the following two points:(1)This thesis firstly proposes a fault detection method based on broad learning system and convolutional neural network combined with auxiliary principal component analysis.This method can deal with nonlinear and non-Gaussian data,which is difficult to be processed by traditional multivariate statistical methods.Two kinds of neural networks are used to fit the offline data,and principal component analysis is used to process the output data of the neural network.Finally,the reliability of this data-driven method is verified in the simulation experiment,For the three classic faults,the fault detection rate is all above 90%.(2)This thesis proposes a data-driven approach based on broad learning system assisted canonical correlation analysis.Due to the high sampling frequency of the sensor under actual working conditions,and the measured data does not comply with the Gaussian distribution,it is difficult for the general method based on neural network or multivariate statistics to obtain the ideal fault detection effect.In this thesis,broad learning system is used to enhance the robustness of the algorithm,which theoretically proves that the method has excellent fault diagnosis ability.The golden section method is used to select the appropriate threshold faster,which can maximize the fault detection rate of the algorithm while ensuring the acceptable false positive rate.This threshold is suitable for processing data with high sampling rates.Compared with the previous neural network-based method,it can effectively improve the computing efficiency and ensure good fault detection ability.Finally,the effectiveness and feasibility of the proposed method are verified on the simulation platform of high-speed train traction drive control systems,the false alarm rate decreased to 0% when no faults were injected.