Uso de inteligência artificial para quantificar e classificar impurezas em cana-de-açúcar para fins industriais

• Autor(a) principal: Santos, Lucas Janoni dos
• Data de Publicação: 2021
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://hdl.handle.net/11449/214015
• topic: Artificial Neural Networks; Digital images; Sugarcane; Impurities; Classification; Prediction; Cana-de-açúcar; Impurezas; Classificação; Predição; Redes neurais artificiais; Imagens digitais

Artificial neural networks (ANNs) are computational models similar to the structure and functioning of the human brain. They are systems in which basic processing units (neurons) are distributed and interconnected by different connections (synapses), implemented by different mathematical functions, vectors and matrices. They present the ability to acquire and record information from the processing of a database, having a high capacity to map and solve non-linear problems. Due to these characteristics, RNAs were used to quantify and classify the mass percentage of sugarcane and impurities typically found in their harvest, since they can cause increased energy consumption in the preparation of the cane, extraction losses, difficulties in the broth’s treatment and equipment wear. The quantification and classification of the 146 samples were made using digital information from digital images converted into color scales (RGB, HSV, rgb and L), a non-invasive data acquisition method. The classification involved dividing the samples into two classes: 90-100% (class 1) and 41- 89% (class 2) of sugarcane, by weight. The implementation of artificial neural networks were performed in the MATLAB R2018a software with the nnstart tool, using the Levemberg-Marquadt algorithm (trainlm) for the prediction models and the scaled conjugated gradient algorithm (traincsg) for the classification models. The neurons’ number in the intermediate layer was adjusted by trial and error until the best result by the ANNs was achieved. In both processes there were a random division of the samples: 70% for the training of the network, 15% for its validation and the remaining 15% for the ANN test. The results obtained by the ANNs were very promising, since it was possible to predict the amounts of sugarcane, plant material and soil with low absolute errors and high correlation coefficients. It was also possible to classify all samples correctly, even in the face of a complex and multivariate data set such as digital image parameters.