Abstract
The purpose of this research is to establish a method of detecting and diagnosing faults in heat source systems by applying machine learning. In a previous paper, a fault database that includes six types of fault conditions was generated through detailed simulation, the features of the faults were extracted from the database using a Convolutional Neural Network (CNN), and it was elucidated that Fault Detection and Diagnosis (FDD) was possible by the learned CNN. However, preprocessing of the raw fault database was not examined for more appropriate diagnosis of the real data called Building Energy Management System (BEMS) data. The objective of this study is to propose a method for preprocessing the raw fault data and to elucidate the effectiveness of the FDD with the proposed preprocessing method compared to BEMS data analysis. The operation data collected in 1 day was regarded as 1 data to be learned or diagnosed. We targeted a heat source system that comprises two chillers and cooling towers, pumps for chilled water and condenser water respectively, and secondary chilled water pumps. First, we generated the fault database that has 39 types of faults such as improper set values and sensor errors. The simulation for the database was run based on the input from real data over one year in 2013. Then, we conducted four case studies on the preprocessing of the database as follows: -Case1, a case that utilizes the normalized database -Case2, a case that utilizes the normalized discrepancy between the condition without faults and the fault data -Case3, a case wherein extracted data that is different from the condition without faults in addition to Case2 -Case4, a case where the number of each fault data is equal in addition to Case3. As for the learning accuracy in the cases, the CNN learned the fault database with high accuracy in Case3 (98.5%) and Case4 (98.3%). Then, in each case, BEMS data over one year in 2014 was diagnosed and the BEMS data was analyzed based on the diagnosis. As a result of comparing the diagnosis by the CNN and the BEMS data analysis, Case4 diagnosed the BEMS data most properly, and it mainly diagnosed five types of faults according to the period when the faults actually occurred. From the above results, it was clarified that even though the system has a fault, the features of the fault do not always emerge in the data, and it emerges according to the operation of the system. This phenomenon was identified from both the diagnosis by the CNN and the BEMS data analysis. In the diagnosis of the BEMS data, there was a period when only one fault was diagnosed despite the occurrence of two types of faults. Therefore, the development of a method that enables detecting and diagnosing multiple faults more accurately is a future task.
