Preprint
Article

This version is not peer-reviewed.

Can LLMs Detect Numerical Outliers: LLM Aided Machine Prognosis (LAMP)

  † These authors contributed equally to this work.

Submitted:

17 November 2026

Posted:

19 November 2026

You are already at the latest version

Abstract
LLMs are well known for their ability to retrieve knowledge from data sources and generate text at scale taking benefit from their Decoder based architecture; however, their capability to handle long, numerically intensive telemetry datasets remains largely unexplored. In this paper, we study the behavior of LLMs in the analysis of industrial machinery telemetry to explore their ability to identify patterns in bearing vibration datasets. For the vibration data, characteristic frequency metrics have been derived using classical physics-based formulas, and synthetic anomalies have been introduced through controlled signal perturbations. Anomaly detection in industrial machinery telemetry is essential for early fault identification and predictive maintenance across a wide range of equipment types. In our experiments, LLMs demonstrated varying performance depending on the model's capabilities in reasoning and mathematical tasks. We benchmarked LLMs' performance against established machine learning models. Since machine learning (ML) models are based on numerical computations and loss function minimization, they are generally expected to perform better in detecting anomalies. However, in our experiments, we observed that LLMs, when asked for multishot serialized telemetry data, matched the precision and recall of traditional ML models and, in some cases, exceeded the accuracy by up to 9.09%. We propose an LLM-aided machine prognosis (LAMP) framework, to utilize the growing capability of LLMs for analysis of complex machinery telemetry dataset using conversational natural language.
Keywords: 
;  ;  ;  
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.