Multi-source domain adaptation (MSDA) has demonstrated superior performance in intelligent fault diagnosis (IFD) compared to single-source domain adaptation (SSDA), as it can provide more comprehensive and diverse information from multiple fully-labeled source domains. However, in many real industrial scenarios, acquiring multiple fully-labeled source domains is challenging because labeling all the source domains is as expensive and laborious as labeling the target domain. Given this concern, a cross-supervised multisource prototypical network (CSMPN) is proposed for multi-source few-shot fault diagnosis. Specifically, a domain-shared and a domain-individual branch are constructed to realize shared domain alignment across all the source and target domains and individual domain alignment of source-target domain pairs, respectively. Within two branches, domain alignment is realized by the designed prototypical contrastive learning (PCL) module. In the PCL module, we propose a prototype calibration strategy to address the issue of biased prototype estimation owing to outlier samples. In addition, a two-stage pseudo-labeled sample selection mechanism is proposed to enhance the feature representation ability of two branches. At the end of the two branches, we design a cross-supervised learning (CSL) module to realize mutual and collaborative learning between the two branches, which can further improve the diagnosis performance on the target domain. Experiments on two different bearing datasets are implemented to verify the superiority of the proposed method compared with the comparison methods. Our code is available at https://github.com/YNWA-Zhang/CSMPN .

The development of convolutional neural networks (CNN) has led to revolutionary progress in the resolution of the offline handwritten Chinese character recognition (HCCR) problem. As the recognition rate on a standard offline HCCR testbed is outstanding, a few samples that remain misclassified have kindled our interest. In this paper, with the help of human recognition results, we present a comprehensive analysis of the samples misclassified by a state-of-the-art CNN model. We performed the analysis based on the top-1-votes, which are obtained from the statistical analysis of human recognition results, and derived the following conclusions: (1) the majority of samples with high top-1-votes were mis-labeled. Besides, by comparing the results of human recognition with that of CNN, some limitations of CNN that provide scope for further improvement are presented; (2) in the samples with medium top- 1-votes, it is shown that the samples with different confidence level have different characteristics. Specifically, some samples could be regarded as multi-label samples; (3) the samples with low top-1- votes are either wrongly written or written extensively in cursive style, which are difficult to match their given ground-truths; (4)the relationship between writing styles and misclassifications are also introduced in the paper. We believe this work should provide some insights and brings new clues on designing new classification methods to deal with these challenging samples.