Deep learning GAN-based fault detection and diagnosis method for building air-conditioning systems

Operational data of building air-conditioning systems can provide a lot of useful information for intelligent control and fault diagnosis of air-conditioning systems. This study presents a semi-supervised fault detection and diagnosis (FDD) method for air-conditioning systems using transfer entropy (TE), deep convolutional generative adversarial network (DCGAN), and convolutional neural network (CNN) methods. The FDD method consists of a transfer-entropy-based fault feature variable extraction module, a DCGAN-based training data construction module, and a CNN-based fault diagnosis module. Firstly, both transfer entropy and cause-effect diagram are utilized to select fault feature variables for different faults. Subsequently, the DCGAN method is utilized to construct training data of fault diagnosis models. The CNN fault diagnosis models are trained by using a training dataset consisting of labeled, unlabeled, and generated fake samples. Finally, the CNN fault diagnosis models trained with the extended training dataset are used to diagnose faults in air-conditioning systems. The FDD method was validated by using the operational data of nine types of faults involving two existing buildings. The validation results indicated that the proposed FDD method can diagnose faults in air-conditioning systems accurately. The enhanced CNN fault diagnosis models were robust to the operational data from air-conditioning systems in different buildings.