An imbalanced data learning method for tool breakage detection based on generative adversarial networks

Tool breakage in manufacturing procedures can damage machined surfaces and machine tools. It is crucial to detect tool breakage in time and promptly respond to it. Due to the safety restrictions imposed in production, failure samples are significantly scarcer than normal samples, and this disequilibrium results in difficulty of failure detection. Therefore, we propose a new imbalanced data learning method for tool breakage detection. The key strategy is to balance the data distribution by producing valuable artificial samples for the minority class using an adversarial generative oversampling model based on a generative adversarial network (GAN). Unlike previous studies using GAN, we use the discriminator to screen samples generated by the generator and achieve effective oversampling. Multiple classifiers are adopted as the decision-making models to perform tool breakage detection. The proposed method is applied to a set of imbalanced experimental tool breakage data collected in a workshop. Compared with the best results of other oversampling solutions, the proposed method improves the breakage detection rate from 93.6% to 100%, which shows its practicability and validity. Additionally, evaluations are performed based on 12 imbalanced benchmark datasets. The results further substantiate the superiority of the proposed method to existing sampling methods.