Multisource Domain Separation Network for Industrial Intelligent Monitoring in Unseen Conditions.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Industrial Monitoring

Background:

• Industrial time-series prediction is vital for intelligent monitoring but suffers from distribution shifts due to changing operational conditions.
• Existing domain generalization methods often fail in regression tasks or degrade prediction accuracy by obscuring relevant information.
• Current domain generalization techniques can be sensitive to extreme data distributions, impacting model robustness.

Purpose of the Study:

• To develop a robust domain generalization framework for industrial time-series regression tasks.
• To address limitations of existing methods, including lack of regression support, information loss, and sensitivity to extreme data.
• To improve the accuracy and generalization capabilities of industrial monitoring models.

Main Methods:

• A multisource domain separation network (MS-DSN) was proposed to separate features into domain-private and -shared spaces.
• A supervised contrastive loss was defined to retain label-relevant information in the domain-shared space.
• A directional risk extrapolation (DREx) method was employed to handle extremely distributed data.

Main Results:

• The MS-DSN effectively filters domain-specific information while preserving domain-invariant features.
• The proposed method maintains prediction precision by preserving task-relevant information.
• Experiments on CMAPSS and N-CMAPSS datasets demonstrated the effectiveness of the MS-DSN.

Conclusions:

• The MS-DSN offers a robust solution for industrial time-series prediction under distribution shifts.
• The method enhances model generalization and accuracy in dynamic industrial environments.
• The proposed approach overcomes limitations of existing domain generalization techniques for regression tasks.