Deep Transfer Learning Based Fault Diagnosis for Electromagnetic Pulse Valve Faults Under Small Sample

• Proposes a deep transfer learning method for electromagnetic pulse valve fault diagnosis under small sample conditions, using simulated data to augment real data. • Identifies an optimal parameter transfer strategy that fine-tunes both feature extractor and classifier, combined with an attention mechanism to enhance multi-sensor feature integration. • Achieves high transfer accuracies of 93.5% and 94.2% with only 7.9% and 8.2% of the total dataset as small sample data, respectively. • Addresses the practical challenge of scarce fault data in industrial settings, offering a cost-effective and reliable diagnostic solution.