基于角域重采样的宽度迁移学习算法

Broad transfer learning algorithm based on angular domain resampling

针对在变工况中采集到的信号通常为非平稳信号,并且采集到的振动数据分布不一致等情况,提出了基于角域重采样的宽度迁移学习(ADR-BTL)算法,用于轴承故障诊断。首先将非平稳的时域振动信号转换为角域平稳信号,将平稳化处理后的信号进行特征提取,构建多域特征数据集,然后将不同工况下的源域数据和目标域数据通过平衡分布自适应(BDA)方法进行领域适配来减小域间的分布差异,最后构建宽度迁移学习模型。实验首先验证了角域重采样方法可以将振动信号进行平稳化处理,然后通过仿真样本分析得出BDA方法能够解决数据分布不一致问题,最后通过实验结果得出,提出的ADR-BTL识别率达到了98.9%,识别效果是最好的,证明了所提的方法在轴承故障诊断方面是有效的。

An broad transfer learning algorithm based on angular domain resampling(ADR-BTL)is proposed for bearing fault diagnosis in response to the fact that the signals collected in variable operating conditions are usually non-stationary signals and the distribution of the collected vibration data is inconsistent.Firstly,the non-stationary time domain vibration signal is converted into an angular domain stationary signal,and the smoothed signal is feature extracted to construct a multi domain feature dataset.Then,the source domain data and target domain data under different operating conditions are domain adapted using the balanced distribution adaptation(BDA)method to reduce the distribution differences between domains.Finally,a broad transfer learning model is constructed.The experiment first verified that the angular domain resampling method can smooth the vibration signal,and then through simulation sample analysis,it was found that the BDA method can solve the problem of inconsistent data distribution.Finally,the experimental results showed that the proposed ADR-BTL recognition rate reached 98.9%,and the recognition effect was the best,proving that the proposed method is effective in bearing fault diagnosis.