Model-Based Fault Detection and Diagnosis Using Convolutional Neural Networks in Heat Source System with Water Thermal Storage Tanks

4th Asia Conference of International Building Performance Simulation Association

土木建筑工程

Miyata, S.^1 ; Akashi, Y.^1 ; Lim, J.^1 ; Tanaka, K.^2 ; Tanaka, S.^3 ; Kuwahara, Y.^4

School of Engineering, University of Tokyo, 7-3-1 Hongo, Bunkyo Tokyo

1138654, Japan^1

Tokyo Electric Power Company Holdings Inc., 1-1-3 Uchisaiwaicho, Chiyoda Tokyo

1008560, Japan^2

TEPCO Energy Partner, Incorporated, 1-11-1 Kaigan, Minato Tokyo

1050022, Japan^3

MTD Co. Ltd., 1-2-6 Minamiooi, Shinagawa Tokyo

1400013, Japan^4

关键词: Convolutional neural network;    Expert knowledge;    Fault detection and diagnosis;    Fault severities;    Heat exchanger fouling;    Model-based fault detection;    System behaviors;    Using probabilities;   

Others  :  https://iopscience.iop.org/article/10.1088/1755-1315/238/1/012070/pdf
DOI  :  10.1088/1755-1315/238/1/012070

学科分类：土木及结构工程学

来源: IOP

【 摘 要 】

Since a heat source system that produces and conveys the heat for air conditioning comprises various devices, and has complex controls, faults that impair performance occur. Heat source systems are customized, however, and are complicated to the extent that a 'one-fits-all' approach to fault detection and diagnosis (FDD) has not been established. Here we propose a novel method for FDD, using a fault database generated by a simulation, and using convolutional neural networks (CNNs) trained by the database. A real system, with a water thermal storage tank, was the object of this research. Firstly, system behaviors in response to faults were calculated, using a detailed simulation, and then a database was generated, using the simulation results with fault labels. There were 16 fault types in total, which included a condition without faults, four types of faults, and their combinations. The assumed four types of faults: were chiller deterioration by condenser fouling, improper sewage pump set value, heat exchanger fouling, and temperature sensor error at the supply side of the heat exchanger. Then, we preprocessed the database, and converted the data into images, with two axes of time series, and with items from one 24-hour data set as a representative image. Then, CNNs were trained by the database, and trained CNNs were tasked with the diagnosis of real data. The CNNs performed with 98.7% accuracy in training, and diagnosed the real data using probabilities. We reviewed the analysis of the real data, where the probability indicated the likely presence of a fault, and how the real data was similar to the fault severity assumed in the simulation. We concluded that this FDD method will help analyze real data, because it indicates faults emerging in the real data with probability, whereas conventional data analysis requires checking the data using expert knowledge.