Multisource Domain Separation Network for Industrial Intelligent Monitoring in Unseen Conditions
针对工业时间序列预测中工况变化导致分布偏移的问题，提出多源域分离网络，通过分离域私有与共享特征、保留标签变化趋势的域不变信息，并引入方向性风险外推方法，提升模型在未知工况下的鲁棒性和泛化能力。
Industrial time-series prediction is one of the crucial tasks of industrial artificial intelligence, which has been extensively adopted in intelligent monitoring. However, continuous variations in the working conditions of industrial processes lead to continuous changing in monitoring data, resulting in distribution shift. Industrial time-series prediction models may not be able to maintain satisfactory accuracy. Domain generalization (DG) methods can effectively improve the robustness and generalization of the models for unseen or dynamic distribution data. However, existing methods have drawbacks limit their applicability in practice. First, most DG methods are designed for classification tasks and lack a general framework for regression tasks. Second, extracting domain-invariance may obscure task-relevant information, thereby degrading the prediction precision. Third, DG methods are generally sensitive to extremely distributed data, potentially compromising the robustness. To address these issues, a multisource domain separation network (MS-DSN) is proposed. First, by constructing a DSN to separate the features into domain-private and -shared spaces, domain-specific information can be filtered while domain-invariant features are preserved. Second, a supervised contrastive loss is defined to preserve the domain invariant information related to the label changing tendency in the domain-shared space. Finally, a directional risk extrapolation (DREx) method is proposed to represents the extreme distributed data. Experiments on two datasets, CMAPSS and N-CMAPSS, verify the effectiveness of our method.