A Dynamic Printing Equipment Production Status Anomaly Detection Model

In the digital transformation of the printing industry, monitoring the equipment status during the printing production process is the key task. Researchers establish equipment anomaly detection models on historical and real-time data collected from sensors. However, in the actual production process, the data from equipment lacks clear status labels. Manually labeling the status of each time point is a challenge, it makes supervised learning methods difficult to apply. Unsupervised methods are proposed to automatically detect anomalies. But current methods are sensitive to noise and highly depend on data quality. In the realistic scenario, the timestamp intervals are uneven, and the insufficient sensor precision may produce uncorrected speed data. In response to the above issues, this article proposes a new unsupervised dynamic printing equipment production status anomaly detection model, which combined autoencoder and Gaussian model. The model, considering complexity in real scenes, is composed of three components: window feature extraction, anomaly scores measurement and confidence calculation. Compared with traditional anomaly detection using Kmeans with fixed window features, our model improves the classification accuracy of normal samples in real printed data from 90% to 93%.