Scholarlink

<p>In the era of Industry 4.0, critical industrial equipment frequently operates under complex, non-stationary conditions, where identifying novel failure modes remains a formidable challenge. While Large Language Models (LLMs) demonstrate exceptional cognitive capabilities, their deployment in safety–critical industrial fault diagnosis is fundamentally hindered by the closed-set data assumption, prohibitive cloud-inference latency, and strict data privacy constraints. To break the intelligence-resource trade-off, this work proposes EC-Distill-ZeroDiag, a hierarchical edge-cloud collaborative framework tailored for description-based zero-shot fault diagnosis within a tiered edge computing architecture. Specifically, a lightweight Edge Student Agent, physically deployed on an industrial edge workstation, utilizes a patch-based convolutional tokenizer to continuously process raw vibration signals. Empowered by a self-reflection mechanism, it resolves routine tasks locally while autonomously offloading highly uncertain patterns to the cloud. At the cloud level, a teacher agent leveraging the Retrieval-Augmented Generation (RAG) paradigm grounds its analysis in expert maintenance manuals to deduce unseen fault physics without hallucination. Crucially, a Chain-of-Thought (CoT) distillation strategy is designed to transfer not merely the final predictions, but the complete deductive reasoning logic from the teacher to the student. By aligning continuous sensor patches with discrete textual rationales, the edge model progressively internalizes the ability to diagnose entirely unseen signal patterns based purely on semantic descriptions. Extensive experiments across two distinct industrial datasets demonstrate that the proposed framework significantly outperforms state-of-the-art baselines, achieving over 75% accuracy in strict zero-shot scenarios. Furthermore, validated by a novel 6-dimensional LLM-as-a-Judge metric, the edge student successfully retains 95.8% of the teacher’s reasoning quality while delivering a real-time amortized latency of 68&#xa0;ms, proving its exceptional potential for reliable, high-throughput deployment under data-scarce and resource-constrained conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EC-Distill-ZeroDiag: a cloud-edge collaborative framework for zero-shot industrial fault diagnosis via large language model distillation

  • 期刊论文

  • 发表日期：

    2026年6月14日

查看全文
  • Zhiwen Yao,
  • Yuhui Xu,
  • Tangbin Xia,
  • Zhen Chen,
  • Ershun Pan,
  • Lifeng Xi

摘要

In the era of Industry 4.0, critical industrial equipment frequently operates under complex, non-stationary conditions, where identifying novel failure modes remains a formidable challenge. While Large Language Models (LLMs) demonstrate exceptional cognitive capabilities, their deployment in safety–critical industrial fault diagnosis is fundamentally hindered by the closed-set data assumption, prohibitive cloud-inference latency, and strict data privacy constraints. To break the intelligence-resource trade-off, this work proposes EC-Distill-ZeroDiag, a hierarchical edge-cloud collaborative framework tailored for description-based zero-shot fault diagnosis within a tiered edge computing architecture. Specifically, a lightweight Edge Student Agent, physically deployed on an industrial edge workstation, utilizes a patch-based convolutional tokenizer to continuously process raw vibration signals. Empowered by a self-reflection mechanism, it resolves routine tasks locally while autonomously offloading highly uncertain patterns to the cloud. At the cloud level, a teacher agent leveraging the Retrieval-Augmented Generation (RAG) paradigm grounds its analysis in expert maintenance manuals to deduce unseen fault physics without hallucination. Crucially, a Chain-of-Thought (CoT) distillation strategy is designed to transfer not merely the final predictions, but the complete deductive reasoning logic from the teacher to the student. By aligning continuous sensor patches with discrete textual rationales, the edge model progressively internalizes the ability to diagnose entirely unseen signal patterns based purely on semantic descriptions. Extensive experiments across two distinct industrial datasets demonstrate that the proposed framework significantly outperforms state-of-the-art baselines, achieving over 75% accuracy in strict zero-shot scenarios. Furthermore, validated by a novel 6-dimensional LLM-as-a-Judge metric, the edge student successfully retains 95.8% of the teacher’s reasoning quality while delivering a real-time amortized latency of 68 ms, proving its exceptional potential for reliable, high-throughput deployment under data-scarce and resource-constrained conditions.