Data-driven unsupervised anomaly detection of manufacturing processes with multi-scale prototype augmentation and multi-sensor data

Accurate anomaly detection (AD) of machine tools is crucial to ensure the quality and efficiency of the manufacturing processes. Due to the lack of tool anomaly information, it is difficult for AD model to precisely capture the distribution of health states and then obtain a discriminative decision boundary. Current methods try to reconstruct the normal data distribution without restricting the abnormal, resulting in the unacceptable overlap between normal and abnormal regions and finally leading to high false alarm rate. To tackle these issues, a hierarchical augmented autoencoder is proposed for AD of machine tools during manufacturing. First, a skip-connected autoencoder is built to basically learn the normal representations of multi-sensor data in an unsupervised manner. Then, to improve further emphasis the reconstruction on normality and suppress that on anomalies, we propose hierarchical memory modules to store multi-scale normal prototypical patterns, using them as a prior to guide the reconstruction with preference. Finally, A compound metric loss function is designed to measure data similarity considering both distance and angle perspectives, which can restrain noise interference and enhance model robustness. Extensive experiments are conducted on real-world CNC machine tool datasets, the proposed method achieves better performance for unsupervised AD compared with other typical methods.