Study Of Railway Track Irregularity Time Series Data Mining And Trend Forecasting

Track condition is directly related to the safe operation of trains. Track irregularity is not only an important indicator of track condition, but also the basis for measuring the quality of train operation. Where there is track irregularity, speed limit should be paid attention, and worst of all, overturning might occur. As a result, it is urgent for railway departments to study the law of track irregularity changes so as to master trends of track state changes and to take prevention measures. In this context, this doctoral dissertation analyses track irregularity data, explores the underlined rules of track irregularity, predicts future trends and ultimately, it provides support to relevant railway departments in models about data and track state changes, to ensure safety of railway transportation.During the track irregularity data analysis, this doctoral dissertation first analyzes the characteristics of track irregularity data systematically based on data mining, employs clustering method to recognize abnormal data, proposes data variation calibration algorithm based on trends similarity, and algorithm of partial abnormal data recognition and noise elimination based on abnormality, preprocesses data. Next, wavelet decomposition and reconstruction model is proposed, so as to lay the data base for modeling for track irregularity time series. Finally, since track irregularity data reflects the dynamic characteristics of track state changes, it is an important temporal data. This doctoral dissertation conducts clustering analysis on track irregularity sequence and discovers pattern features with the application of data mining and algorithms. In concrete research, it conducts case analysis of data clustering based on the original data, standard deviation data, and standard deviation data after wavelet decomposition, similar standard deviation sequence, and finally discovers the trend in data changes.During the track state prediction analysis, this doctoral dissertation first focuses on the inertia characteristics of the track state changes. Since track state has a memory effect, latest track state and the nearest previous state shares similarity, and the inspection state of the adjacent time points has a similar trend. From the macro perspective, the track state presents nonlinear changes throughout the whole life cycle of the track, but from the micro perspective, in a short time, track state changes at adjacent time will be close to linear features. Based on this assumption, combined with the non-isochronous features of track irregularity time series data, this paper proposes Piecewise Linear Recursive Model based on Short-term Historical Trends model (PLRMSHT). Also, to improve the prediction accuracy of PLRMSHT, residual correction is conducted based on Fourier delta transformation. Through case studies, models have achieved satisfactory results. Next, since wavelet transform can refine signal (function) in multiple aspects gradually through local analysis and stretching and panning on space (time) frequency, and ultimately realize time subdivision at high frequency points and frequency subdivision at low frequency points. This feature can adapt to signal analysis of time-frequency, and can focus on every detail of signal, and can divide the function into a series of simple basis functions. As a result, it is of significance both in theory and in practice. Therefore, based on wavelet decomposition-reconstruction, the thesis divides track irregularity time series, and finds the best fit forecast model to the details signal and the approximate signal of the wavelet decomposition. During concrete research, ARIMA forecast model after residual correction is of higher accuracy compared to the original ARIMA forecast model, but the correction itself increases the amount and complexity of computing, and can not reflect the inadequacy of changes. This thesis proposes Piecewise Linear-ARMA Recursive based on Wavelet Decomposition and Reconstruction model (PL-ARMARWDR). After wavelet decomposition and transformation, the low-frequency approximation sequence will be smoother, and its trend will be more obvious, and the high-frequency detail sequence will be more stable. In the model, the low-frequency approximation sequence uses linear recursive model and high-frequency detail sequence uses ARM A model. Since track status change is not a linear trend, in fact, the trend is nonlinear. Studies have shown that it more in line with exponential trends of track irregularity changes. Because the gray model is an exponential function approximation model, the thesis proposes Piecewise Gray-ARMA Recursive based on Wavelet Decomposition and Reconstruction model (PG-ARMARWDR). As an important non-linear modeling method, neural network model has been widely used. Because neural network model can be approximated by any non-linear process, so this thesis proposes Piecewise ANN-ARMA Recursive based on Wavelet Decomposition and Reconstruction model (PANN-ARMARWDR). In the model, the low-frequency approximation sequences uses recursive neural network model, and high-frequency detail sequence uses ARM A model. After the decomposition-modeling-recon structuring process, the proposed models have realized accurate prediction on trends of track irregularity state changes.In the final part of this doctoral dissertation, the accuracy of four prediction models are analyzed by examples, and MSE and MAPE are used as forecasting accuracy index and the results show that all models belong to high-precision prediction, and the proposed models have achieved satisfactory results. Also, the applicability of all models are also compared and analyzed.