Improving training efficiency for scalable automated fault detection and diagnosis in chilled water plants by transfer learning

Abstract

Automated fault detection and diagnosis (AFDD) is crucial for improving the energy efficiency of air conditioning systems, and its quantitative effects have become clearer in recent years. Although many AFDD methods are proposed in the literature, there is insufficient research on their scalability to different systems. The authors previously proposed an AFDD method using a fault dataset created by simulation that can be used to train convolutional neural network (CNN); however, this AFDD method is difficult to apply to different systems because it requires significant computation time to generate a fault dataset by simulation and train the CNN. In this study, we focused on transfer learning and examined whether reducing the size of a fault dataset and re-training a CNN with only some parameters of the originally-trained CNN would reduce the computation time and make training more efficient. Specifically, we ran simulations, calculated the fault dataset, trained the CNN, and applied transfer learning using two real chilled-water plants. As a result, it was confirmed that by reducing the fault dataset by a certain amount and the limiting the number of epochs for re-training, it is possible to achieve reasonable diagnostic performance while reducing the computation time to less than half compared to our previous method. This study also has value in that it examined the application of transfer learning using multiple real systems.