Study On Preprocessing And Feature Extraction Of High Speed Rail Vibration Data Based On Cloud Computing

With the further speed raise of China High Speed Train and improvement of the information system, more signal data can be collected nowadays. Currently lots of sensors are deployed on the running train to collect a variety of data. The vibration data is one kind of them. Vibration affects not only the travelling comfort experience, but also reflects the running state of the train components. Hence analysis and processing high-speed vibration data is of great significance to enhance the service and ensure safety of the trains. However, traditional ways of extracting feature and processing the vibration data are based on stand-alone machine, which are vulnerable facing the large amount data collected by numerous sensors. In specific, computing time and manual intervention are suffering and they are unable to handle large files. Cloud computing technology offers a solution to these problems. MapReduce is a major cloud computing model, which can automatically assign task and balance the load. It is regarded as a convenient and effective cloud computing framework. Thus, it is of important pratical value in the application of MapReduce to analyze the vibration, to enhance the processing ability of the vibration data and to respond to challenge of big data. This thesis aims at studying algorithms for preprocessing the high speed vibration data and extracting features from the vibration data.The major contributions of this thesis are as follows. The parallel computing platform based Hadoop is established as preparation. After carefully studying the preprocessing methods, two algorithms based on MapReduce are proposed for preprocessing the original vibration data. Outlier removal algorithm locates the outliers and processes them. Linear trend removal algorithm removes the linear drift from the original data. After preprocessing work, we present a MapReduce algorithm to separate the mixed data into a unique data file. Each file corresponds to one single data channel and the data sequence is kept unchanged. Then, we extract the data feature of the vibration data using MapReduce. At last, on the one hand, the data distribution of the vibration data is studied. We find that its parameter varies when different vibration components become invalid. The result can contribute to the fault diagnosis of the vibration components. On the other hand, the extracted features are used to mine the relation between the train speed and standard deviation of the vibration data with the linear regression. The vibration data can be classified into three categories, namely, the normal data, the data to be observed and the alarming data. This work makes the processing of vibration data more targeted and then the processing efficiency is improved.All algorithms in this thesis are verified on Hadoop parallel computing platform to evaluate their parallelism. The experimental results show that the parallel algorithms perform well and are capable of performing the task for analysis of big high-speed rail vibration data.