Cross-domain fault diagnosis leverages knowledge from multiple source domains to improve diagnostic accuracy in target domain. However, existing methods align source and target domains either jointly or independently, often neglecting the distributional discrepancies among source domains, which hinders effective knowledge transfer. To address this issue, we proposes a hierarchical task-building-based multisource adaptive meta transfer learning (HTB-MSAMTL) framework. First, the semantic alignment bidirectional embedding module pretrains meta-learning parameters through label-feature embedding alignment. Second, a task split strategy is designed for hierarchical meta-tasks, creating multiple domain pairs for learning long-term embeddable high-level meta-knowledge. Finally, a prototype feature reprojection network is developed to optimize calibrated prototypes and minimizes cross-domain distribution discrepancies through a differentiable closed-form solver. Moreover, learnable matrices replace fixed prototypes to enable adaptively tuned prototype representations. HTB-MSAMTL is evaluated on Tennessee Eastman Process, Case Western Reserve University, and Aluminum Electrolysis Process datasets. Results demonstrate HTB-MSAMTL outperforms existing methods in multisource cross-domain fault diagnosis under scarce labeled data scenarios.

Shape shift refers to the phenomenon where by amplitude and scaling of samples of the same fault vary with the environment. How to incrementally diagnose industrial faults under shape shift samples with directly interpretability is a challenging problem. Therefore, a new convertible shapelet learning with incremental learning capability method is proposed in this study. First, incrementally learning under shape shift samples is formulated as how to learn new scaling shift and amplitude shift samples in industrial process. A shape shift detection method is proposed to detect new shape domain. Second, a new parameter initialization method for shapelet learning is proposed, which employs pretraining technology to obtain better initial shapelets. Third, the convertible shapelet-transformed representation is proposed to enhance informativeness of shapelets. An incremental convertible shapelet learning loss is proposed to alleviate catastrophic forgetting in incrementally training new shape time series. By comparing with other state-of-the-art methods, experiments on the Tennessee Eastman Process and real-world aluminum electrolysis process demonstrate the superior performance of the proposed method in accuracy, alleviating catastrophic forgetting and interpretability.

Early classification predicts the class of the incoming sequences before it is completely observed. How to quickly classify streaming time series without losing interpretability through early classification method is a challenging problem. A novel memory shapelet learning framework for early classification is proposed in this article. First, a memory distance matrix is introduced to store the historical characteristics of streaming time series, which can alleviate repetitive calculations caused by the growing length of time series. Second, early interpretable shapelets are extracted in the proposed method by optimizing both accuracy objective and earliness objective simultaneously. The proposed method employs end-to-end learning, which allows the model to directly learn early shapelets without the necessity of searching for numerous candidate shapelets. Third, an objective function of memory shapelet learning is proposed by overall considering accuracy and earliness, which can be optimized by gradient descent algorithm. Finally, experiments are conducted on benchmark dataset UCR, Tennessee Eastman process, and real-world aluminum electrolysis process in China. Comparable results with other state-of-the-art methods demonstrate the superior performance of the proposed method in interpretability, accuracy, earliness, and time complexity.

Few-shot new faults are constantly emerging due to the dynamic environments and operations in the industrial process. It is a challenge for existing fault diagnosis methods to diagnose few-shot new faults without forgetting old faults by fine tuning the base model. This article defines this challenge as few-shot fault incremental learning problem, and proposes a multiview shapelet prototypical network to solve this problem. First, a multiview metalearning framework that simulates true incremental tasks and combines multiview information is proposed in this article to build a generalizable feature space for unseen classes. Second, a multiview shapelet prototypical classifier is proposed to enhance the generalization ability of shapelets in adapting new faults with few samples. Third, multiview metacalibration modules based on transformers are proposed to fuse multiview information and calibrate prototypes and embedded features into a distinguishable space. Finally, experiments are conducted on the benchmark Tennessee Eastman process and the real-world aluminum electrolysis process. Experimental results illustrate that the proposed method is better than the existing methods in terms of interpretability, alleviating catastrophic forgetting, and reducing time complexity.