Physics-Informed Autoencoders With Intrinsic Differential Equations For Anomaly Detection In Industrial Processes

An autoencoder for anomaly detection is presented that is enhanced by a differential equation and a wavelet retransformation stage. The algorithm is applied to cool down data from steel coils that have four distinct classes. Using automatic differentiation, the differential equation is included in the loss of the encoder side of the autoencoder network. Training now lets the network converge to a solution where one of the latent neurons contains the solution of the differential equation. As result, the classes of the similar cooling curves are successfully separated by the prediction of the autoencoder. Using this clustering anomalous behaviour can be early detected in the process.