Hybrid intelligent system for acoustic-based surface roughness prediction in milling using group-stratified ensemble learning and CNN regressors

Abstract In this study, an acoustic signal-based intelligent framework is proposed for noncontact surface roughness prediction in milling. This framework integrates both classical machine learning and deep learning approaches. The study utilizes a milling acoustic dataset containing 7448 recordings collected under varying feed rates, spindle speeds, and depths of cut. Each recording is paired with an experimentally measured surface roughness value (Ra). Handcrafted acoustic descriptors, including Mel-frequency cepstral coefficients (MFCCs), zero-crossing rate, spectral centroid, spectral bandwidth, and root mean square energy, are extracted from the milling audio signals. Four machine learning regressors, namely random forest, extra trees (ETs), support vector regression, and K-nearest neighbours (KNN) are evaluated. The ET demonstrates the highest predictive performance, achieving an R 2 of 0.9329, mean absolute error of 0.1076 μ m, and root mean squared error of 0.2703 μ m. Feature importance analysis reveals that higher-order MFCC descriptors, particularly MFCC_Mean_11, are the most discriminative predictors. To further assess the robustness of the ensemble framework, GroupKFold cross-validation is implemented, with the ETs model achieving a mean R 2 of 0.9672. In parallel, a two-dimensional convolutional neural network (CNN) is trained on Mel spectrogram images, achieving a competitive R 2 of 0.9279. Comparative analysis indicates that the CNN outperforms several classical regressors, demonstrating the strong representational capacity of Mel spectrogram features for acoustic roughness characterization. Collectively, these findings confirm that acoustic signals contain sufficient spectral information for reliable, real-time surface roughness estimation, establishing a strong foundation for intelligent quality monitoring in smart manufacturing and Industry 4.0 environments.