A novel few-shot meta-learning strategy for fault diagnosis of wastewater treatment process

Challenges in diagnosing faults in wastewater treatment processes (WWTPs) arise from the limited availability of fault samples and the complexity of multidimensional time-series data with low information density. Current deep learning methods predominantly rely on large-scale labeled data, whereas traditional meta-learning approaches often fail to effectively capture heterogeneous temporal patterns in sensor measurements. To address these issues, this paper proposes a novel few-shot fault diagnosis framework that integrates prototypical networks with an enhanced TimesBlock module. First, a Meta-TimesBlock (MTB) is designed to adaptively weight multidimensional time-series features based on XGBoost-driven importance analysis, addressing the limitation of equal treatment of variables in existing methods. Second, a parameter-efficient MetaLearner module is introduced to reduce computational complexity while maintaining feature extraction capability. Third, an auxiliary loss function enhances prototype discrimination by encouraging orthogonality between class representations. Experiments on the BSM1 benchmark under three weather scenarios demonstrate that the proposed method is superior, achieving average accuracies of 94.72%, 98.24%, and 82.75% in 8-way 5-shot tasks. Compared with state-of-the-art methods, the proposed approach achieves performance improvements of 3.39%-8.43%. These results validate the effectiveness of combining domain-specific feature engineering with meta-learning for few-shot fault diagnosis in industrial processes.