A dual-objective contrastive learning approach with dynamic self-adaption for zero-shot fault diagnosis

Fault type classification and fault severity identification are two critical and complementary tasks in fault diagnosis of industrial machines, providing essential information for the maintenance and safety of the machines. However, variable operating conditions in industrial settings make it hard to collect comprehensive fault data covering all possible types and severities, thereby limiting diagnostic efficiency. To overcome these challenges, a novel multi-task network approach is proposed to detect fault type and severity simultaneously even with zero novel samples. Discriminative features are extracted through a contrastive network with task-specific projection heads, enabling the capture of distinct representations for fault type and severity. Two zero-shot mapping spaces are constructed to diagnose fault types and severity by aligning feature representations with the semantic information of fault types and severity. A dynamic self-adaptation optimization mechanism is introduced considering the dependency of fault severity on fault types. It enhances the identification of fault severity. The proposed method was evaluated on two bearing datasets. It achieved up to 89.4 % accuracy for fault type and 83.42 % for fault severity under zero-shot settings, outperforming baselines and demonstrating strong real-world applicability.