Tackling data scarcity in machine learning-based CFRP drilling performance prediction through a broad learning system with virtual sample generation (BLS-VSG)

Abstract

Machine learning (ML)-based data-driven method has emerged as a powerful tool for predicting the manufacturing performance of carbon fibre reinforced plastic (CFRP), particularly in CFRP machining, where physics-based models are computationally expensive. However, the effectiveness of ML models are often constrained by limited datasets, due to the high cost and time required for experimental data acquisition. To address this, this paper presents the first study to apply virtual sample generation (VSG) techniques to enlarge the training dataset and mitigate data scarcity in the prediction of CFRP drilling performance. A novel hybrid ML framework integrating Broad Learning System (BLS) and VSG (BLS-VSG) is proposed to combine the capability of BLS in small dataset prediction with the enlarged dataset generated by VSG. The model has been employed to predict the drilling thrust force and delamination damage under various drilling conditions (spindle speed, feed rate, point angle). Three different VSG methods (SMOTE, MD-MTD and CVT) and the number of virtual samples were evaluated in detail. Results show that VSG can effectively enlarge the training dataset and improve the prediction performance of the ML model. Specifically, VSG reduced the mean square error (MSE) and mean absolute percentage error (MAPE) for thrust force prediction by 39.0 % and 12.9 %, respectively, compared to the benchmark without VSG. For delamination factor F_da prediction, MSE and MAPE were reduced by 22.6 % and 16.5 %, respectively. The proposed BLS-VSG model outperforms other conventional ML models (BPNN, ELM, SVR and RT) for both scenarios (with/without VSG), providing a robust and data-efficient solution for CFRP drilling performance prediction.