Abstract
<title>Abstract</title> <p>In smart manufacturing, accurate tool wear prediction is often limited by data deficiency and cross-condition adaptation. Conventional data-driven generative models, such as GANs, synthesize high-frequency dynamic signals without physical guidance. This may cause frequency shifts and negative transfer in downstream prediction tasks. To address this issue, this study proposes a cutting bending moment generation framework that integrates a mechanistic cutting bending moment model with a diffusion model. Different from conventional physics-informed or hybrid ML frameworks that usually impose physical equations through loss functions or simulation-assisted data generation, the proposed framework uses the ideal cutting bending moment as a physical input condition and injects the tool wear state through Feature-wise Linear Modulation (FiLM) during diffusion-based denoising. The proposed method uses a 1D-CNN U-Net as the denoising backbone. This design allows the model to generate cutting bending moment signals that preserve the main physical periodicity and wear-related signal variations. Cutting experiment results show that, using only 5% of real training data, the proposed method improves the coverage of wear-related features in the training set. In cross-condition tool wear prediction, the proposed method reduces the Root Mean Square Error (RMSE) by 60.45% compared with the baseline model trained only with real data. These results show that the proposed bending moment generation framework can support tool wear prediction under unseen cutting conditions and reduce the need for additional full-life tool wear experiments for tools with the same number of flutes.</p>