基于贝叶斯优化的SWDAE-LSTM滚动轴承早期故障预测方法研究

Early fault prediction method combining SWDAE and LSTM for rolling bearings based on Bayesian optimization

针对滚动轴承的早期故障特征较弱,在强噪声背景下难以有效提取以致生命周期很难准确预测的问题,提出了一种基于贝叶斯优化(BO)的滑动窗堆叠去噪自编码器(SWDAE)和长短期记忆(LSTM)网络的早期故障预测模型。使用滑动窗算法保留具有非线性特征和时序特征的历史正常数据,输入到模型中进行训练,使模型学习滚动轴承的正常运行状态趋势。将滚动轴承运行的数据输入到训练好的SWDAE-LSTM模型中进行实时在线监控,利用模型的预测值与真实值的残差检测滚动轴承早期故障。针对模型超参数组合选择困难的问题,使用贝叶斯优化算法对模型的超参数进行调优。最后,使用美国辛辛那提大学智能维护中心(IMSCenter)的轴承全生命周期数据以及机械故障综合模拟实验装置获取的数据进行仿真实验验证。结果表明,使用贝叶斯优化算法进行智能调参的模型和基于时域指标的方法对比,可以更早的有效检测出滚动轴承的早期故障并具有很强的鲁棒性。与其余深度学习方法比较,其模型的诊断准确率高于其他方法,进一步证明了其有效性和可靠性。

The early fault characteristics of rolling bearings are usually rather weak and it is difficult to effectively extract under the background of strong noise so that the life cycle is difficult to accurately predict.Aiming at this,an initial fault prediction model based on Bayesian optimization(BO)consisting of sliding window stacked denoising auto encoder(SWDAE)and long short-term memory(LSTM)network was proposed.The sliding window algorithm was used to retain historical normal data with non-linear characteristics and time series characteristics,so,when the processed data were input into the model for training,the model was able to learn the normal running state trend of the rolling bearing.The data of the rolling bearing operation were then input into the trained SWDAE-LSTM model for real-time online monitoring,and the residual of the predicted value and the true value of the model was used to detect the early failure of the rolling bearing.Aiming at the difficulty in selecting the hyperparameter combination of the model,a Bayesian optimization algorithm was used to tune the hyperparameter of the model.Finally,the bearing full life cycle data from the University of Cincinnati Intelligent Maintenance Center(IMSCenter)and the data from a mechanical failure integrated simulation experiment device were used for simulation verification.The results show that the model of intelligent parameter adjustment using Bayesian optimization algorithm and the method based on time domain index can detect the early fault of the rolling bearing effectively and have strong robustness.Compared with other deep learning methods,the diagnostic accuracy of the model is higher than that of other methods,which further proves its validity and reliability.