ARSLoss: a new perspective on quantitative risk-guided intelligent fault diagnosis

Existing intelligent fault diagnosis methods focus primarily on improving identification accuracy but neglect fault severity and diagnosis prioritisation – a misalignment with real-world operational needs. Collecting sufficient expert knowledge and maintenance data to assess risk levels is time-consuming, and quantifying/mapping such multidimensional information into a unified framework remains challenging. For machine/deep learning methods, adaptively leveraging this information to enhance high-risk fault detection while maintaining overall accuracy poses another unresolved problem. To address these gaps, this study proposes a quantitative risk-guided intelligent fault diagnosis strategy that aligns model training objectives with practical requirements from an economic perspective. Specifically, we quantify fault severity via multi-tiered economic loss metrics and design a risk-sensitive matrix to characterise diagnosis risks under diverse scenarios. More critically, we propose an Adaptive Risk-Sensitive Loss Function (ARSLoss) that transforms risk levels into differentiable loss values, synergistically optimised via orthogonal regularisation and an attention annealing mechanism to balance multiple loss components. Finally, an Adaptive Risk-Sensitive Graph Feature Extractor (ARSGFE) for mechanical fault diagnosis is developed and trained with ARSLoss. Experimental results demonstrate its practicality: in three case studies, the proposed ARSGFE reduces economic losses by up to 18.32%, validating both the methodological innovation and engineering value of ARSLoss.