| Abstract [eng] |
The aim of this project was to present the possibility of artificial neural networks (ANN) application for manufacturing efficiency forecasting. By analyzing the nowadays practices of applying ANN in manufacturing processes, five main areas was identified - optimization, modeling/simulation, quality control, image recognition and forecasting. Also the process of artificial neural network creation in the Matlab computing environment was analyzed and four key phases was defined: data collection, processing and division into functional groups; network structure and architecture selection; optimal training function selection; network training, validating, valuating and optimizing. Operating data from manufacturing process of UAB Mars Lietuva used to create ANN driven forecasting model. Using the product type, format, production volume, shift and production line as input data and efficiency index as an output. During the research, more than one hundred neural networks with different structure trained, validated and valuated using actual UAB Mars Lietuva production process data. Fast-forward and cascade-forward networks verified with five main Matlab multilayer neural network training functions. In addition, ANN optimized by finding the optimal number of neurons in a hidden layer and using early-stopping methodology. Based on the results, the best structure for forecasting model was obtained by using fast-forward artificial neural network with Levenberg-Marquardt (trainlm) algorithm, with seven neurons in the hidden layer and using early-stopping. The predicted percentage result of the generated multilayer ANN model differs from the actual value by an average of 3.17 percent and is 54 percent more accurate than the model (± 6,9%) currently used by the company. Company could use the created model in the future for more precise forecasting of production efficiency. |
