Anomaly detection of gear wear: introducing a novel health indicator through experimentation and physical investigation

Abstract

Monitoring gear wear is important for gear diagnosis, although it remains a challenging endeavor. Unlike localized tooth faults, there are still physical patterns in the vibration signature associated with distributed wear faults that should be unfolded. This study contributes a novel framework for gear wear diagnosis, supported by physical principles and validated against traditional methods through extensive experimentation. First, we characterize gear wear in the vibration signature, utilizing data from dozens of seeded, realistic degrading wear cases across various rotational speeds collected through unique controlled-degradation tests. We examine conventional features that have proven useful for diagnosing localized tooth faults and demonstrate their inadequate performance in diagnosing gear wear. We propose a sensitive spectral analysis of the gear mesh and sideband spectral energy in the spectrum to recognize wear manifestation and monitor its degradation. Drawing on the insights gained in this study, we introduce a novel health indicator for anomaly detection of incipient gear wear, grounded in physical understanding and employing sophisticated yet simple feature engineering techniques. We analyze the performance of the proposed health indicator, demonstrating its superiority over state-of-the-art methods, and underscore the importance of considering assembly and dismantling operations to avoid the issue of test-training leakage. Additionally, we highlight the potential of the proposed physical-based framework combined with a deep-learning approach to detect wear in its early stages. To the best of our knowledge, this is the first study to conduct extensive controlled-degradation tests, rather than endurance tests, to investigate gear wear and offer a hybrid physical data-driven framework that combines advanced signal processing, feature extraction, and feature engineering techniques for early anomaly detection.