Sample Augmentation for Intelligent Milling Tool Wear Condition Monitoring Using Numerical Simulation and Generative Adversarial Network

Recent advances in artificial intelligence (AI) technology have led to increasing interest in the development of AI-based tool condition monitoring (TCM) methods. However, achieving good performance using these methods relies heavily on large training samples, which are both expensive and difficult to obtain in practical TCM applications. This article addresses this issue by employing a much smaller training sample composed of a non-exhaustive sampling of experimentally measured cutting force signals in conjunction with a novel data augmentation method that combines numerical simulation with a generative adversarial network (GAN). First, cutting force signal samples not present in the experimental dataset are obtained by numerical simulation using a finite element method simulated based on the Johnson-Cook model. Second, the GAN is employed to synthesize additional samples that are similar to both the simulated samples and the experimentally measured samples. The synthesized samples are combined with the measured and simulated samples to produce an appropriately large dataset necessary for the effective training of an AI classifier. The proposed sample augmentation method is applied in milling TCM experiments, and the classification accuracies obtained with several AI classifiers trained with the augmented dataset were all close to or equal to 100%.

Automated summary

This article discusses a method of augmenting the training dataset for AI classifiers by combining numerical simulation and generative adversarial networks, addressing the challenge of obtaining large training samples in TCM applications. Through this approach, experiments show that the classification accuracies of several AI classifiers trained on augmented datasets are close to or equal to 100%.