Aero-Engine Remaining Useful Life Prediction Based on Bi-Discrepancy Network.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Unsupervised domain adaptation (UDA) commonly uses adversarial learning for feature distribution alignment.
• Existing UDA methods often require auxiliary models, increasing computational costs.
• Current approaches overlook fine-grained domain discrepancies, hindering detection of target samples with significant domain shifts.

Purpose of the Study:

• To propose a bi-discrepancy network for effective cross-domain prediction.
• To address computational costs and the oversight of fine-grained domain discrepancies in UDA.
• To enhance the detection and processing of target samples with significant domain shifts.

Main Methods:

• A bi-discrepancy network is developed, incorporating a dual regressor to detect samples with significant domain shifts by maximizing output discrepancy.
• Adversarial training between a feature generator and the dual regressor facilitates global domain adaptation.
• Local Maximum Mean Discrepancy (MMD) is employed for fine-grained feature alignment across different degradation stages.

Main Results:

• The proposed method demonstrated significant improvements on 12 cross-domain prediction tasks on the C-MAPSS dataset.
• Average root-mean-square error (RMSE) reductions of 77.24%, 61.72%, 38.97%, and 3.35% were achieved compared to four mainstream UDA methods.
• The results validate the effectiveness of the bi-discrepancy network in handling domain shifts.

Conclusions:

• The bi-discrepancy network offers an efficient and effective solution for unsupervised domain adaptation.
• The method successfully tackles both global and fine-grained domain discrepancies.
• This approach significantly enhances performance in cross-domain prediction tasks, particularly in scenarios with substantial domain shifts.