Improving the training efficiency of automated fault detection and diagnosis for central chilled water plants by transfer learning

Shohei Miyata, Yasunori Akashi, Yasuhiro Kuwahara, Katsuhiko Tanaka

Proceedings of Building Simulation 2021: 17th Conference of IBPSA, pp.3044-3051, September2021

https://doi.org/10.26868/25222708.2021.30357

Abstract

During the operation phase of heating, ventilation, and air conditioning (HVAC) systems, faults that deteriorate the system performance often occur. Therefore, automatic fault detection and diagnosis (AFDD) is essential for proper operation. The authors propose an AFDD method, which utilizes a fault database generated by a system simulation and convolutional neural networks (CNN) as classifiers. Although this method has demonstrated high performance, it requires a large number of human and calculation resources. In this study, to reduce the calculation resource for network training, transfer learning is applied to the AFDD at different sites. The subjects were central chilled water plants, which are the waterside of HVAC systems. It was confirmed that by applying transfer learning reasonable diagnosis results were achieved and the computational resources during training were halved. These results are expected to contribute to the widespread use of the AFDD method.