IIR document

A study on leakage fault prediction of propane room air conditioner under cross-conditions based on transfer learning.

Author(s) : YANG J., WU Z., WU J., HUANG D., DAI X., CHEN Z.

Type of article: IJR article

Summary

To facilitate the extensive application of the flammable refrigerant propane (R290), accurate fault prediction is paramount. However, historical studies have predominantly relied on ‘refrigerant undercharge’ experiments—static states with fixed low charges—as surrogates for leakage. This approach fails to capture the
dynamic transient behavior of an active leak. Addressing this gap, this research proposes a predictive model validated through real-time dynamic leakage tests. These tests were utilizing capillary tubes to simulate continuous refrigerant loss, thereby providing a rigorous dataset for model validation. Experimental data from
actual refrigerant leakage scenarios were subsequently utilized to eliminate theoretical characteristic parameters exhibiting Pearson correlation coefficients lower than 0.85. We utilized transfer learning to develop the predictive model. The BiLSTM-FT model demonstrated robust adaptability, achieving a 17.65% reduction in Mean
Absolute Error (MAE), a 16.67% decrease in Root Mean Squared Error (RMSE), and a 4.50% increase in the coefficient of determination (R²) compared to standard RNN and LSTM baseline models. Furthermore, the CNNFT model for leakage location identification exhibited superior generalization capabilities, identifying leak
sources with an accuracy exceeding 0.91 across all tested conditions. Notably, compared to traditional classification models (RF, SVM), the CNN-FT model improved diagnostic accuracy by 46.74% under maximum operational cooling conditions and by 62.13% under minimum operational cooling conditions. The findings demonstrate the models' robust applicability to air conditioners at cross-conditions, providing a powerful tool for enhancing safety and efficiency in the management of these systems.

Available documents

Format PDF

Pages: 18 p.

Available

  • Public price

    20 €

  • Member price*

    Free

* Best rate depending on membership category (see the detailed benefits of individual and corporate memberships).

Details

  • Original title: A study on leakage fault prediction of propane room air conditioner under cross-conditions based on transfer learning.
  • Record ID : 30035264
  • Languages: English
  • Subject: Technology
  • Source: International Journal of Refrigeration - Revue Internationale du Froid - vol. 189
  • Publication date: 2026/09
  • DOI: http://dx.doi.org/https://doi.org/10.1016/j.ijrefrig.2026.107011

Links


See other articles in this issue (30)
See the source