Research on few-shot fault diagnosis and feature extraction mechanism based on ADPCW-ELCNN

Aiming at weak features masked by noise, disconnected parameter-feature adaptation, and unbalanced generalization-efficiency in rotating machinery few-shot fault diagnosis, an Adaptive Dual-Parameter Collaborative Wavelet Convolutional Neural Network is proposed. Different from existing frameworks relying on data augmentation, transfer learning, or blind structural optimization, it adopts a physics-informed, parameter-collaborative, feature-faithful paradigm. Core innovations include quantitative correlation between wavelet scale and convolution kernel length, global-local adaptive strategy, and cascaded nonlinear enhancement. Experimental results on bearing and gearbox datasets show the model has 9.19 K parameters and 0.02 × 10³ M FLOPs, with 99.98% average accuracy. It maintains F1-score over 94% under 20-sample and - 10 dB SNR conditions, providing an efficient solution for limited-sample fault diagnosis.