Variable refrigerant flow (VRF) systems are widely used in buildings. Soft faults are common and difficult to identify during VRF operation

making the system less efficient. In this study

a soft-fault level identification model for VRF was proposed based on a classifier chain using one-dimensional convolutional neural networks as the base classifiers. The structure and parameters of the base classifiers were set according to a fault diagnosis model using the experimental data of fouling faults in the outdoor unit; two new methods for encoding data labels were proposed. After establishing the initial soft-fault level identification model

the number of convolution kernels in the base classifiers was further adjusted and a magnification factor was proposed to improve the label encoding. The results showed that the improved classifier chain model can diagnose fouling faults in the outdoor unit with an accuracy greater than 96%

corresponding to an increase of 2%–3% from the baseline. The encoding methods proposed in this study did not diagnose faulty conditions as normal and are suitable for use in the classifier chain model.