Anomaly Detection in Multivariate Industrial Signals: LLMs, TSFMs, or Classical Deep Learning

Abstract

Large language models (LLMs) offer several distinctive advantages over other machine learning models. First, they are trained as general-purpose models and are readily available, which eliminates the need for task-specific training and allows them to improve rapidly over time. Second, they can be applied directly, without constructing domain-specific or signal-specific models. Third, they are easy to integrate into existing systems and can be deployed without requiring an additional training step. Finally, they are inherently interactive because users can direct them with natural language. In this paper, we investigate whether LLMs can achieve multivariate anomaly detection. To fully exploit the aforementioned benefits, we define a set of guiding principles (such as avoiding pre-learning or representation learning on the signals) to ensure the LLMs remain general-purpose models. Based on these principles, we then propose several algorithmic approaches for building multivariate anomaly detection pipelines. We compare our approaches with two alternatives: (i) classical deep learning pipelines trained specifically for anomaly detection, and (ii) a foundation-model-based approach, in which domain-specific or general purpose time-series foundation models are trained without explicit supervision for anomaly detection but are then used for this purpose. The comparison highlights trade-offs along three key dimensions: anomaly detection accuracy, computational cost, and the amount of domain knowledge required to develop the pipeline. We evaluate our methods through two case studies. The first uses a benchmarking testbed designed for anomaly detection, while the second examines real-world data from wind turbines with known anomalous events.