Blind Source Separation And Time Series Prediction Of Environmental Vibration Data From Urban Metro Vehicles

With the development of urban metro,the environmental vibration caused by the metro has become increasingly prominent.The installation of accelerometer s to monitor environmental vibration is an effective technique for grasping environmental vibration and evaluating environmental comfort,building safety.However,the vibration signals acquired along the subway include not only the vibration caused by the metro,but also the vibration caused by other factors.This article refers to the former as the metro-induced vibration,and the latter as the non-metro-induced vibration.How to separate the two kinds of vibrations so as to conduct the the metro-induced vibration is an important issue that needs to be solved.In this paper,based on independent component analysis principle,the Fast ICA method is used to blindly separate the vibration data collected along the subway in a certain city.After evaluating the results of the separation,the metro-induced vibration was used for the time series study.The paper mainly discusses from the following three aspects:(1)For measured data,the preconditions for the blind source separation using the Fast ICA method is validated,a Butterworth bandpass filter bank is desighned to filter the raw data into components in the frequency domain.Th ese components were applied to Fast ICA analysis,and the independent subcomponents were represented by the sub-spectral basis vector A and the time-course partial coefficient W;The two sets of test data obtained 83 and 86 ICA subcomponents,the narrow-band ICA sub-components were clustered by the NCUT method into metro-induced vibration and non-metro-induced vibration;One set of test data espectively contained 64 metro-induced vibration ICA subcomponents and 19 non-metro-induced vibration ICA subcomponents,another set of test data espectively contained 60 metro-induced vibration ICA subcomponents and 26 non-metro-induced vibration ICA subcomponents;(2)Based on the principle of BP neural network,BP neural network is constructed by four statistical features of the signal: kurtosis,RMS,and peak frequency.The classifier is designed based on BP neural network toolbox of Matlab.500 metro-induced vibration signals,500 mixed vibration signals,and 500 non-metro-induced vibration signals are regarded as the learning sample sets,metro-induced vibrations signal subcomponents and non-metro-induced signal obtained form ICA were put into the classifier.The results were divided into three types: excellent,good,and poor.Excellent and good ratios in the two sets of test data accounted for 80% and 75% of the total sample number,respectively,demonstrating that the Fast ICA program can operate normally;(3)The metro-induced vibration obtained by blind source data separation is used for time series prediction.With the first three segments of data,The forth segment of data in the time domain were successfully predicted by the WNN time series prediction program.The results of the time series prediction are added to the sample set for learning to evaluate the prediction algorithm until the prediction of the seventh segment of data was failed.