Symmetrized dot pattern enhanced cross-attention DDPM method for tool condition monitoring

In the current intelligent transformation of manufacturing, the tool condition monitoring (TCM) method based on deep learning can accurately analyze the complex signal characteristics presented by cutting tools at different wear stages. However, these methods require a large number of training samples to obtain significant results. In practical applications, samples are often unlabeled and difficult to obtain, resulting in a low model accuracy and insufficient generalization ability. To solve this problem, this study proposes a cross-attention diffusion model-enhanced recognition method that uses small samples. Combined with the symmetrized dot pattern (SDP), the cutting force signals of different tool wear conditions were first converted into SDP images. By improving the noise addition algorithm in the forward process of the denoising diffusion probabilistic model (DDPM) and adding the U-Net network with the cross-attention mechanism, high-quality pseudo-samples were generated using an improved DDPM model. A mixed sample set was created by combining the original and generated SDP image samples for sample augmentation and then inputted into ResNet18 to significantly improve the recognition accuracy. The experimental investigation of the milling TCM experiment demonstrated that the recognition accuracy of the proposed method for tool conditions can be achieved at 96.4% with small samples.