基于油液光谱LSSVR-AR模型的发动机故障预测

Engine Fault Prediction Based on Oil Spectrum Data LSSVR-AR Model

针对传统油液光谱数据预测模型精度有限的不足,提出了一种基于最小二乘支持向量回归(LSSVR)与AR模型相结合的非平稳时间序列建模方法(LSSVR-AR),并应用于某型履带车辆发动机油液光谱数据及故障的预测。首先对非平稳时间序列进行最小二乘支持向量回归,得到非平稳时间序列的趋势项及剔除趋势项后的随机项;然后对随机项建立AR模型并与趋势项的LSSVR模型组合,得到非平稳时间序列模型;最后用所建模型对油液光谱数据及发动机故障进行预测。用所提建模方法对Fe、Cu、Pb、Si光谱数据预测的平均绝对百分比误差分别为1.987%、2.889%、2.343%、6.860%,明显低于其他模型。实例证明,所提模型能对发动机故障进行准确预测。

To improve prediction accuracy of traditional oil spectrum models,a LSSVR-AR（least square support vector regression and auto-regression model）based non-stationary time series model is proposed and applied in the oil spectrum data to predict the fault of a specified track vehicle engine.Firstly,least square support vector regression is used to non-stationary time series to abstract the tendency.The stochastic components are obtained after the tendency is eliminated.Secondly,stochastic components are modeled with auto-regression model and then combined with LSSVR model to form the non-stationary time series model.Finally,the constructed model is used to predict oil spectrum data and engine fault.Prediction mean absolute percentage error of Fe,Cu,Pb and Si spectrum data are 1.987%,2.889%,2.343%,6.860%,respectively,and they are obviously lower than the other models.The proposed model has reliable precision in predicting engine fault.