Machine learning, neural networks

Data preprocessing for building a neural network model to predict the state of a technical object

Yu. E. Kuvayskova, A. A. Nemykin

Ulyanovsk State Technical University, Ulyanovsk, Russia

Abstract: In order to prevent emergency situations in the operation of a technical facility, it is necessary to predict its state. To solve this problem, neural network models are used in the work. However, for effective training of models and obtaining more accurate forecasting results, the input data should be preprocessed. In this paper, a new technique is proposed for preprocessing the initial data when constructing neural network models, which includes algorithms for finding outliers, restoring missing values, and removing correlating factors. A special program in the Python programming language was written to implement the proposed technique. The study of the effectiveness of the proposed data preprocessing technique for predicting the state of a technical facility was carried out using two objects as an example: a turbojet engine and a lithium-ion battery. The following approaches were used to compare the results: the data preprocessing technique from the AutoKeras library and the method based on the use of a compactness profile. It is shown that the use of the proposed data preprocessing approach increases the forecasting accuracy of neural network models by approximately 3–4 times compared to the other two approaches.

Keywords: technical object, forecasting, neural networks, data preprocessing, isolation forest method, MICE method, principal component method.

DOI: 10.14357/20718594250106

Bibliographic databases:

•