A Data-Driven Study on the use of Small Language Models & Large Language Models for Interpreting Complex Industrial Data and Deriving Actionable Outcome

Abstract

Modern industrial systems generate unprecedented volumes of heterogeneous data, spanning high-frequency sensor streams, SCADA outputs, unstructured maintenance logs, and operational reports. Integrating and interpreting these disparate data sources creates a persistent challenge for timely decision-making, as traditional predictive models often lack the contextual reasoning required to translate raw metrics into actionable operational insights. This paper presents a comprehensive, data-driven study evaluating the efficacy of Small Language Models (SLMs) and Large Language Models (LLMs) in interpreting complex industrial data and deriving actionable outcomes. We systematically compare how these models differ in their capacity to summarize complex operational patterns, explain transient anomalies, infer root causes, and prescribe maintenance decisions. Crucially, this evaluation grounds model capabilities against practical industrial constraints, including inference latency, computational cost, data privacy, and edge-deployment feasibility. A data-driven comparative framework is proposed to rigorously evaluate model performance across interpretation quality, domain-specific accuracy, robustness to noisy data, and the tangible usefulness of generated recommendations. Furthermore, this study investigates the viability of a hybrid, tiered architecture—combining the localized, low-latency efficiency of SLMs for immediate processing with the advanced, high-level reasoning capacity of LLMs for complex diagnostics—to optimize both reliability and operational value. Ultimately, the findings are intended to inform the design of language-model-based industrial analytics systems that are computationally efficient, contextually aware, and readily deployable in real-world settings.