Intelligent fault detection of zero-sample rotating machinery with embedded physical knowledge of vibration envelope and time-frequency analysis

Data-driven fault detection methods for rotating machinery have achieved impressive performance. Nevertheless, their practical deployment faces substantial challenges, including the high cost of acquiring fault data and inherent difficulties in constructing accurate models. This paper integrates domain knowledge of vibration signal analysis and proposes a physical knowledge-driven modeling method for rotating machinery fault detection with zero fault sample. First, Hilbert envelope analysis is introduced to attenuate the impact of fundamental frequency components. Subsequently, multi-dimensional evaluation metrics are used to select and filter multiple time-frequency analysis methods, thereby constructing a robust time-frequency knowledgebase. Then, three novel loss function driven by zero-fault samples is designed based on the differences between the selected time-frequency analysis methods and physical knowledge regarding the similarity among sliding window samples in monitoring signals. Finally, an end-to-end intelligent fault detection algorithm is developed based on the trained feature encoder and the introduced physical knowledge. The effectiveness of the proposed method is validated on both the rolling bearing experimental platform and the turbine experimental platform. The validation results demonstrate that the proposed method can achieve intelligent fault detection modelling without any fault samples, attaining fault detection test accuracies of 98.97 % and 96.19 % in the two respective case studies.