Multi-mode frequency response prediction of milling robot based on feature transferring with small sample sets

Industrial robots are increasingly used in machining due to their cost-effectiveness and larger work envelopes. However, their relatively low structural stiffness makes them vulnerable to machining chatter, which negatively impacts both process stability and surface quality. Accurate prediction of the multi-mode frequency response function (FRF) of robotic milling systems is crucial to ensure process stability. Traditional FRF prediction approaches, however, often require extensive experimental procedures, are complex, and are time-consuming. To address these challenges, this study proposes an innovative feature-transfer-based method for multi-mode FRF prediction in milling robots, requiring only a minimal set of impact tests. The method organizes measured FRFs into second-order complex tensors, facilitating the transfer of features between different postures. Multi-mode parameters of the tool-tip FRF under the source posture are extracted using the least-squares complex exponential (LSCE) method and assembled into a label vector. A complex-kernel extreme learning machine with augmented inputs (CKELM-AI) is then trained to predict the tool-tip FRF under the target posture. Additionally, a virtual sample generation strategy based on CKELM-AI and feature augmentation, including statistical, frequency, and time-frequency features, is applied to enhance prediction accuracy. Experimental validation on a milling robot demonstrates that the proposed method significantly improves both prediction efficiency and accuracy, establishing a new, more efficient approach for predicting multi-mode FRFs without the need for extensive testing.