Prediction of Technical Object State In Conditions of Unbalanced Classes on the Basis of Machine Learning Models

This article describes the method of predicting the state of a technical object in conditions of unbalanced classes in order to provide timely warning about the violation of the objects operability and the possibility of an emergency situation on the basis of machine learning models. The method uses random forest and support vector algorithms as basic machine learning models. However, when applying machine learning models to predict critical situations in the operation of an object, there is often a problem of class imbalance in the source data, i.e., a small amount of information about the emergency state of the object. In such a case, the accuracy of predicting the state of the object is significantly reduced. To improve prediction accuracy, it is necessary to build the model in such a way as to minimize the number of errors in classifying critical situations in the operation of the object under conditions of class imbalance. To solve this problem, the proposed method uses sampling algorithms: the algorithm using Tomek links and the ADASYN algorithm. F- score and AUC ROC criteria are used to evaluate the quality of prediction. The study of the effectiveness of the method of predicting the state of a technical object in conditions of unbalanced classes is carried out on the example of assessing the state of a hydraulic system pump. It was shown that the application of the proposed method improved the value of quality metrics by 3.5%-21% compared to models without data balancing.