A method for detecting anomalies in a physical system generates from a set of physics rules and a process graph representing the system a set of candidate physics models that assign physics rules to portions of the process graph representing sensors. Candidate physics models are rejected if an error between the models and sensor data exceed a predetermined error tolerance. Supervised learning is used to train a machine learning model to predict an error between the physics models and the sensor data. The predicted error and predicted sensor measurements from the physics models are then used to detect anomalies using unsupervised learning on a distribution of error between the predicted sensor measurements and the sensor data.