基于Bi-TCN-LSTM的滚动轴承剩余使用寿命预测方法

Prediction Method for Remaining Useful Life of Rolling Bearings Based on Bidirectional Temporal Convolutional Network and Long Short-Term Memory Network

由于时间卷积网络(temporal convolutional networks, TCN)感知场不足,轴承的关键退化信息常常被忽略,导致轴承剩余使用寿命(remaining useful life, RUL)预测结果不佳;而长短期记忆网络(long short-term memory, LSTM)随着数据量及序列长度的增加,长期依赖问题仍可能得不到很好解决。因此,课题组提出了一种基于双向时间卷积网络和长短期记忆(Bi-TCN-LSTM)的滚动轴承寿命预测方法。首先对多传感器数据进行归一化并做融合处理,然后采用Bi-TCN-LSTM进行数据特征提取与深度学习,其中对TCN模块引入卷积注意力机制(convolutional attention module, CAM),将LSTM的3个门简化为1个门,有效加快了预测模型学习的速度并提高了预测模型的精确度;采用IEEE PHM 2012轴承数据集作为实验数据集,进行了RUL预测实验。结果表明:与其他先进的预测模型相比,Bi-TCN-LSTM方法预测结果的误差相对较低,预测性能较好。

Due to the insufficient sensing field of the temporal convolutional networks(TCN),the key degradation information of the bearing is often ignored,which results in poor prediction of the remaitning useful life(RUL)of bearings.Moreover,the long-term dependence problem of long short-term memory(LSTM)may not be well solved with the increase of data volume and sequence length.Therefore a new prediction method based on Bidirectional temporal convolutional network andLong short-term memory(Bi-TCN-LSTM)was proposed.Firstly,the multi-sensor data was normalized and fused,and then the Bi-TCN-LSTM was used for data feature extraction and deep learning,in which the convolutional attention mechanism(CAM)was introduced into the TCN module,and the three gates of the LSTM were simplified into one gate.It effectively accelerated the learning speed of the prediction model and improved the accuracy of the prediction model.The IEEE PHM 2012 bearing dataset was used to carry out the RUL prediction experiments.The results show that compared with other advanced prediction models,the Bi-TCN-LSTM method has relatively lower prediction error and better performance.