Diagnostics and prognostics system design integrating hybrid Siamese and model-driven learning

Abstract

The article introduces a hybrid fault detection and prognosis system design for motor bearings, which is crucial for the reliability and efficiency of rotating machinery. The focus is on the concern of bearing fault detection with limited data quantity. Traditional approaches struggle with fault detection depending on the quantity and quality of historical data samples, especially in dynamic conditions, often failing to capture long-term dependencies and uncertainties in bearing health states, thereby limiting precise fault detection and remaining useful life predictions. To overcome these challenges, the research proposes a novel hybrid framework that combines data-driven and model-driven learning. This framework employs the Siamese neural network (SNN) containing two identical subnetwork models models for fault detection and Bayesian estimation for prognosis purposes, bridging the gap between data-driven and model-based approaches. SNNs excel in learning from limited labeled data and are suitable for real-time fault detection. Experimental validation with the PRONOSTIA-FEMTO bearings dataset confirms the framework’s superior performance in fault detection, promising improved maintenance practices and equipment reliability in industrial processes.