Contextual Anomaly Detection in Hot Forming Production Line using PINN Architecture

This paper presents a physics-informed neural network (PINN) architecture for contextual anomaly detection in a hot forming production line. It enhances widely used proximity- or distribution-based anomaly detection approaches for industrial processes through the consideration of contextual process data. The physical model is built using a priori process knowledge and thermodynamic equations. This model is then injected into the loss function of a neural network. The network is trained on data from the production line and constantly regularized by the physical loss term. Within inference, the PINN predicts the resulting temperature of the produced blank given the contextual process data. The anomaly detection is performed using the unsupervised local outlier factor algorithm on the error between actual and predicted blank temperature. This makes it possible to assess whether the achieved product temperature appears normal or abnormal based on the database. The main advantage of this novel approach is that it can detect contextual anomalies that remain otherwise undiscovered.