Прогноз Состояния Объекта На Основе Применения Фильтра Калмана И Глубоких Нейронных Сетей

Abstract

The paper presents an algorithm for predicting an object state based on data from different sources (for example, video cameras) coming in the form of images aimed at critical technological zones. The proposed algorithm is based on the consistent use of a deep artificial neural network and the Kalman filter. A neural network is designed to reduce the input data dimension (images) performing the function of an encoder, which gives of an observation vector of the object state on the output. Based on these observations, the object state is evaluated by a recurrent filter. Using the filter directly for images would lead to a large dimension of the problem; it would be impossible to perform it practically due to computational difficulties. The program that implements the proposed algorithm was developed in Python 3.6 using the Spyder integrated environment from the Anaconda assembly for the Linux operating environment. The choice of a programming language is due to the availability of powerful libraries for machine learning TensorFlow from Google, as well as the convenient Keras framework for creating and working with deep neural networks. The paper describes the results of a model experiment on using the proposed algorithm for predicting an object state, which consisted in attributing the obtained observations to a particular class. The experiment also involved generating sets of images belonging to different classes, differing in their texture. A line-by-line horizontal pixel shift simulated the noise in the images. The comparative analysis of the predicted results with and without using the Kalman filter has shown that filtering reduces the number of false classifications. The developed algorithm might be used in decision support systems and automated process control systems.