Generation and Evaluation of Classification Models Using Multi-Layer Neural Networks Utilizing Maintenance Records of Air Conditioning Heat Source Facilities with Thermal Storage

Abstract

Because the frequency of repair and renewal of the air conditioning facilities is typically greater than that for the buildings in which these facilities are installed, air conditioning facilities require efficient maintenance. In this study, classification models using multi-layer neural networks that utilized the maintenance records of air conditioning heat source facilities were generated and evaluated. First, the failure histories and recorded values were organized and labeled as “failures.” Next, classification models were developed on the basis of an original multi-layer neural network. The generated classification models classified the training data well, and the boundaries of the classification corresponded to the changes in the recorded values identified during labeling. In the evaluation of the ROC curve and area under the curve (AUC), the AUC of most of the classification models was over 0.9, this is than that of the recorded values, thereby confirmed the high accuracy of the models.