A Generative AI approach to improve in-situ vision tool wear monitoring with scarce data

Most aerospace turbine casings are mechanised using a vertical lathe. This paper presents a tool wear monitoring system using computer vision that analyses tool inserts once that the machining process has been completed. By installing a camera in the robot magazine room and a tool cleaning device to remove chips and cooling residuals, a neat tool image can be acquired. A subsequent Deep Learning (DL) model classifies the tool as acceptable or not, avoiding the drawbacks of alternative computer vision algorithms based on edges, dedicated features etc. Such model was trained with a significantly reduced number of images, in order to minimise the costly process to acquire and classify images during production. This could be achieved by introducing a special lens and some generative Artificial Intelligence (AI) models. This paper proposes two novel architectures: SCWGAN-GP, Scalable Condition Wasserstein Generative Adversarial Network (WGAN) with Gradient Penalty, and Focal Stable Diffusion (FSD) model, with class injection and dedicated loss function, to artificially increase the number of images to train the DL model. In addition, a K|Lens special optics was used to get multiple views of the vertical lathe inserts as a means of further increase data augmentation by hardware with a reduced number of production samples. Given an initial dataset, the classification accuracy was increased from 80.0 % up to 96.0 % using the FSD model. We also found that using as low as 100 real images, our methodology can achieve up to 93.3 % accuracy. Using only 100 original images for each insert type and wear condition results in 93.3 % accuracy and up to 94.6 % if 200 images are used. This accuracy is considered to be within human inspector uncertainty for this use-case. Fine-tuning the FSD model, with nearly 1 billion training parameters, showed superior performance compared to the SCWGAN-GP model, with only 80 million parameters, besides of requiring less training samples to produced higher quality output images. Furthermore, the visualization of the output activation mapping confirms that the model takes a decision on the right image features. Time to create the dataset was reduced from 3 months to 2 days using generative AI. So our approach enables to create industrial dataset with minimum effort and significant time speed-up compared with the conventional approach of acquiring a large number of images that DL models usually requires to avoid over-fitting. Despite the good results, this methodology is only applicable to relatively simple cases, such as our inserts where the images are not complex.