earticle

영어

Generators continue to deteriorate in performance due to aging and result in increased failure rates and reduced reliability. Therefore, studies are being conducted on anomaly prediction models for generator engines to prevent potential accidents during operation. However, there are problems in designing the models due to class imbalance and manual input of maintenance history. This study labels data from the time an anomaly occurs up to 60 minutes before the occurrence as anomalies to solve these problems. Data from the time an anomaly occurs up to 30 minutes before the occurrence were also added as derived variables to reflect the warning signs of anomalies in model training. The anomaly prediction models were created using engine log and maintenance history data and applying Random Forest(RF), eXtreme Gradient Boosting(XGB), Linear Support Vector Classifier(LSVC), and Deep Neural Networks(DNN) algorithms. The performance of the models was evaluated by F1-Score and Recall. XGB showed excellent performance in terms of F1-Score, and DNN in terms of Recall. As a result of comparing the F1-Scores to sort the optimal model for each system, XGB was optimal for systems 1, 2, and 4, and RF was optimal for systems 3 and 5. System 5 showed excellent performance when only the derived variable condition was applied, and the other systems showed excellent performance when applying the derived variable and labeling.