Abstract
For diagnosing the root cause of a fault, we developed an automated fault detection and diagnosis (AFDD) method for cooling plants, which are water-side HVAC systems. In our initial paper (part 1), we described in detail the proposed AFDD method which uses convolutional neural networks (CNN) based deep learning approach and demonstrated its effectiveness. First, the original simulation program of the cooling plant is described. In addition, we show that the system behavior during faults can be calculated by providing the fault conditions. Based on such conditions, a fault database is generated, and the features of which learned using CNN to determine which faulty behavior is close to the actual data. In our second paper (part 2), we expanded the number of fault types from 6 in the first paper to 39 and discussed the data preprocessing method to achieve a high diagnosis performance. In addition, the severity setting of the faults and the diagnostic characteristics of multiple faults were examined and analyzed in a previous related paper. As building and energy management system (BEMS) data continue to be accumulated, sufficient operational data exists for AFDD. However, there is a possibility that multiple faults, including sensor errors, will occur simultaneously in an actual system, and there are only a few effective methods for such data, and only a few applications of AFDD for BEMS data. Based on the above, we constructed a demonstration system for the proposed method and improved the operation based on the AFDD results. The primary objective of this paper is to provide an evaluation of the demonstration system and operational improvements. In the demonstration system, AFDD is automatically executed every day, and the diagnosis results and related system behaviors can be confirmed on a web browser. After discussions with local operators, the lower limit of the condenser water pump frequency, which had been diagnosed as a fault, was adjusted lower. The analysis and simulation of the BEMS data confirmed that an energy saving of approximately 3% can be achieved. The proposed method requires the construction of a simulation program, the generation of a fault database, and the training of CNN for each system, which is expensive in terms of the number of human resources and computations. Therefore, the second objective of this paper is to apply transfer learning to a similar system to reduce the computational cost of learning, and to confirm the accuracy and effectiveness of transfer learning. In addition to the trained CNN created using the proposed method, we also tried to use a pre-trained network called VGG-16 for image recognition, which is trained on more than one million images. As a result, it was shown that transfer learning reduces the computation time by approximately half when introducing AFDD to a new system, while still providing useful a BEMS data diagnosis performance. However, when transfer learning was applied to VGG-16, a sufficient performance could not be obtained. This can be attributed to the fact that the characteristics of the fault data of a cooling plant are unique to a general image. The scalability of the method is important for the widespread use of AFDD. We believe that the application of transfer learning will contribute to this scalability. Maintaining a fault database of an HVAC system is expected to contribute to an improved the scalability of the AFDD method because its features can be used for transfer learning.
