In large complex systems, the system may encounter fault categories for which no training samples are available. This article introduces a large language model (LLM)-guided framework for zero-shot fault diagnosis that leverages the reasoning and representation capabilities of LLM. First, structured textual attribute descriptions are defined for faults, and a reasoning chain mechanism converts numerical operational data into text using the DeepSeek-70B model. Both attributes and textualized data are then embedded via the FlagEmbedding model. In addition, a shared representation space for samples and attributes is then established. System complexity is quantified by compressing the textualized operational data, and the complexity prior, combined with an adaptive fuzzy-boundary adjustment strategy, guides each fault class to form a prototype-centered distribution in the latent space, achieving fine-grained semantic alignment. Finally, an unseen fault classifier is trained solely on attribute information. Experimental results on the ground source heat pump system and Chiller system demonstrate that the proposed method achieves near-zero error in zero-shot fault diagnosis, highlighting the potential of LLM in intelligent fault diagnosis of complex systems.

The few-shot fault diagnosis methods have achieved competitive results in complex systems, but existing approaches fail to achieve human-like one-shot diagnostic learning. This work uses the large language model (LLM) to enhance fault diagnosis with human-like reasoning. First, we transform sensor data into language-based representations for more detailed feature descriptions. Then, we introduce prior knowledge into the language-based diagnostic data through the inference process of the LLM to achieve information gain. Furthermore, we apply arithmetic coding to encode predictive probabilities, forming an entropy-based compressor approximating the information distance derived from Kolmogorov complexity. We achieve the one-shot fault diagnosis by using information distance as the learning metric for human-like few-shot learning. Experimental results on chiller systems show that our method, empowered by the LLM, enables zero-error human-like fault diagnosis. The proposed method holds promise for significantly enhancing intelligent fault perception and maintenance capabilities in complex industrial systems, while also offering a novel research paradigm for the application of LLMs in industrial fault diagnosis.

Federated continual learning (FCL) tackles scenarios of learning from continuously emerging task data across distributed clients, where the key challenge lies in addressing both temporal forgetting over time and spatial forgetting simultaneously. Recently, prompt-based FCL methods have shown advanced performance through task-wise prompt communication. In this study, we underscore that the existing prompt-based FCL methods are prone to class-wise knowledge coherence between prompts across clients. The class-wise knowledge coherence includes two aspects: (1) intra-class distribution gap across clients, which degrades the learned semantics across prompts, (2) inter-prompt class-wise relevance, which highlights cross-class knowledge confusion. During prompt communication, insufficient class-wise coherence exacerbates knowledge conflicts among new prompts and induces interference with old prompts, intensifying both spatial and temporal forgetting. To address these issues, we propose a novel Class-aware Client Knowledge Interaction (C$^2$Prompt) method that explicitly enhances class-wise knowledge coherence during prompt communication. Specifically, a local class distribution compensation mechanism (LCDC) is introduced to reduce intra-class distribution disparities across clients, thereby reinforcing intra-class knowledge consistency. Additionally, a class-aware prompt aggregation scheme (CPA) is designed to alleviate inter-class knowledge confusion by selectively strengthening class-relevant knowledge aggregation. Extensive experiments on multiple FCL benchmarks demonstrate that C$^2$Prompt achieves state-of-the-art performance. Our code will be released.

In the real chiller fault diagnosis tasks, the training data of target faults is commonly unavailable. This scenario restricts the diagnostic performance of data-driven methods. The existing zero-shot learning-based diagnostic models have conquered this challenge by introducing sparse fault semantic attributes. However, the sparse semantic representation misses the interconnections between attributes, which causes difficulty in representing holistically the latent associations between categories. In this article, an innovative zero-shot fault diagnosis (ZSFD) framework for chillers based on sentence-level text attributes is proposed to overcome this difficulty. The defined fault attributes contain two textual labels for the cause and consequence. Subsequently, all text attributes for each category are combined into a sentence (in text form), which is named sentence-level text attributes. Afterward, a semantic representation model, namely, sentence-level bidirectional encoder representations from transformer (SBERT), is employed to project sentence-level text attributes as the dense semantic attributes. Later, a cross-modal contrastive embedding (CMCE) model is established to embed semantic attributes and collected sensor data into a common latent space, where the distributions of attributes and data are aligned. Meanwhile, a hybrid reconstruction classification strategy is designed in CMCE to fully integrate the intrinsic characteristics of categories from both modalities. Finally, the unseen semantic attributes embedded by the CMCE are employed for training zero-shot learning classifier. Extensive experiments are designed and executed on the chiller dataset. The results demonstrate that the proposed framework can improve the distribution consistency between data and attributes compared with the ZSFD methods based on sparse representation, and achieves satisfactory diagnostic performance.