Data-Driven Sensor Fault Diagnosis Based on Nonlinear Additive Models and Local Fault Sensitivity

Linear dependence on variables is a common assumption in most fault diagnosis systems for which a large number of powerful methodologies has been developed over the years. In case the system nonlinearities are relevant fault diagnosis tools relying on the linear assumption may provide unsatisfactory results in terms of false alarms and missed detections. In this paper, non-linear Additive Models are proposed to characterize non-linear redundancy relations existing between the system signals. These relations have been identified directly from data using the Multivariate Adaptive Regression Splines (MARS) algorithm. Then, the non-linear redundancy relations have been linearized in order to derive a local (time-dependent) fault signature matrix. The fault signature matrix is then used within a directional residual-based Fault Isolation procedure to isolate additive sensor failures. A quantitative analysis has been performed exploiting real multi-flight data of a semi-autonomous aircraft, making a detailed comparison with a state-of-the-art Fault Isolation method based on linear redundancy relations.