Share

Recommended by the IIR

Using machine learning for feature selection in automated fault detection and diagnosis of split system air conditioners.

Number: 2351

Author(s) : CHEN Y., EBRAHIMIFAKHAR A., YUILL D.

Summary

Because of the advancement of smart buildings and smart sensors, enriched data with respect to building operation is now available, making the data-driven approach an appealing method for fault detection and diagnosis in air-conditioners. Data-driven methods, unlike many other fault diagnosis methods, necessitate a substantial quantity of data to develop a model, which limits their deployment in the field. Another issue with using such an approach is the variables' availability. From a cost-effectiveness standpoint, the input variables need to be chosen very carefully, since they constitute a large portion of the cost of applying fault diagnostics. This study provides two feature selection strategies for use in machine learning-based diagnostic techniques, using real-world practical considerations in combination with the SVM classifier, to investigate how the reduction of features affects the fault diagnosis performance of the data-driven approach. This study used a high-fidelity data set with a total of 15 variables, in which seven faults were simulated under various driving conditions. The full data set was split into a training set, with 70% of the original data, and a test set, accounting for 30%. During the training process, the 10-fold cross-validation method was employed to tune the parameters of the model. Compared to the balanced accuracy of 98.9% when all features were used, the removal of the features by sequential backward selection process produced 97.2% balanced accuracy which was only slightly lower. In terms of the feature selection based on real-world availability, 94.7% of balanced accuracy was obtained, indicating that this method can still achieve acceptable accuracy. Precision, recall, and F1 scores are also provided in the paper, to further describe the diagnostic performance for each fault in the data.

Available documents

Format PDF

Pages: 10 p.

Available

Free

Download

Details

  • Original title: Using machine learning for feature selection in automated fault detection and diagnosis of split system air conditioners.
  • Record ID : 30030687
  • Languages: English
  • Subject: Technology
  • Source: 2022 Purdue Conferences. 19th International Refrigeration and Air-Conditioning Conference at Purdue.
  • Publication date: 2022

Links


See other articles from the proceedings (224)
See the conference proceedings

Indexing