Robust-MBDL: A Robust Multi-Branch Deep-Learning-Based Model for Remaining Useful Life Prediction of Rotating Machines

Abstract

Accurately predicting the remaining useful life (RUL) of rotating machinery is a cornerstone of proactive industrial maintenance, essential for preventing catastrophic failures, optimizing service schedules, and reducing operational costs. However, industrial equipment operates under highly dynamic conditions, generating complex, multi-channel sensor data heavily corrupted by environmental noise and operational variations. Traditional RUL prediction methods often rely on handcrafted features or single-stream deep learning architectures that struggle to capture the simultaneous spatial-temporal degradation patterns across different sensors. To overcome these challenges, this paper introduces Robust-MBDL, a robust multi-branch deep learning framework engineered specifically for complex machinery prognostics. The core architecture utilizes a parallel, multi-branch design where distinct network streams are dedicated to extracting heterogeneous feature representations—such as localized spatial correlations and long-term temporal dependencies—directly from raw sensor signals. By integrating robust optimization strategies and advanced feature fusion mechanisms, the Robust-MBDL model effectively filters out industrial noise and handles shifting operational contexts. This comprehensive approach bridges the gap between raw, multi-sensor data stream processing and high-precision asset degradation tracking, providing a highly reliable and stable solution for modern predictive maintenance systems.