Hybrid Data Augmentation Combining Screening-Based MCGAN and Manual Transformation for Few-Shot Tool Wear State Recognition

Deep learning has been widely applied in fault diagnosis and monitoring, but obtaining labeled data under abnormal conditions is challenging. This limitation makes the best current deep learning methods seem powerless as they require a large amount of labeled data for training. When the training dataset is small, overfitting is highly likely to occur, leading to performance degradation of deep neural networks. To address this issue, this article proposes a hybrid data augmentation mechanism (HDAM) that utilizes a multicategory generative adversarial network (MCGAN) model and similarity-based selection criteria to generate high-quality data in few-shot scenarios. The selected samples are mixed with augmented samples that undergo image flipping, rotation, and noise addition to form the training set, enhancing the generalization capability of the recognition model after training. To validate the effectiveness of the proposed method, experiments are conducted on the PHM2010 dataset and the TC4 titanium alloy dataset. The experimental results demonstrate a significant improvement in the recognition accuracy of tool wear states in few-shot scenarios.