Structured neural networks use relationships between manufacturing events to guide model architecture before training. This approach supports supervised and unsupervised anomaly detection while reducing unnecessary connections and, in reported experiments, improving detection performance.