Determining the application domain of machine learning using anomaly detection and application to refrigerant leak detection

Abstract

Predictions using machine learning often make prediction errors for data outside the range of the training data. When correct predictions are needed, it is sometimes better to output unpredictable rather than erroneous predictions. However, a lot of effort is required to find predictable conditions and filter out unpredictable data. We propose a method to calculate the applicable domain (AD: Applicable Domain) for a learned prediction model (regression model) using an anomaly detection algorithm that can output anomalous values numerically. The results of applying this method to a refrigerant leakage estimation model for air conditioners show that the prediction accuracy is equivalent to that under conventional filtering conditions, while the filtering conditions can be relaxed and the number of predictable data can be increased.