Predictive modelling of residual stress in turning of hard materials using radial basis function network enhanced with principal component analysis

This study proposes a hybrid machine learning (ML) model that combines a radial basis function network (RBFN) with principal component analysis (PCA) to predict residual stress (RS) in the machining process. Higher temperatures and plastic deformation can generate RS conditions in the machining of hard materials, significantly influencing the quality of machined parts, particularly their surface integrity. It is crucial to evaluate and guarantee machining conditions that ensure reliable surface integrity that yields compressive RS conditions. Incorrect parameter settings can lead to poor surface quality, resulting in tensile RS conditions that impact both the quality and lifetime of manufactured products. The methodology involved experimental machining trials for data acquisition, with conditions selected for hard material machining tests conducted using a computer numeric control (CNC) lathe under dry cutting conditions. The cutting force components were measured, and RS was evaluated using X-ray diffraction. A total of 60 trials were conducted. Data preprocessing, including normalization and removal of outliers, was executed before constructing the ML algorithm, leaving 59 valid tests for model training and testing. Separate datasets for training (70 % of the data) and testing (30 %) were randomly selected from the valid experimental tests. PCA enhanced the RBFN model’s generalization by reducing data dimensionality and providing the hidden units parameter. The eigenvectors obtained through PCA served as an efficient initial reference for RBF centroids, highlighting principal directions of variation. The PCA-RBFN model, using a multiquadric radial function, exhibited robust performance in capturing underlying patterns during training and demonstrated balanced results in the tests. In the PCA-RBFN model training stage, it exhibited a coefficient of determination (R2) of 76.66 %, a mean average error (MAE) of 72.15 MPa, and a root mean square error (RMSE) of 84.94 MPa. In the testing stage, the model showed increased metrics, indicating its ability to predict RS values accurately, with an R2 of 80.54 %, MAE of 51.25 MPa, and RMSE of 70.36 MPa. The evaluation between the training and testing stages highlighted the model’s coherence in performance, demonstrating its ability to generalize well to new data. A second comparative analysis was conducted with a conventional RBFN approach using stochastic method for centroid determination. The hybrid PCA-RBFN method exhibited lower computational effort, requiring only three iterations compared to 221,923 iterations in the conventional approach. Despite the conventional approach yielding better metrics in the training stage, it exhibited signs of overfitting, a common concern in machine learning models. The hybrid PCA-RBFN consistently outperformed it regarding generalization and fit during testing. The proposed methodology holds practical implications for the manufacturing industry, allowing the integration of a pre-trained predictive model into machine controllers or smart sensors. The hybrid PCA-RBFN’s computationally lightweight approach also supports incorporating real-time adaptive algorithms into production systems, enabling continuous real-time adjustments to the model factors and weights, thereby improving automated control.