Fault diagnosis of variable refrigerant flow system based on the deep reinforcement learning method

Variable refrigerant flow (VRF) air-conditioning systems are widely deployed in modern commercial and residential buildings. Faults in these systems can significantly reduce energy efficiency and increase operation and maintenance costs. To address the scarcity of FDD annotation data for VRF systems and the insufficient accuracy of traditional methods under various fault conditions, this paper is the first to apply deep reinforcement learning (DRL) to the fault diagnosis of VRF systems. Through the experimental system, four common faults were simulated: indoor unit fouling, outdoor unit fouling, refrigerant overcharge, and refrigerant undercharge. 13 key feature variables were selected, and operating data under eight operating conditions were collected for model training and testing. The two algorithms, value-based DQN and policy-based PPO, were systematically compared. The results show that DQN outperforms PPO in terms of convergence speed and diagnostic stability. By optimizing the three hyperparameters of network depth, training iteration steps, and learning rate, the optimal model configuration was determined as a 5 or 6-layer hidden network, with the learning rate ranging from 5 × 10^-5 to 10^-3. The optimized DQN model achieved excellent diagnostic performance: the fault diagnosis accuracy (FDA), geometric mean accuracy (GMA), false alarm rate (FAR), and missing alarm rate (MAR) reached 98.46%, 98.19%, 0.94%, and 0.85%, respectively. The proposed DQN model demonstrates superior performance compared to traditional methods such as SVM, ANN, DT, and K-means clustering, as well as recent research results.