Collaborative and trustworthy fault diagnosis for mechanical systems based on probabilistic neural network with decision-level information fusion

Abstract

Fault diagnosis is a critical component of prognostics and health management, enhancing machinery reliability and ensuring operational efficiency by enabling proactive maintenance strategies. However, achieving this requires high data fidelity to accurately predict the full spectrum of faults and structural degradation for reliable assessments. AI-driven fault diagnostics based on machine learning often face challenges in reliability due to uncertainties arising from variations in data distribution, caused by changing operating conditions and noise interference. These factors undermine the trustworthiness of such methods. To address these challenges in accuracy and reliability for mechanical systems, such as gearboxes, this study proposes a Trustworthy Intelligent Diagnostic (TID) model. The TID model incorporates a multi-scale probabilistic neural network, and a decision fusion module based on uncertainty quantification (UQ). Specifically, three UQ-based decision fusion strategies are introduced to enhance diagnostic reliability by effectively managing uncertainty in fault diagnosis. Building upon the TID model, a cooperative fault diagnosis framework is further proposed to facilitate fault knowledge sharing and alleviate the limitations posed by data scarcity. The proposed approach is validated using both experimental data and real-world wind turbine gearbox failure datasets, demonstrating significant improvements in diagnostic accuracy and a notable reduction in false alarm rates. These results highlight the effectiveness, reliability, and superiority of the proposed method.