Prediction of CNC machining vibration signals based on conditional generative adversarial network

Vibration in high-speed, high-precision CNC machining significantly impacts machining quality and efficiency. While existing condition monitoring techniques can detect process abnormalities, they suffer from response lag. Physics-based modeling approaches are constrained by deterministic prediction bias, and low computational efficiency, hindering real-time, accurate vibration state prediction. To address these limitations, this paper proposes a generative modeling method for cutting vibration based on a Conditional Generative Adversarial Network. It establishes an end-to-end, non-deterministic mapping from machining parameters to vibration signals. And to address the complex heterogeneity of processing parameters as input conditions, the feature-wise linear modulation method is introduced to fuse conditional inputs. Through this integration, the trained model can directly generate high-fidelity signals for new target parameters without requiring additional physical trials. To enable model development, an experimental platform was constructed to acquire representative vibration data, providing dedicated training/evaluation datasets. Finally, 5-fold cross-validation demonstrates the model’s performance: the average Euclidean distance between generated and actual signals is as low as 1.932, and the average Pearson correlation coefficient reaches 0.944. These results robustly confirm high accuracy and strong generalizability. This study presents a novel data-driven paradigm for the pre-process prediction and active suppression of CNC machining vibration, supporting process parameter optimization and machining state assessment.