Abstract
Effectively identifying and preventing anomalies in the melt process significantly enhances production efficiency and product quality in industrial manufacturing. Consequently, this paper proposes a study on a melt anomaly identification system for pelletizers using autoencoder technology. It discusses the challenges of detecting anomalies in the melt extrusion process of polyester pelletizers, focusing on the limitations of manual monitoring and traditional image detection methods. This paper proposes a system based on autoencoders that demonstrates effectiveness in detecting and differentiating various melt anomaly states through deep learning. By randomly altering the brightness and rotation angle of images in each training round, the training samples are augmented, thereby enhancing the system's robustness against changes in environmental light intensity. Experimental results indicate that the system proposed has good melt anomaly detection efficiency and generalization performance and has effectively differentiated degrees of melt anomalies. This study emphasizes the potential of autoencoders in industrial applications and suggests directions for future research.