Application of Machine Learning Technology to Management Methods

Abstract

Multicomponent meat system -brine injected pork and cooked sausage "Doctorskaya" were was analyzed using neural network technologies and the conditions of uncertainty and risk of human error in a decision-making process in time domain were identified. The formation of a situational classifier (digital image-based histology -meat sample sections with a detailed description) and the system's knowledge base was described. The general steps of a histological section image processing are: 1) preprocessing of section images (noise removal, palette optimization, etc.); 2) color segmentation based on palette minimization; 3) approximation of boundaries of the areas highlighted in the image; 4) area size determination; 5) particle shape determination; 6) particle color determination; 7) identification of the presence of counterfeits; 8) results' output regarding the determination of the presence of counterfeits. The Jupyter Notebook and Colaboratory software environment was used to study and compare the influence of several activation functions (ReLu, tanH, eLu, sigmoid, softPlus, softSign) on the generated DataSet. The best result was obtained when with ReLu (0.9843) activation function, followed by SoftPlus (0.9765) and eLu (0.9687) activation functions. This stage of the study considered a kind of convolutional neural network (CNN) architecture with two layers of convolution (Convolutional, C-Layer) and pooling (Subsampling, S-Layer). An algorithm of the Error Back Propagation gradient was applied to train CNN. This is the first research stage for convolutional neural network applications in solution management.