Multi-objective optimization based on machine learning and non-dominated sorting genetic algorithm for surface roughness and tool wear in Ti6Al4V turning

Abstract

Titanium alloys are notoriously difficult to machine. They are used in the manufacture of various types of lightweight components. It is therefore important to improve their machinability and thus achieve sustainability in machining such alloys, by selecting appropriate influential factors: cutting parameters, tool material, geometric form, coolant types, and hybrid machining methods, to deliver efficient output. Nowadays, meta-heuristic algorithms effectively solve multi-objective optimization in machining problems instead of single-objective one. Along with that, the mathematical predictive models used for single-objective optimization are gradually being replaced by machine learning algorithms, which are highly robust and efficient in terms of prediction performance. Therefore, this work addresses the prediction and optimization of average surface roughness (Ra) and tool wear (VB) in Ti6Al4V alloy turning, using a WC tool coated by chemical vapor deposition (CVD) and physical vapor deposition (PVD), with dry machining. We apply a two-pronged approach combining machine learning (ML) and Non-Dominated Sorting Genetic Algorithm (NSGA-II), to model and optimize Ra and VB. The four ML models – Linear Regression (LIN), Support Vector Machine Regression (SVR), Extreme Gradient Boosting (XGB), and Artificial Neural Network (ANN) – are used to predict Ra and VB. The input variables of the turning process – feed rate, depth of cut, cutting speed, cutting time, and tool materials – are the major factors affecting surface quality and tool wear. By the error metrics such as root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R2), ANN is found to offer the best predictive performance. An ML and NSGA-II-based approach is then employed for multi-objective optimization of cutting parameters in Ti6Al4V turning. Fifty Pareto solutions are identified in the range of Ra and VB between (1.332 and 1.441 µm) and (0.100 and 0.125 mm), respectively. In this work, the Pareto solutions are selected based on their ranked performances. This aligns with the decision criterion employed to select the most robust cutting parameters. The definitive optimal Ra and VB are obtained by formulating a robust decisive multi-criterion function which integrates performance, preferred decision criterion, and trustworthiness. Finally, this produces the optimal solution for Ra and VB − 1.439 µm and 0.100 mm, respectively. Experimental validation confirms that the final optimum solution is within the acceptable range.