Research on fault degree recognition technique based on edge distribution prediction and network fusion

To address the poor identifiability of weak fault features in rotating machinery under complex working conditions and the low accuracy of few-shot fault severity identification, this paper proposes a fault identification algorithm combining edge distribution prediction and a hybrid deep network. The method integrates fast spectral kurtosis transform and an improved edge distribution prediction to enhance weak fault features and recover missing information in vibration signals. A dual constraint mechanism with feature matching loss and perceptual loss is designed to alleviate feature shift and refine fine-grained feature extraction. A hybrid CNN&Bi-LSTM structure with an information connection layer is built to deeply fuse spatial and temporal features, strengthening the boundedness of fault feature spatial edges and boosting the model's attention to fault-related information. Experiments on a self-built mechanical pump dataset and the CWRU bearing dataset show that the proposed method achieves 96.78% fault identification accuracy and outperforms mainstream approaches in few-shot scenarios. With strong generalization and interpretability, it provides a new technical solution for accurate few-shot fault severity identification of rotating machinery.