Kalman Supervised Network for Improved Model Predictions

In this paper, we propose a Neural Network to improve long-term state predictions without measurements based on Kalman filter observations. It is well known that the Kalman Filter is an iterative algorithm composed of two phases: predict and update. The update corrects predictions based on measurements. Predictions rely exclusively on the embedded physical model. This research aims to learn the underlying dynamics of the system under observation from the estimates of a standard Kalman Filter that supervises a Neural Network. Then, the Kalman Supervised Net (KSN) can be used to improve predictions learning from Kalman filter corrections. Numerical results show the advantages of the proposed solution when predicting the state of a springmass-damper system without using acceleration measurements.