Semi-Supervised Positive Unlabeled Fault Diagnosis Based on Multimodal iMambaformer Network

With the increasing level of industrial automation, the importance of timely diagnosis of mechanical equipment failures has become increasingly prominent. In recent years, deep learning has shown remarkable performance in the field of fault diagnosis due to its excellent feature learning capabilities. However, the feature extraction from a single modality has limited generalization ability, and the overhead introduced by complex neural networks can be substantial. Additionally, in actual industrial production, fault samples are very limited, which restricts supervised learning guided by a large number of labeled samples. This paper proposes a semi-supervised positive-unlabeled fault diagnosis framework based on a multi-modal iMambaformer network (PMMTN). By integrating multi-modal data from time series and spectrograms, PMMTN effectively extracts spatiotemporal features and reduces computational costs. Furthermore, we introduce an adaptive positive-unlabeled learning strategy tailored for multi-class tasks, overcoming the traditional deep learning models' dependency on large amounts of labeled data and enhancing the model's fault diagnosis capability under conditions of limited labeled samples. Extensive experimental validation demonstrates that PMMTN outperforms existing methods on both public and real-world datasets, particularly exhibiting exceptional robustness and accuracy when handling long sequence data and complex fault patterns.