Thermal damage diagnosis in micro grinding processes based on raw acoustic emission signals and convolution neural networks

A micro-grinding process is usually the last machining operation among the various processes employed in the manufacturing of metallic components. It is generally the first option when the manufactured products require a combination of low surface roughness values and narrow dimensional tolerances. However, one of the main challenges of a micro-grinding process is avoiding high specific energy and the corresponding intense heat generation, as this can damage the workpiece. The energy consumed during micro-grinding is almost entirely converted into heat, making the component susceptible to thermal damage. As the occurrence of this damage compromises the service life of manufactured components, the implementation of monitoring systems is essential to guarantee the quality of the manufactured products improving process efficiency. In this study, a new approach based on convolution neural networks (CNNs) is proposed to predict the occurrence of thermal damage in the micro-grinding process. This approach uses raw acoustic emission (AE) signals as inputs into CNN, which is a more technically feasible approach to performing real-time monitoring. To obtain the AE data, micro-grinding experiments were conducted on N2711 grade steel (which is susceptible to thermal damage) under different cutting conditions (a_e = 5–50 µm). The obtained AE signals were labeled based on the offline analyses performed on the workpieces, which identified the occurrence of thermal damage through visual inspections, microhardness, and microstructural analysis. For the conditions employed in this work, thermal damage was obtained after grinding with depth of cut values > 10 µm. Total 3960 AE signals were obtained of which 2200 of them were associated with thermal damage. The results depict that the proposed CNN model was able to successfully classify normal and thermal damage AE signals reaching an accuracy of 98.6% over the dataset (0.4% false positives and 1.0% false negatives).