Program
0830-0900 Coffee
0900-0930 Introduction to GenAI models for time series - Alireza Naziri Khansari?
Abstract: Time series tasks include classification, clustering, forecasting, imputation, anomaly detection, segmentation, pattern detection, and change point detection.
These tasks are essential for understanding trends and dynamics in time-dependent data across various domains. Applications in subsurface domain range from production forecasting, decline curve analysis, well logs imputation, pattern recognition as well as other applications such as oil price prediction, forecasting green energy production and consumption and more.
Time series methods can be categorized into classical statistical approaches, mathematical methods, machine learning/deep learning models, hybrid techniques and, more recently, generative AI models.
We'll cover architecture of new generative AI models for time series, such as TimeGPT (Nixtla), TimesFM (Google), MOIRAI (Salesforce), Chronos(Amazon), Tiny Time Mixers (IBM), and MOMENT (Carnegie Mellon & University of Pennsylvania) and metric evaluation and comparison with classical time series methods.
0930-1000 Possibilities and challenges of 2D timeseries from fiber optic monitoring - Silje Fuglerud Schwermer
Abstract: In Equinor, we monitor several of our assets using optical fibers, which can cover several km and give data with a spatial resolution down to 0.25 m and temporal resolution on the order of less than milli seconds or minutes (based on technology). The data is highly correlated in both space and time, and dependent on both the asset being monitored and the monitoring equipment (interrogators). This talk will show some data examples and discuss challenges with this type of data related to data quality, processing, interpretation, and subsequent classification. In addition, future possibilities of analysis and applications will be discussed.
1000-1030 Omnia.Prevent - Predictive Maintenance at Scale - Hugo Bettencourt Machado
Abstract: Omnia.Prevent is Equinor’s flagship ML product whose value proposition is to predict machine failures early, allowing for timely planning of maintenance activities therefore keeping installation downtime to a minimum. It has been running in production for 5 years with lots of success stories. With many machines to monitor, a large amount of time series sensor data to ingest, numerous ML models to manage, and an organization used to do things the traditional way, this talk will be about how we make it all work.
1030-1045 Break
1045-1105 Maintenance strategies, challenges and opportunities - Arne Bang Huseby (UiO)
Abstract: Predictive maintenance (PM) is a data-driven maintenance strategy that leverages advanced analytics, machine learning, and modern sensor technology to anticipate equipment failures before they occur. By continuously monitoring asset health through real-time data collection and analysis, PM helps organizations optimize maintenance schedules, reduce downtime, and lower operational costs. Unlike reactive or preventive maintenance, which rely on scheduled inspections or breakdown response, PM enables proactive interventions based on condition-based insights. In this presentation we will give an overview of the methodology, and discuss some important challenges.
1105-1130 Advances in Change and Anomaly Detection: Challenges and Statistical Methods - Per August Moen (UiO)
Abstract: Change and anomaly detection are active areas of research at the University of Oslo, as well as within the broader fields of statistics and machine learning. This presentation will provide an overview of key challenges encountered in these domains and advocate for the adoption of statistically founded methods. I will showcase some recent advancements and practical applications of these methods through three case studies particularly relevant for dScience partners.
1130-1200 Light lunch