Causality-Preserving Domain Generalization via Adaptive Fourier Mixup for RUL Prediction.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Time Series Analysis

Background:

• Domain generalization (DG) faces challenges in time series due to domain shift.
• Strict DG settings lack target domain data during training, hindering model adaptability.

Purpose of the Study:

• To propose AFM-CIR, a unified framework for robust time series domain generalization.
• To address the challenge of domain shift in strict DG settings.

Main Methods:

• Integrated semantic-similarity-guided Adaptive Fourier Mixing (AFM) with Causality-Inspired Regression (CIR).
• Constructed a domain-invariant, order-preserving guidance embedding for adaptive modulation and phase perturbation.
• Employed correlation factorization and adversarial masking within CIR for invariance and causal sufficiency.

Main Results:

• AFM-CIR generated label-consistent and causally coherent augmented samples.
• Theoretical guarantees for phase intervention controllability were established.
• Consistently achieved state-of-the-art performance on four benchmark industrial datasets.

Conclusions:

• AFM-CIR effectively addresses domain generalization challenges in time series.
• The framework outperforms existing ERM, general DG, and task-specific DG baselines.
• Demonstrates the potential of integrating Fourier mixing and causal regression for robust DG.