Multi-view contrastive learning framework for tool wear detection with insufficient annotated data

The tool is an important part of CNC machine tools, and its wear degree affects the operating efficiency of CNC machine tools. Predicting tool wear can be a challenging task when only unlabeled samples or a limited number of labeled samples are available. Supervised learning is not able to handle this type of data, and in this case, unsupervised learning frameworks have shown competitive performance. When there are only a few labeled samples, the unsupervised learning framework can be used as a flexible and effective method to explore the latent characteristics of the data. In this paper, a novel self -supervised contrastive learning framework is proposed for tool wear prediction with insufficient annotated data. The framework consists of two steps, namely the pre -training phase and the fine-tuning phase. During the pre -training phase, multiple data augmentation techniques are first employed to provide multiple augmented views of the original input data. Then, a multi -scale CNN -based model is utilized as the feature extractor to extract multi -level features from these multi -views. Furthermore, a temporal representation module and a distribution representation module are utilized to learn invariant representations from the multi -view data. The KL divergence evaluation function is employed to guide the learning process and ensure the acquisition of meaningful representations. During the fine-tuning stage, we use a small amount of annotated data to fine-tune the parameters of the classifier, thereby achieving precise tool wear recognition tasks. The proposed method is validated through comparative and ablation experiments conducted on one publicly available dataset and one self -collected dataset. The results clearly demonstrate the superiority of the proposed method, achieving highly precise fault diagnosis even when faced with limited annotated data.