Extrapolative sample generation: A novel framework for imbalanced multi-source fault diagnosis

Multi-source data provide rich information for bearing fault diagnosis, yet sample imbalance between critical fault classes and healthy conditions severely restricts model generalization. To address this problem, this study proposed a multi-source fault diagnosis framework based on extrapolation sample generation and an attention-enhanced multivariate information bottleneck. First, a fault-type node structure was constructed according to fault types, and multi-source data were organized by fault category. Second, a novel data augmentation strategy, termed extrapolation sample induced by mixup, was developed. This strategy divided the same-type multi-source data into subdomains and generated interpolation domains through Dirichlet mixing. A reverse mixup mechanism combined with adaptive instance normalization style transfer was then introduced to generate plausible out-of-distribution samples, thereby expanding the fault distribution and reducing the effect of class imbalance. Third, a dedicated convolutional neural network extractor and projection layer were designed for the augmented multi-source data to map heterogeneous features. An attention-enhanced multivariate information bottleneck model was then constructed to map heterogeneous features into a shared representation space, dynamically learn cross-source correlations and optimize feature representation according to information bottleneck criteria. ... mehrFinally, experiments on bearing datasets under imbalance ratios from 2:1 to 5:1 demonstrated that the proposed method achieves 97–99 % accuracy, demonstrating its potential of imbalanced fault diagnosis in engineering applications.