Automatic grinding burn recognition based on time-frequency analysis and convolutional neural networks

The grinding process is employed to provide a high-quality surface finish and tight dimensional tolerances to the manufactured components. However, it has the disadvantage of generating a large amount of heat during machining that is mostly transferred to the workpiece when employing conventional grinding wheels, which makes it highly susceptible to thermal damage. In terms of the different thermal damage associated with this process, grinding burn deserves special attention, as it affects the aesthetic aspect of the machined components. Since there is an increasing demand for productivity and high-quality products, the use of systems to monitor grinding burn becomes crucial when global competitiveness is in evidence. In this study, a novel approach, based on time-frequency images of acoustic emission signals and convolutional neural networks was proposed to monitor grinding burn. Experimental data were obtained from grinding tests on N2711 grade steel under different cutting conditions. Three different time-frequency analyses, including the short-time Fourier transform, the continuous wavelet transform, and the Hilbert-Huang transform, were used to generate the images that served as input for the CNN models. Through the proposed approach, grinding burn was successfully recognized, as the highest accuracy obtained by the models was 99.4% on the test dataset. This result is superior when considering those reported in the literature, in which conventional machine learning techniques are employed for grinding burn monitoring.