TRUST-TREND: robust and generic trend detection

Contact person: Johannes Voll Kolstø

Status: Pre-project.

The pre-project explores robust, generic methods for trend detection in time series: deciding whether a noisy signal is flat or contains a meaningful increase or decrease. The core idea is to recast this as a supervised classification task, training deep learning classifiers on large sets of weakly annotated pairs (X, y), where X is a time series and y a single trend-type label such as "flat" or "contains trend".

This matters because trend detection underlies early event detection across domains, ranging from machine faults to disease outbreaks. Classical change point methods require manual preprocessing and problem-specific modelling, and can perform unsatisfactorily on real, messy data.

The novelty lies in replacing explicit modelling of change and noise processes with a data-centric approach: weak labels make large-scale annotation feasible, capture the context-dependent nature of meaningful trends, and lay the groundwork for future foundation models for trend detection. This pre-project will deliver open-source implementations, pilot studies on data from TRUST partners, and, subject to partner approval, an open benchmark dataset.

Collaboration with user partners is central: the pre-project is motivated by needs from Statkraft and ABB (Action Cluster AC10) and may well be relevant for a variety of TRUST partners. The 6-month pre-project (July 2026 – December 2027) may lead to a larger follow-up project on foundation models for change point detection.