Akdulum, Aslan; Siyambaş, Yusuf; Çakıroğlu, Ramazan; Uzun, Gültekin

This study develops a machine learning-based hybrid optimization framework to predict energy consumption from cutting temperature in turning of Ti6Al4V alloy under different cooling environments. Due to the poor thermal conductivity of titanium alloys, heat accumulation in the cutting zone increases energy consumption and reduces process efficiency. Cutting temperatures were recorded in real time using a thermal camera, and energy consumption was measured with a power analyzer. Statistical features extracted from temperature data served as inputs for energy prediction. A hybrid optimization framework was developed by integrating feature selection methods, machine learning algorithms, and hyperparameter optimizers. Among six algorithms evaluated, CatBoost combined with SHAP-based feature selection and Optuna hyperparameter optimization achieved the best performance. The proposed hybrid model (CAT+SHAP+Optuna) provides a 91% improvement in Mean Squared Error compared to the benchmark model, demonstrating high accuracy in energy consumption estimation. This approach offers significant potential for enhancing energy efficiency and promoting sustainable manufacturing in the machining of difficult-to-cut materials.

Siyambaş, Yusuf

Nickel-based superalloys are commonly employed in the aerospace and defense industries because of their high temperature resistance and exceptional mechanical qualities. However, poor heat conductivity and a high tendency for work hardening make these materials difficult to process, leading to reduced surface quality and increased energy consumption. This study aims to predict surface roughness (Ra) and carbon emissions (CE) of Inconel 718 utilizing machine learning methods based on electrical signals obtained from a power analyzer during turning. Experiments were conducted under dry conditions at three cutting speeds, three feed rates, three depths of cut, and two cutting lengths. Forty-seven statistical features and truncation parameters extracted from current (Irms), total harmonic distortion (Ithd), active power (AP), and fundamental active power (FAP) signals were used as input to the model. Total CE has been calculated, taking into account energy consumption, cutting inserts, materials, and chip-related components. RF, SVR, XGBoost, and MLP algorithms were compared with SHAP, RFE, PFI, and PCC-based feature selection methods. The results showed that progress in Ra estimation was the most dominant parameter and that FAP and AP-based features had high representational power. With the XGBoost–RFE combination, a MAPE of 5.82% and an R2 of 0.9728 were obtained for Ra. In CE estimation, cutting length and depth of cut were the key factors; XGBoost–PCC achieved a MAPE of 0.0367% and an R2 of 0.9999. The findings reveal that power analyzer signals can be effectively used in the simultaneous estimation of quality and sustainability indicators.

Siyambaş, Yusuf; Akdulum, Aslan

Despite their superior mechanical properties, titanium alloys cause excessive temperatures in turning operations due to their low thermal conductivity. This situation reduces workpiece quality and shortens tool life. Therefore, it is necessary to determine the type of coolant that will provide the desired product quality at the optimum cost. Surface roughness (Ra) is one of the most important indicators of product quality. The goal of this work is to accurately predict the Ra in various cooling conditions prior to turning the Ti6Al4V alloy. Models built using various machine learning algorithms were used to compare the cutting parameters and features taken from sensor signals. While the lowest Ra values were obtained in dry cutting, the highest Ra values were obtained with vortex CO2. The highest performance success was achieved with the XGBoost algorithm. The model created by enhancing the cutting parameters performed better in predictions than the models created using statistical information taken from the sensor signals. The suggested approach eliminates the need for extra sensors and allows for the low computational cost and high success rate estimation of Ra.