Fault Diagnosis of Uninterruptible Power System Based on Gaussian Mixed Model and XGBoost

Abstract

Uninterruptible power system (UPS) is an important equipment for guarantee data security. The previous research on UPS fault detection has established a mathematical model of the circuit system and combine with machine learning algorithm for fault diagnosis. Because the UPS circuit structure is complex and easy influence by environment temperature humidity, the research method of establishing mathematical model is suitable for the diagnosis of some internal circuits, but not for the whole UPS machine. Based on the analysis of UPS historical data, this paper has established a method that can catch and classify the UPS fault based on the Gaussian Mixed Model (GMM) and the eXtreme Gradient boost (XGBoost). This design is divided into two parts. The first part has used GMM to calculate the logarithmic summation probability of UPS data, and analyze the relationship between UPS logarithmic sum probability and fault working condition. Then set the threshold to capture the UPS fault according to the fault data recall rate. The second part was fault recognition by XGBoost model. The extremely unbalanced fault data makes it difficult to classify, so XGBoost algorithm has been used to do the classification. combine learning curve with grid search to optimize XGBoost parameters. Experimental results show that GMM can accurately detect equipment faults.