关键词: 刀具破损监测, 增量学习, 半监督学习, 不平衡数据, 工况

Abstract: Data-driven tool breakage detection methods suffer from the limited availability of anomaly samples and variable distribution of normal samples.A semi-supervised incremental learning method was proposed for the online detection of cutting tool breakages.A certain number of normal samples were acquired,and the low-dimensional representation features of the normal samples were learned by an autoencoder.Based on the distribution of the representation features,a breakage detection model was established and used for online detection.During breakage detection,the reconstruction errors of the trained autoencoder were used for concept drift detection of the observation samples.When concept drift was detected or a sample was falsely recognized,the model should implement incremental learning with the latest sample to retain ideal performance.The proposed method was applied to an experimental cutting tool breakage dataset and compared with the standard autoencoder method,one-class support vector machine method and local outlier factor method.The experimental results demonstrated that the proposed method significantly improved the recognition performance on samples acquired under different conditions.

Key words: tool breakage detection, incremental learning, semi-supervised learning, imbalanced data, operation conditions