基于相关分析和组合神经网络的退化预测

产品剩余寿命预测是加速退化试验和故障预测与健康管理两大热点领域中的关键技术之一.为了解决复杂退化的预测问题,提出了一种新型预测方法,对退化轨迹能够实现较长距离的预测.此方法首先对复杂退化数据进行小波变换,通过Durbin-Watson方法和偏相关图分析各级分解序列的自相关性,最后根据序列的特点,组合BP(Back Propagation)和小波神经网络对退化轨迹进行预测.为了验证所提组合神经网络方法的有效性,采用小波神经网络的预测结果进行对比分析.实际退化数据的预测结果表明,所提方法比单独采用小波神经网络,具有更小的均方差(MSE,Mean Square Error),对剩余寿命(RUL,Remaining Useful Life)也具有更高的预测精度.

Degradation prediction based on correlation analysis and assembled neural network

Efficient prognosis for remaining useful life of product is critical for both accelerated degradation testing and prognostics and health management, which are two hot points in recent years. A novel degradation prediction model was proposed to improve the long prediction capability for complex degradation path. Durbin-Watson method and partial correlation graph were utilized to analyze the decomposition results of wavelet transformation. Then, according to the characters of series, back propagation(BP) and wavelet neural network were assembled to predict degradation path. To verify the proposed method, wavelet neural network was selected as comparison. A practical degradation result demonstrates that this model can offer smaller mean square error (MSE) and higher prediction accuracy for remaining useful life (RUL).