Abstract: In practical engineering, it is difficult to acquire sufficient available fault samples of gearbox with the same distribution, in addition, the acquired unlabelled samples will inevitably be mixed with some out-of-distribution unknown interference samples, which will bring challenges to the existing research on intelligent fault diagnosis of gearbox. A new method based on an improved semi-supervised prototype network is proposed for cross-domain fault diagnosis of gearbox between different working conditions under out-of-distribution interference samples. First, a label allocation criterion is designed, which can fully exploit the information and assign pseudo-labels of the unlabelled samples to effectively suppress out-of-distribution interference samples. Then, a modified cost function is defined based on label smoothing and metric scaling to fully evaluate the similarity between fault samples, which can exploit the generic characteristics of the meta-learning task, and further improve network generalisation. The proposed method is used to analyse the experimental data of gearbox under different health states, and then different few-shot cross-domain diagnosis scenarios and out-of-distribution distribution samples are set up for comparison and verification. The experimental results show that the proposed method can more effectively achieve the cross-domain fault diagnosis of gearbox under different working conditions with few samples compared with the existing methods.

Key words: improved semi-supervised prototype network, gearbox fault diagnosis, out-of-distribution interference samples, label allocation criterion, modified cost function