Exemplar-free class-incremental fault diagnosis via adversarial drift prediction and directionally diverse boundary repulsion

Deep learning-based fault diagnosis methods have achieved remarkable performance under the static closed-set assumption. However, in industrial scenarios, new fault classes constantly emerge, requiring models to incrementally expand their diagnostic capability without forgetting old ones. Although existing rehearsal-based methods mitigate catastrophic forgetting, retaining historical samples is infeasible under the constraints of data privacy and limited on-device storage. Meanwhile, feature drift and overlap also arise: 1) Feature extractor updating causes old prototypes to deviate from their true distributions; 2) New-class features gradually invade the old-class feature space. To address these issues, this paper proposes an exemplar-free class-incremental learning (EFCIL)-based fault diagnosis method integrating adversarial drift prediction and directionally diverse boundary repulsion. During incremental training, the old feature extractor generates adversarial samples through optimizing new-class features toward old prototypes. The feature differences between the old and new extractors are calculated to dynamically train a drift predictor for calibrating old prototypes and boundary features. Meanwhile, a directionally clustered strategy selects boundary features from the margins of each directional cluster as repulsion anchors, and a repulsion loss pushes new-class features away from old-class regions. Experiments demonstrate the proposed method effectively mitigates catastrophic forgetting across all tasks, achieving superior diagnosis performance compared with other methods.