Graph-structured data generation and analysis for anomaly detection in an automated manufacturing process

During automated manufacturing processes, multiple sensors are attached to facilities to collect and analyze analog data for detecting operational anomalies. However, owing to facility devices being interlinked by a control system, simultaneous examination of the control system and analog data enhances the accuracy of anomaly detection and diagnosis of root causes. We proposed a system detecting anomalies by integrating an internal control system with external analog data and representing it in a graph structure. The system generates and combines the adjacency and feature matrices for training a convolutional autoencoder model to identify operational anomalies. Performance tests revealed distinct operational patterns in the cycle data flagged by the model as anomalies. The system diagnosed the root cause of anomalies, such as control operation sequencing, timing variances, and shifts in analog or video signals. This approach may enhance the productivity and quality of the manufacturing processes by facilitating anomaly detection and cause diagnosis.