Partager

Perspectives de généralisation d’un algorithme de classement des défaillances par apprentissage automatique s’appuyant sur un ensemble pratique de fonctionnalités pour les unités de toiture.

Generalizability of a machine-learning fault classifier utilizing a practical set of features for rooftop units.

Numéro : 2178

Auteurs : UDDIN R., YUILL D. P., WILLIAMS R. E.

Résumé

Soft faults in rooftop units (RTUs) degrade the system performance, impacting equipment, economics, and the environment. Automated Fault Detection and Diagnosis (AFDD) can use a data-driven approach by fitting machine learning classifiers that can predict typical soft faults using suitable inputs. Among several other issues, a challenge for fault detection and diagnosis protocols is getting a reasonable trade-off between the number of false alarms and missed detections. In addition, the practical deployment of machine-learning fault classifiers will require the most practical set of feature inputs for rooftop units. Finally, practical machine-learning classifiers will need to be able to predict faults from a system different from those with which the classifiers were trained. To obtain a more generalizable fault classifier, this study proposes a machine-learning classifier that was trained by using simulation data for multiple rooftop units, with a limited set of input features, addressing these practical challenges of application. The proposed classifier was tested using: (i) existing laboratory measurement data, and (ii) field data from a faulty RTU. The results are promising. This indicates that the proposed classifier could be generalizable for diagnosing common soft faults in RTUs. This study focused on fixed orifice (FXO) equipped systems because of the availability of training data, but the results suggest that the method could be adapted to the more common TXV-equipped systems if training data is available. The paper shows the diagnostic accuracy in terms of false alarms, missed detections, and misdiagnoses, and describes some of the important methods required to achieve good accuracy with machine-learning-based diagnostics, such as rebalancing the training dataset and selecting meaningful features.

Documents disponibles

Format PDF

Pages : 10 p.

Disponible

Gratuit

Télécharger

Détails

  • Titre original : Generalizability of a machine-learning fault classifier utilizing a practical set of features for rooftop units.
  • Identifiant de la fiche : 30033066
  • Langues : Anglais
  • Sujet : Technologie
  • Source : 2024 Purdue Conferences. 20th International Refrigeration and Air-Conditioning Conference at Purdue.
  • Date d'édition : 17/07/2024

Liens


Voir d'autres communications du même compte rendu (187)
Voir le compte rendu de la conférence

Indexation