LaplaceDiag uses actuator effects for fault diagnosis

The method relies on LaplaceNet, a neural network that operates in Laplace space to learn a healthy plant model. It separates nonlinear prediction based on measured history from an explicit matrix describing how actuator values affect sensor values. By keeping the history predictor fixed, current actuator effects can be estimated via linear regression without retraining the neural network. This allows the system to compare current states against a healthy reference, accounting for estimation uncertainty and normal variation to detect changes and generate diagnostic hypotheses.

LaplaceDiag supports both online and offline diagnosis, serving as an intermediate step between general anomaly detection and specific component identification. The authors report successful evaluations of the method on two datasets, demonstrating its potential to narrow the search for faulty equipment by leveraging the specific nature of actuator impacts.