考虑多种贝叶斯神经网络分布形式组合的设备剩余寿命预测方法

On RUL Prediction Method Based on Bayesian Neural Network with Different Distribution Combinations

基于dropout NN(dropout Neural Network)的设备剩余寿命(RUL)预测方法因使用具有固定分布形式的先验与近似后验的贝叶斯神经网络(BNN)导致预测精度较低。为解决该问题,提出基于高斯近似后验BNN的RUL预测方法和基于混合高斯-伯努利网络的设备RUL预测方法,前者引入混合高斯分布作为先验,通过对参数梯度进行无偏蒙特卡罗估计以优化BNN,后者引入一种离散化的高斯先验分布以正确地定义KL散度,进而可以优化BNN。在PHM 2012轴承数据集上的验证结果表明所构建的混合高斯-高斯网络效果好于dropout NN,证明了改变分布组合可以获得更好的预测效果。

The equipment Remaining Useful Lifetime(RUL) prediction method based on dropout Neural Network(dropout NN) has low precision since it uses a Bayesian Neural Network(BNN) with fixed priori distribution and approximate posterior distribution.To solve the problem, we proposed a RUL prediction method based on BNN with Gaussian approximation posterior distribution and a RUL prediction method based on mixed Gaussian-Bernoulli network.The former introduces the mixed Gaussian distribution as priori distribution and then optimizes BNN by unbiased Monte Carlo estimation of parameter gradient, while the latter introduces a discretized Gaussian prior distribution to define KL divergence correctly, and then optimizes the BNN.The verification results on PHM 2012 bearing dataset show that the mixed Gaussian-Gaussian network has better effect than dropout NN,which proves that BNN with the changed distribution combination can obtain better prediction effect.