Fault diagnosis of industrial robot joints based on generative domain adaptation

The data-driven diagnosis of faults in industrial robots has already yielded promising results. However, the fault scenarios observed in production environments are frequently more intricate, and some unidentified faults cannot be replicated in a laboratory setting. In order to address these limitations, a novel generative domain adaptive network-based model is proposed as an alternative to the traditional data-driven model. This model aims to overcome the challenges posed by unbalanced small-sample fault diagnosis. The initial stage of the data processing involves the application of a wavelet transform, which serves to filter out noise and employ other methods. Subsequently, a sample generation network (GDAN) and a domain adaptive network are constructed to supplement the scarce samples and align the feature distribution of the data between domains. Finally, the Mixup method is introduced to enrich the spatial features of a few samples in the target domain and improve the diversity of samples in the source domain. The experimental results demonstrate that the diagnostic accuracy and generalizability of GDAN are superior to those of the comparison methods when targeting small sample data with unknown defect types.