Research On Fault Diagnosis Of High Speed Running Gear Based On Deep Learning Under Cloud Platform

With the development of High Speed Train (HST), its security issues have received increasingly attention. HST running gear as an important component of the train will be directly concerned with the safety of HST. Therefore, it is significant for HST to identify faults of the train running gear. On the vehicle body, corbel, axle box and some other place of the train, a large number of many types of sensors are fixed to collect vibration data, through which we can determine the operational status of running gear. However, how to quickly and efficiently extract features from these massive vibration data and fault diagnosis is a problem need to be solved.On the one hand, deep learning is one of the most powerful data characteristic expression technologies and Deep Belief Networks (DBNs) are pioneers in building this kind of deep architectures. It is considered as a highly complex nonlinear feature extractor in which each hidden layer learns to represent features that acquire higher order correlations in the original input data. It brings hope to solve the deep structure of the related optimization problems in handling highly complex data structures, outstanding performance and other data processing large volumes of data. This thesis analyzes the time domain and frequency domain characteristics of vibration signals of several conditions firstly. On this basis, an approach based on FFT-DBNs of HST running gear feature extraction and fault diagnosis is proposed. The approach automatically extracts high-level features from HST vibration signals and recognizes the faults. DBNs are trained greedily, layer by layer, using a model referred to as a Restricted Boltzmann Machine (RBM). The vibration signals are preprocessed by FFT and the FFT coefficient-vectors are used to set the states of the visible units of DBNs. N label units are connected to the "top" layer of the DBNs to identify different faults. The experimental results show that the method may learn useful high-level features from vibration signals and diagnose the different faults of HST. Then, by combining the advantages of KNN and DBNs, an improved K-DBNs deep learning algorithm is proposed, and simulation experiment is carried out. Finally, a method of optimizing each layer in deep network is introduced and conducted simulation experiment.On the other hand, a huge challenge is confronted when the traditional vibration signal processing way is used for handling the increasing big data. However, cloud computing as a new Internet-based computer technology, has excellent performance in large data computing and network storage. On the Hadoop platform, HST vibration data analysis and processing method is realized using the Spark-based distributed cluster framework which has fault-tolerant processing, automatic load balancing and so on. It offers a solution to these problems during the procedure of HST vibration data processing, such as the incomplete use of the original data and slow processing speed. Furthermore, the parallel efficiency is discussed. Experimental results show that the proposed parallel algorithm design performs well and is capable of performing the task for analysis of big HST vibration data.