Logo image
Back
Uncertainty-aware continual learning for class-added multimodal industrial diagnosis with time-series and textual data
Journal article   Peer reviewed

Uncertainty-aware continual learning for class-added multimodal industrial diagnosis with time-series and textual data

Yi Ding, Chao Wang, Hanxiao Zhang and Xingchen Liu
Reliability engineering & system safety, Vol.278(February 2027), 113569
09/28/2026
DOI: 10.1016/j.ress.2026.113569

View Online

Abstract

In practical industrial systems, maintaining reliable operation requires fault diagnosis that not only identifies known fault categories without costly manual annotations, but also recognizes and incrementally learns newly emerging fault types under uncertain conditions. Meanwhile, maintenance logs recorded during operation contain semantic information that complements sensor data and guides fault understanding. Motivated by this, this paper proposes an uncertainty-aware continual learning framework for class-added multimodal industrial diagnosis without manual annotations. First, to bridge the heterogeneity between sensor dynamics and textual semantics in multimodal diagnosis, a Hard-Soft Cross-Modal Semantic Alignment module learns shared representations by combining instance-level pairing with soft relational consistency across modalities. Second, to improve reliable novel-class recognition in open and evolving environments, an Uncertainty-Score-Based Dual-Modal Novel Class Inference module converts modality-specific prototype distances into uncertainty scores and fuses them through a disjunctive probabilistic rule. Third, to enhance the few-shot continual adaptation of newly emerging classes, an Uncertainty-Informed Text-Guided Prototype Rectification mechanism refines prototypes with textual semantic cues, where the rectification strength is adaptively controlled by semantic-transfer uncertainty. A compact replay memory with herding is further adopted to preserve representative historical samples and reduce catastrophic forgetting. Experiments on two datasets validate the proposed framework.
 Continual learning Maintenance logs Multimodal industrial diagnosis Novel class recognition Sensor-text fusion Uncertainty-aware learning

Details

Metrics

1 Record Views
Logo image