Semantic Clustering and LLM‐Based Recommendation for Industrial Fault Diagnosis

OpenAlex İngilizce

ABSTRACT Industrial maintenance systems generate large volumes of fault records that are typically stored as unstructured textual descriptions, making automated analysis and solution recommendation challenging. This study presents an AI‐assisted framework with two complementary components for Turkish industrial maintenance data: semantic clustering of fault descriptions and LoRA‐based generation of corrective‐action suggestions. Fault descriptions are transformed into semantic representations using multiple embedding methods, reduced with Uniform Manifold Approximation and Projection (UMAP), and clustered with HDBSCAN. Under the fixed UMAP‐HDBSCAN configuration reported in this study, Language‐Agnostic Sentence Representations (LASER) obtained the strongest internal clustering results, including a Silhouette score of 0.8529. For recommendation generation, Meta‐LLaMA‐3‐8B, Phi‐4, Gemma‐3n‐E4B, and T5‐base were fine‐tuned with Low‐Rank Adaptation (LoRA) and evaluated using ROUGE metrics. Among the evaluated models, Phi‐4 achieved the highest lexical‐overlap scores (ROUGE‐1 = 0.3320, ROUGE‐2 = 0.1915, ROUGE‐L = 0.3128, and ROUGE‐Lsum = 0.3127), followed by Meta‐LLaMA‐3‐8B. In the reported generation experiment, cluster identifiers, LASER distances, and retrieved historical actions were not supplied as additional Large Language Model (LLM) input; consequently, the present results should not be interpreted as evidence that clustering causally improves generation quality. Moreover, ROUGE measures lexical similarity rather than technical correctness, executability, or safety. The framework is therefore positioned as a decision‐support approach whose generated recommendations require domain‐expert verification before operational use.