Active Learning-based Online Coupling of Sawmill Simulators and Their Surrogate Model: The Effect of Sampling Bias on Concept Drift Detection
Résumé
Coupling numerical models with different computational costs and fidelity levels is a promising solution for developing efficient industrial digital twins able to process fast data streams. Such a coupling strategy between a high fidelity but computationally intensive simulation model and its machine learning-based surrogate model in the context of the sawmill industry had been proposed in a previous publication [1]. The strategy proposed is inspired by active learning and based on a measure of prediction confidence. It relies, however on the assumption that the input data stream models make predictions upon is stationary. Concept drifts are, however, frequent in an industrial context. It is, therefore, necessary to integrate into the proposed strategy a mechanism to detect such drifts. The present article evaluates the usage of several drift detection methods in conjunction with the previously proposed coupling strategy and highlights the issue caused by sampling bias. In particular, it can increase the time needed to detect the drift. Four datasets are used for evaluation. The first contains sawing simulation results, while the three others are benchmark datasets from the literature.
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |