A Fault Feature Learning Model for Zero-shot Intelligent Diagnosis of Compound Faults in Rotating Machinery

Due to the intricacy of rotating machinery and the interconnections among its systems, diagnosing compound faults in equipment poses a major challenge in this research area. Existing intelligent compound fault diagnosis methods rely on having sufficient compound fault data for training. However, obtaining complete compound fault data in industrial scenarios is an arduous task, and compound faults are often considered as unknown faults. This paper introduces a zero-shot intelligent diagnosis model that focuses on discriminating unknown compound faults. The method comprises four stages: data acquisition and preparation, fault features learning, fault semantics construction, and feature semantics mapping. Our primary focus is on the feature learning stage. The fault feature learning model of the zero-shot intelligent diagnosis method (FFL-ZID) incorporates a feature-level contraction block and a decision-level contraction block. The former filters valid information to extract essential features from the raw data, while the latter promotes feature decoupling to ensure independence among selected features and reduce redundancy. Experimental results on the mechanical comprehensive diagnostic simulation platform (MCDSP) demonstrate the effectiveness of the proposed method for identifying unknown compound faults in rotating machinery.