Using machine learning and optimization for controlling surface roughness in grinding of St37

In the context of the industrial grinding process, the quality of products is often assessed by the final surface roughness, which is influenced by various parameters in the industrial environment. Previous studies lacked a feasible formulation based on a kinematical and statistical model to explain the uncertainty and non-linearity of grinding conditions, particularly concerning the cooling method, leading to significant discrepancies between the formulated and real results. This study introduces a novel strategy that combines deep learning and optimization to establish a suitable framework. It employs an artificial neural network to simulate and predict surface roughness, considering various dressing and cooling parameters in the industrial grinding of St37 steel alloy. Initially, an analysis of variance (ANOVA) is conducted to determine the correlation between input and output data. Subsequently, a neural network approach with one and two hidden layers, incorporating various activation functions, is employed. Therefore controlling and improving the accuracy of surface roughness predictions in industrial grinding processes can be automated. The mean squared error (MSE) metric is applied to each implementation to identify the best network architecture for the dataset. Upon selecting the network with the lowest MSE, the final algorithm predicts a set of randomly selected data from the dataset, achieving an overall accuracy of 80%. When compared to the accuracy of the formulated implementation, the neural network approach demonstrates a significantly higher accuracy of up to 30%, surpassing conventional analytical formulation in predicting final surface roughness. These results underscore the considerable potential and feasibility of deep learning approaches for industrial applications.