客货共线无砟轨道平顺状态预测模型

客货共线无砟轨道的轨道质量指数(TQI)具有随时间长期缓慢变化并伴随平稳波动的特点,而现有的预测模型难以预测这种变化。基于小波和时间序列分析预测方法,提出ARMA-BP神经网络和ARMA-SVR预测模型。通过小波分析将TQI时间序列分解为高频和低频2个部分,采用ARMA模型对高频部分建模,分别采用BP神经网络和支持向量回归SVR模型对低频部分建模,最后对高频和低频进行综合预测。此方法可根据具体情况对具有不同特性的TQI时间序列进行针对性建模,提高预测精度。运用此方法对包西线和太中线10个无砟轨道区段TQI时间序列进行预测,结果表明:ARMA-BP神经网络与ARMA-SVR的建模精度平均值分别为98.1%和98.5%,后验差分别为0.31和0.21,均达到1级;前者对已知数据的拟合精度高,而后者对未知数据预测能力较强、泛化能力更突出。

Prediction Model for Rail Regularity State of Ballastless Track on Passenger-Freight Mixed Lines

The track quality index(TQI)of ballastless track on passenger-freight mixed line is characterized by slowly changing for long-term with time and accompanied by stationary fluctuations.However,the existing prediction models are incompetent to predict this variation.Based on wavelet analysis and time series prediction methods,two prediction models named ARMA-BP neural network and ARMA-SVR are put forward.Wavelet analysis algorithm is utilized to decompose the original TQI time series into high and low frequency components.Then ARMA model is used for high frequency component modeling,BP neural network and support vector regression(SVR)model are respectively used for low frequency component modeling.Finally,comprehensive prediction is made for high and low frequency components.This method can model TQI time series with different characteristics according to specific situations,which can improve prediction accuracy.Using this method,the TQI time series samples of ten ballastless track sections on Baotou-Xi an line and Taiyuan-Zhongwei line are predicted.Results show that the average modeling precisions of ARMA-BP neural network and ARMA-SVR are 98.1%and 98.5%,and the posterior error ratios are 0.31 and 0.21 respectively,both reaching level 1.The former model has a high fitting precision for the known data,and the latter has a strong ability to predict the unknown data,showing more outstanding generalization ability.