Zero-Shot Anomaly Detection in Alphanumeric Vehicle Data using Large Language Models : a Design Science Approach

Abstract

This master’s thesis examines the use of large language models for zero-shot anomaly detection in alphanumeric vehicle datasets, filling a gap where traditional statistical methods face limitations. While numerical data can be reliably assessed with algorithms like Local Outlier Factor or Isolation Forest, the high-dimensional nature of alphanumeric serial numbers makes them difficult to model with established algorithms. Using an iterative design science approach, this study develops and tests a Proof-of-Concept Python application that uses state-of-the-art large language models to detect anomalies in real-world vehicle datasets. Besides some prompt engineering, the models are intentionally not fine-tuned, enabling application without in-depth knowledge of large language models. The theoretical background covers data management, anomaly detection, and the core principles of large language models. Results show that large language models, especially Google’s Gemini 2.5 Pro, can effectively identify anomalies in both numerical and alphanumeric data. Compared to statistical algorithms, large language models offer the benefit of processing alphanumeric inputs, adding a valuable extension to the anomaly detection toolkit. However, challenges like hallucination, inconsistent length counting, and sensitivity to highly anomalous datasets highlight current limitations. Additionally, statistical methods remain more efficient, scalable, and cost-effective for purely numerical datasets. The findings confirm that large language models can be applied in a zero-shot manner to detect anomalies in alphanumeric datasets. Beyond the automotive industry, these insights can be applied to other fields where alphanumeric identifiers are essential. This work advances both academic discussion and practical applications, providing a foundation for future research on fine-tuned models and industrial implementation.