基于MsTCN-Transformer模型的轴承剩余使用寿命预测研究

Research on bearing remaining useful life prediction based onan MsTCN-Transformer model

剩余使用寿命(remaining useful life, RUL)预测是PHM的核心问题之一,复杂的运行工况往往导致设备部件经历不同的故障退化过程,给RUL准确预测带来了巨大挑战。为此,提出了一种多尺度时间卷积网络(multi-scale temporal convolutional network, MsTCN)与Transformer(MsTCN-Transformer)融合模型用于变工况下滚动轴承RUL预测。该方法设计了一种新的多尺度膨胀因果卷积单元(multi-scale dilated causal convolution unit, MsDCCU),能够自适应地挖掘滚动轴承全寿命信号中固有的时序特征信息;然后构建了基于自注意力机制的Transformer网络模型,在克服预测序列记忆力退化的基础上,准确学习时序特征与轴承RUL之间的映射关系。此外,通过对轴承不同故障退化阶段所提取的时序特征可视化分析,验证了所提方法在变工况下所提取的时序特征泛化性较好。多种工况条件下滚动轴承RUL预测试验表明,所提方法能够较为准确地实现变工况下轴承的RUL预测,相比当前多种方法RUL预测结果准确性更高。

Remaining useful life(RUL)prediction is one of the key issues to be solved in prognostics and health management(PHM).Equipment components suffer from different failure degradation processes due to the influence of complex operating conditions,which poses great challenge for accurate RUL prediction.a novel joint mode of multi-scale temporal convolutional network(MsTCN)and Transformer network(MsTCN-Transformer)was proposed for RUL prediction of rolling bearings under variable operating conditions.The method presents a new multi-scale dilated causal convolution unit(MsDCCU)that can adaptively mine the time-sequence feature information inherent in the whole-life signals of bearings.Then a Transformer network based on the self-attentive mechanism was constructed to accurately learn the mapping relationship between the time-sequence features and the bearing RUL by overcoming the memory degradation of the prediction sequence.In addition,the visual analysis for different stages of bearing failure degradation verifies that the proposed method has a better generalization performance of the extracted time-sequence features.Experiments show that the proposed method can achieve more accurate RUL prediction,and the accuracy of RUL prediction is higher compared with other current related methods.