Cross-unit soft fault diagnosis for VRF systems using deep transfer learning: a comparative study across multiple scenarios

Soft faults in VRF systems are difficult to detect, often resulting in air conditioning systems operating in a "sick operation" state, which leads to significant energy waste. This study aims to develop a cross-unit soft fault diagnosis method for VRF systems based on deep transfer learning, addressing limitations in handling cross-condition and cross-unit scenarios. Two distinct transfer learning approaches were investigated and compared for different diagnostic scenarios. First, using 1-D CNN as the base classifier, parameter-based models (FE and FT) were constructed and evaluated under conditions with minimal target domain samples. The FT model achieved an accuracy of 77.4 %. Second, a feature-based domain-adversarial neural networks (DANN) model was constructed with unlabeled target domain data, achieving approximately a 25 % improvement in accuracy over traditional classifiers. These results highlight the potential of deep transfer learning methods for improving diagnostic performance and their applicability in real-world VRF system scenarios.