基于BP神经网络的多应力加速寿命试验预测方法

Life-prediction of Multi-stress Accelerated Life Testing Based on BP Algorithm of Artificial Neural Network

针对多应力加速寿命试验(ALT)中传统的寿命预测方法存在建立加速模型及求解多元似然方程组困难的缺点,基于反向传播(BP)人工神经网络(ANN),利用BP神经网络良好的预测特性,建立了多应力恒定加速寿命试验寿命预测模型。首先,以加速寿命试验中的加速应力水平和通过经验分布得到的可靠度作为网络训练输入向量;以非线性最小二乘法对原始失效数据进行拟合并得到回归方程,利用回归方程生成大量的仿真数据作为训练目标向量;然后,建立3层BP神经网络并对网络进行训练。最后,把正常应力水平和设定的可靠度输入训练好的网络,得到预测的失效时间,进而给出可靠度函数的预测曲线。通过仿真算例对本方法进行验证,预测值和仿真值相比较表明,所建立的网络能反映应力水平、可靠度与寿命的关系,为多应力加速寿命试验的寿命预测提供了新的思路。

To deal with the difficulties of traditional life prediction methods in establishing an accelerated model and solving pluralism likelihood equations, a new model is proposed to predict the life of the items in multistress accelerated life testing（ALT） based on the back propagation （BP） artificial neural network （ANN）. The accelerated stress levels and reliability are used as training input vectors. By the least square fitting, the regression equation of the original data can be obtained, with which a large number of simulation data can be generated as training target vectors. Then a three-layer BP neural network is set up and trained, and failure data can be predicted by putting the normal stress levels and required reliability into the model, and the predicting curves can be drawn. Comparison with the simulation data demonstrates that the model can reflect the relationships among the stress levels, the reliability, and the life of the items. It provides a new route for lifeprediction in multi-stress accelerated life testing.