A Progressive Network With Multiple Classifiers for Fault Diagnosis of Rotating Machinery Under Limited Sample Conditions

The rapid development of fault diagnosis for rotating machinery has provided critical support to industrial production. However, in industrial environments, the high cost of data acquisition and stringent safety requirements make the issue of limited samples a key challenge that restricts the performance of diagnostic models. Therefore, this article proposes a progressive network with multiple classifiers, which effectively integrates multilevel feature information by progressively introducing early, mid, and final classifiers, thereby improving the utilization efficiency of fault features under limited sample conditions. First, a multiscale interactive feature extraction module is proposed, which captures critical information through multiscale feature extraction and interactive fusion, enhancing the ability of the model to perceive subtle fault features. Furthermore, an interactive masked multihead attention mechanism is proposed, which dynamically adjusts the attention of the model to critical information through multispace feature interaction and adaptive enhancement, thereby improving the representation quality of complex fault features. Finally, a dynamic joint optimization strategy is proposed, which dynamically adjusts the ratio of loss weights among classifiers to ensure that the learning objectives at each stage are appropriately focused, thereby enhancing the ability of the model to learn fault features. The results show that, across multiple scenarios in four datasets, the proposed method achieves average diagnostic accuracies of 99.57%, 99.00%, 99.39%, and 92.14%, respectively, all of which outperform the compared methods.