Data-Driven Prediction of Cutting Rate, Residual Stress and Micro-hardness in Wire-Cut EDM of Titanium 6Al-4V ELI Alloy Using an Artificial Neural Network
Main Article Content
Abstract
This paper reports a multi-response investigation and predictive modelling of the wire-cut electrical discharge machining (WEDM) of the biomedical titanium alloy Ti-6Al-4V ELI (Grade 23). Three surface-integrity responses, namely cutting rate, residual stress, and micro-hardness, were studied in relation to six controllable machining variables: flushing pressure, peak current, duty cycle, gap voltage, wire feed rate, and wire tension. A Taguchi L18 (2^1 × 3^5) orthogonal array was adopted as the design of experiments, and regression models relating the inputs to each response were developed in MINITAB-19. Analysis of variance (ANOVA) revealed that peak current, duty cycle, gap voltage and wire tension were the dominant factors, whereas flushing pressure and wire feed rate were comparatively insignificant. Both the cutting rate and the residual stress increased with rising peak current and duty cycle, while the mean micro-hardness of the machined surface was of the order of 208 HV. Higher flushing pressure and feed rate increased the cutting rate but reduced the residual stress (mean of about 212 MPa). The combined action of a high peak current and a high duty cycle was found to degrade the surface texture and to promote surface debris and micro-cracks, thereby elevating the residual stress. The study establishes a compact set of regression models and an optimal parameter window that can guide the WEDM of Ti-6Al-4V ELI for orthopaedic and aerospace applications. Wire-cut electrical discharge machining (WEDM) of the biomedical titanium alloy Ti-6Al-4V ELI (Grade 23) is governed by several interacting electrical and mechanical parameters, making analytical prediction of the machined surface quality difficult. This study develops a data-driven model, based on a feed-forward artificial neural network (ANN), to predict three surfaceintegrity responses, namely the cutting rate, the residual stress and the micro-hardness, from six controllable process variables: flushing pressure, peak current, duty cycle, gap voltage, wire feed rate and wire tension. A Taguchi L18 design furnished eighteen machining conditions; with three replicate measurements per condition, a data set of fifty-four observations was assembled. A separate 6-10-8-1 network was trained for each response using the Levenberg-style Adam optimiser on min-max normalised data, and its performance was benchmarked against ordinary least-squares linear regression. On the held-out test data, the ANN attained coefficients of determination of 0.70, 0.93, and 0.97 for the cutting rate, residual stress, and