Development of a Surface Roughness Prediction & Optimization Framework for CNC Turning

Abstract

Computer numerical control (CNC) machining is an integral element to the manufacturing industry for production of components with requirements to meet several outcome conditions. The surface roughness (Ra) of a workpiece is one of the most important outcomes in finish machining processes due to it’s direct impact on the functionality and lifespan of components in their intended applications. Several factors contribute to the creation of Ra in machining including, but not limited to, the machining parameters, properties of the workpiece, tool geometry and wear. Alternative to traditional selection of machining parameters using existing standards and/or expert knowledge, current studies in literature have examined methods to consider these factors for prediction and optimization of machining parameters to minimize Ra. These methods span many approaches including theoretical modelling and simulation, design of experiments, statistical and machine learning methods. Despite the abundance of research in this area, challenges remain regarding the generalizability of models for multiple machining conditions, and lengthy training requirements of methods based solely on machine learning methods. Furthermore, many machine learning methods focus on static cutting parameters rather than consideration of properties of the tool and workpiece, and dynamic factors such as tool wear. The main contribution of this research was to develop a prediction and optimization model framework to minimize Ra for finish turning that combines theoretical and machine learning methods, and can be practically utilized by CNC machine operators for parameter v decision making. The presented research work was divided into four distinct objectives. The first objective of this research focused on analyzing the relationship between the machining parameters and Ra for three different materials with varying properties (AISI 4340, AISI 316, and CGI 450). This was followed by the second objective that targeted the development of an Ra prediction framework that utilized a kinematics-based prediction model with an ensemble gradient boosted regression tree (GBRT) to create a multi-material model with justified results, while strengthening accuracy with the machine learning component. The results demonstrated the multi-material model was able to provide predictions with a root-mean-square error (RMSE) of 0.166 μm and attained 70% of testing predictions to fall within limits set by the ASME B46.1-2019 standard. This standard was utilized as an efficient evaluation tool for determining if the prediction accuracy was within an acceptable range. The remaining objectives of this research focused on investigating the relationship between tool wear and Ra through a focused study on AISI 316, followed by application of the prediction model framework as the fitness function for testing of three different metaheuristic optimization algorithms to minimize Ra. The results revealed a significant relationship between tool wear and Ra, which enabled improvement in the prediction framework through the use of the tool’s total cutting distance for an indicator of tool wear as an input into the prediction model. Significant prediction improvement was achieved, demonstrated by metrics including RMSE of 0.108 μm and 87% of predictions were within the ASME B46.1-2019 limits. The improved prediction model was used as the fitness function for comparison performance of genetic algorithm (GA), particle swarm vi optimization (PSO), and simulated annealing (SA), under constrained and unconstrained conditions. SA demonstrated superior performance with less than 5% error between the optimal and experimental Ra when constrained to the experimental data set during validation testing. The overall results of this research establish the feasibility of a framework that could be applied in an industrial setting for both prediction of Ra for multiple materials, and supports the determination of parameters for minimizing Ra considering the dynamic nature of tool wear.