Рудик, А. В. та Рудик, В. А. та Rudyk, A. V. та Rudyk, V. A. (2018) РОЗРОБКА СТРУКТУРИ ШТУЧНОЇ НЕЙРОННОЇ МЕРЕЖІ ПРИЛАДОВОЇ СИСТЕМИ МОБІЛЬНОГО РОБОТА. Вимірювальна та обчислювальна техніка в технологічних процесах. с. 99-101.

Перегляд

Text Розробка ШНМ мобільного робота (Рудик А.В.) зах.pdf Download(268kB) | Перегляд

Анотація

Для обчислення з високою точністю відстані до перешкоди за значеннями кількості лічильних імпульсів та швидкості ультразвуку в повітрі проведено розробку та аналіз моделей штучної нейронної мережі (ШНМ) типу Feed-forward (FF) і Cascade-forward (CF). За результатами проведеного аналізу ШНМ типу Cascade-forward показали кращі результати, ніж ШНМ типу Feed-forward, при цьому CF-мережі за точністю мають приблизно однакові результати (для ШНМ з одним прихованим шаром MSE=1.577 мм, MAE=2.035 мм, а для ШНМ з двома прихованими шарами MSE=1.412 мм, MAE=1.892 мм). Однак за рахунок кращих точнісних характеристик для використання в приладових системах МР пропонується ШНМ CF-типу з двома прихованими шарами.

Title in English

DEVELOPMENT THE ARTIFICIAL NEURAL NETWORK OF THE MOBILE-ROBOTS DEVICE SYSTEM

English abstract

To calculate with high accuracy the distance to the obstacle by the number of counting pulses and the speed of ultrasound in the air, the models of an artificial neural network (ANN) of the Feed-forward (FF) and Cascade-forward (CF) type were developed and analyzed. The nature of the nonlinearity and the complexity of the relationship between the parameters are unknown in advance, so the number of hidden layers, neurons in the layers and the activation functions of neurons are chosen experimentally. As a result of the analysis, a set of 69 ANN models of different structures was developed and analyzed. For learning, the method of backward propagation of error was used. Based on the results of the analysis of 20 CF-type ANNs, the most accurate is an ANN consisting of 6 neurons with transfer function (TF) tansig and 9 neurons with TF purelin in the hidden layers and one neuron with TF purelin in the output layer. As a result of the training of this ANN with the duration of the 42nd epoch, the mean square error (MSE) of the training was 0.723 mm, and according to the results of the ANN testing on an additional sample, the MSE was 1.526 mm and the average absolute error was 1.965 mm. Since in the learning process the full sample is divided into parts (60% for training, 20% for testing during training and 20% for testing), then the adequacy of the developed models was not tested on the training data vectors. For the final selection of the ANN structure, the samples for testing and testing were combined in a total size of 50, and the accuracy of the models that had previously shown the best results was checked on this combined sample. Based on the results of the analysis, the ANN of the CF-type showed better results than the ANN of the FF-type, while the CF networks have approximately the same accuracy (for an ANN with one hidden layer MSE=1.577 mm, MAE=2.035 mm, two hidden layers MSE=1.412 mm, MAE=1.892 mm). However, due to the best accuracy characteristics for use in MR systems, an ANN CF-type with two hidden layers is proposed.

Перегляд елементу

Завантажень