Monitoramento Inteligente da Operação de Dressagem de Rebolos Convencionais: Uma Nova Metodologia Baseada em Dados de Emissão Acústica, Análise Tempo-frequência e Redes Neurais Convolucionais

Abstract

This work proposes a methodology based on artificial intelligence to assist the operator in decision-making during the dressing operation. Dressing is essential for the performance of the grinding process, but it relies on the operator's experience to determine the optimal moment to interrupt the operation. The research aims to classify grinding wheels into two conditions: undressed (not dressed, needing reconditioning) and dressed (dressed, suitable for use). For this, dressing tests were conducted on a grinding machine, collecting acoustic emission (AE) signals. These signals were processed to generate Short-Time Fourier Transform (STFT) spectrograms, which were used to train and validate a convolutional neural network (CNN). The proposed CNN has 3 convolutional layers followed by max pooling, a dense layer, and an output layer with 1 neuron using the sigmoid activation function. Despite the relatively small dataset (287 images for training and 71 for validation), the proposed model correctly classified all samples. Previously unseen images were also verified and satisfactorily classified as belonging to dressed or undressed grinding wheels. The neural model effectively classified grinding wheels in dressing tests conducted with different parameters. This work presents a promising methodology to assist the operator in decision-making during dressing, using artificial intelligence and acoustic emission signals, concluding that the model can be improved to ensure fault correction and reduce production costs.