Classification of Grinding Wheel Cutting Ability During Dressing via Acoustic Emission and Neural Network

Dressing is essential for precision in grinding. However, its conventional monitoring relies on the operator’s subjective perception, which compromises process repeatability and reliability. This work presents an indirect monitoring methodology for the automated classification of conventional aluminum oxide grinding wheel conditions into three states (undressed, intermediate, and dressed), investigating the comparative performance between a multiclass Multilayer Perceptron (MLP) neural network and a Mamdani-type Fuzzy Inference System. Dressing experiments were conducted with roughing and finishing parameters on a Sulmecanica RAPH 1055 grinding machine equipped with a single-point CVD diamond dresser. Acoustic emission signals generated by the tool-wheel contact were captured by a sensor attached to the tool holder, amplified by an amplifier module, and digitized at a sampling rate of 2 MHz via an oscilloscope. In the MATLAB environment, the effective contact window was defined, and the counts statistic was applied, which counts the number of times the acoustic emission signal amplitude exceeds a preestablished threshold in each interval, generating time-series vectors smoothed by a moving average filter. For classification, the full counts vector was used as input for the one-dimensional MLP network, while the arithmetic mean and standard deviation extracted from this vector fed the Fuzzy system. The comparison between the two models shows that they successfully estimated wheel sharpness under different dressing parameters. In quantitative terms, the Fuzzy model achieved an accuracy of 97% in classifying the three conditions, while the MLP network achieved a minimum performance of 80%. In this configuration, the Fuzzy system exhibited greater robustness, proving to be the most suitable tool to assist the operator in decision-making during the dressing operation.