Analyse comparative d’algorithmes de classification de séries temporelles multidimensionnelles à partir de jeux de données open source pour la détection et le diagnostic de défaillances dans les systèmes CVC.

Benchmarking multivariate time series classification algorithms on open-source datasets for fault detection and diagnostics in HVAC systems.

Numéro : 3581

Auteurs : ABDOLLAH M. A. F., SCOCCIA R., APRILE M.

Résumé

With the widespread adoption of building automation systems, alongside the progress in data analytics, sensing, and machine learning, the domain of data-driven fault detection and diagnostics (FDD) for building heating, ventilation, and air conditioning (HVAC) systems has attracted increasing interest. Numerous studies have tested various algorithms using diverse data sources, including simulations, laboratory testing, and real building environments. However, there is a notable gap in the literature regarding systematic benchmarking of these algorithms against each other using the same open-source datasets. In this study, we undertake a comprehensive benchmarking of different classification algorithms tailored for multivariate time series classification. We employ a publicly available data set provided by Berkeley Labs, which includes ground-truth data concerning the presence and absence of building faults. This dataset covers a wide spectrum of seasons and operational conditions, encompassing multiple building system types. It also includes detailed information on fault severity and data points indicative of measurements in building control systems, which are typically accessible to FDD algorithms. The data compilation leverages both simulation models and experimental test facilities. Our findings suggested that Canonical Interval Forest (CIF) and K-Nearest Neighbors (KNN) with Dynamic Time Warping (DTW) have the highest average performance over the datasets analyzed with 0.78 and 0.73 respectively. This is particularly notable given the lower computational resources required by these methods compared to deep learning-based classifiers.

Documents disponibles

Format PDF

Pages : 10 p.

Disponible

Gratuit

Détails

  • Titre original : Benchmarking multivariate time series classification algorithms on open-source datasets for fault detection and diagnostics in HVAC systems.
  • Identifiant de la fiche : 30032905
  • Langues : Anglais
  • Sujet : Technologie
  • Source : 2024 Purdue Conferences. 8th International High Performance Buildings Conference at Purdue.
  • Date d'édition : 15/07/2024

Liens


Voir d'autres communications du même compte rendu (63)
Voir le compte rendu de la conférence