From Data to Insight: Neural Networks to Predict Grinding Residual Stresses

Surface integrity, particularly the residual stresses induced by finishing operations such as grinding, plays a fundamental role in the performance of mechanical components, especially gears.However, measuring and controlling these stresses directly on the shop floor remains a major challenge, particularly in high-productivity environments.This scenario highlights the need for innovative and datadriven solutions that can be implemented in real time.This work presents a neural network-based approach to predict residual stresses resulting from the gear profile grinding process.To support the model, an experimental database with 240 grinding conditions, covering a wide range of depth of cut, feed rate, and cutting speed combinations, was build to comprehensively represent process boundaries.Online acquisition of process force and spindle current was conducted without compromising the workpiece, supporting the feasibility of in-situ implementation.Two top-performing models were identified, each achieving a Mean Absolute Error below 30 MPa.This represents only 3.8% of the average residual stress and is even lower than the experimental uncertainty of 38.3 MPa.Additionally, only 25% of the individual predictions exceeded this uncertainty, and even then, the maximum deviation was around 10% of the mean stress level.These results demonstrate the accuracy and robustness of the proposed approach, showing great potential for its integration into intelligent, data-driven manufacturing systems focused on process optimization and quality control.X.