Production Time Series Monitoring Using Vision-Language Models
Hüden Neşe, K. J. Mann, S. Sæten, Lukas Mosser · 2025
Abstract Identifying anomalies in daily time series data can be critical for maintaining the performance and integrity of a petroleum production system. Previously numerous attempts have shown that statistical and machine learning approaches can provide a means to detect and identify the types of anomalies that may occur on a given set of monitored equipment. The recent advances in leveraging generative pre-trained Large Language Models (LLMs) have been shown to be extendable to tasks on time series data by treating the data as a series of textual characters. In this work we investigate the capability of so-called Vision-Language Models (VLMs) for anomaly detection in time series data from production wells. Our approach presents the relevant time series data as an image to the VLM together with a text-based prompt. Due to the pre-training on large-scale datasets and task-instruction pre-training of the used VLMs we are able to leverage the capabilities of these models as flexible way of describing the presented data and querying for anomalies. We constrain the VLM to a structured output format, ensuring adherence to a specified data-model of required attributes e.g. the presence of an anomaly, the description, and the direction of change in the signal. We evaluate our proposed methodology on a dataset of 1000 expert-annotated instances acquired from various operational settings of Aker BPs production wells and its ability to assign an indicator of whether a time series merits inspection by an engineer. We compare against simple statistical baselines, perform a number of ablations, as well as investigate a risk-based evaluation approach enabled by the probabilistic nature of VLMs. Our findings show that VLMs should be considered as an integral and highly flexible component in a production monitoring system.