Anomaly detection for high-speed machining using hybrid regularized support vector data description

Process monitoring in high-speed machining (HSM) is essential to guarantee product quality and improve manufacturing efficiency. Nevertheless, the data acquired from practical machining processes are completely unlabeled and severely unbalanced, which may be seriously insufficient to support deep learning-based anomaly detection. Furthermore, the collected signals are inevitably contaminated by environmental noises and uncertain factors. How to remove these disturbances according to data distribution characteristics remains a challenging issue. To tackle these limitations, a novel interpretable machine learning approach, called hybrid regularized support vector data description (H-SVDD), is proposed for unsupervised anomaly detection during HSM. In this work, an adaptive local kernel density estimate is first constructed to eliminate outlier interferences, and assigns interpretable weights to optimize the SVDD for improving detection accuracy. Subsequently, by introducing the l_p-norm penalty mechanism, a generalized probability density regularized SVDD is innovatively designed to enhance the descriptive capability for complex machining processes. Finally, a hyperparameter tuning strategy based on Bayesian optimization is developed to improve generalizability and stability. The data collected from CNC machines are used to verify the superiority of the proposed method. Experimental results show that the proposed H-SVDD has higher detection accuracy than current SVDD methods and eliminates false alarms caused by noise interferences. This work may provide a useful solution for independently perceiving the health conditions of HSM.