Physics-Informed Autoencoders with Intrinsic Differential Equations for Anomaly Detection in Industrial Processes