A novel hybrid physical AI-based strategy for fault severity estimation in spur gears with zero-shot learning

•Combining physics of failure of machine elements with AI-based approaches for fault severity estimation of tooth breakage.•Comprehensive investigation of the physical phenomena in realistic simulations of the vibration signature.•Proposing a health indicator for fault detection compared to conventional unsupervised-learning anomaly detection algorithms.•Proposing a zero-shot learning strategy for fault severity classification using simulations and estimated transfer functions from experiments.•Performance examination by demonstration on data from an actual testbed. Fault diagnosis of gears by vibration analysis has undergone significant growth in recent years. The traditional approaches for gear diagnostics in the past were focused mainly on fault detection. Improved understanding of physics of gear interaction, together with significant progress in dynamic modelling and the new era of artificial intelligence, make it possible for researchers to take on more challenging tasks such as fault severity estimation (FSE) and gear prognosis. This study presents a novel, hybrid strategy for combining physics of failure of machine elements with AI, specifically FSE of tooth breakage faults in spur gears, by a unique fusion of dynamic modelling, experimental characterization, feature extraction, and unsupervised machine learning algorithms. The novelty of the proposed strategy is its logic flow, which starts with a fundamental and deep understanding of the physical phenomena in realistic simulations. The extracted features are selected based on physical insights. The training dataset includes healthy measured data and simulations of healthy and faulty gear to compensate the lack of faulty data in real case scenarios. Discrepancies between the simulations and reality are reduced by passing the simulated signal through transfer functions estimated from experiments. The severity of the fault is estimated according to a “traffic-light” classification strategy, in which, for the first time, the fault geometry is employed to set thresholds for health indicators (HIs). This study shows that a wide set of sensitive features should be used for an early detection, while a smaller and more basic set of features is sufficient for fault severity classification. The suggested strategy is demonstrated on data from an actual testbed; and it is shown to detect breakage faults and estimate their severity, both by zero-shot learning (that is, without training on any measured faulty i