An Improved Prototypical Network with L2 Prototype Correction for Few-shot Cross-domain Fault Diagnosis

•Coordinate attention (CA)is introduced into the feature extractor, cooperating with the proposed three-channel spectrogram containing positive, negative and the whole spectrograms based on continuous wavelet transform (CWT), which could provide prototypical network better feature maps for classification.•L2 prototype correction, combining l2 normalization and prototype correction, is proposed and introduced in meta-testing, which can not only mitigate the length fluctuations caused by the domain shift, but also make the prototype more accurate, thus improving the generalization performance of the model.•Extensive comparative experiments are conducted on both datasets to demonstrate the effectiveness of the proposed method, especially considering experiments with limited auxiliary samples. Additionally, the construction of the meta-training dataset in the fault diagnosis is also discussed, which treats data from multiple working conditions as auxiliary domain data for the model to learn domain invariant features and avoid overfitting.