АНАЛИЗ МАТЕМАТИЧЕСКИХ МОДЕЛЕЙ ПРИ ПОСТРОЕНИИ НЕЙРОННОЙ СЕТИ ДЛЯ ОЦЕНКИ РАБОТОСПОСОБНОСТИ АГРЕГАТА

Abstract

Prevention of emergencies in aviation technology is largely ensured by diagnostics of the functioning of its units. Often the performance criterion is the level of vibration that influences the decision to adjust the load or shut down the unit. The article discusses mathematical models when applying neural network methods for vibration diagnostics. When using cross-validation, the initial sample with vibration data is divided into several blocks, which are grouped into three samples: training, control, and test. To assess the effectiveness of diagnostics, three different quality criteria were used: mean error in the test sample, AUC, and F-measure. For a given set of initial data, the best fitted configuration turned out to be a neural network of three layers with 18 neurons in each layer, implemented in the MATLAB package. It uses a Bayesian regularization algorithm as a learning function. The percentage of the average error in recognizing the state of the considered aggregate using the neural network turned out to be 4.85, the AUC value was 0.885, and the F-measure was 0.827. Compared to a network built in automatic mode using the Statistics and Machine Learning Toolbox and Neural Network Toolbox machine learning libraries, the F-measure of the fitted network configuration is 6.7% higher.