System identification and tracking using a statistical model and a Kalman filter

The sensitivity of system identification estimates to changing environmental and operational parameters is important for structural health monitoring and inverse force estimation. Damage to a structure can be misidentified or masked by modal shifts as a result of environmental parameters. In this paper a novel approach to reduce the uncertainties and improve the sensitivity of system identification and tracking is presented based on a data driven statistical model and a Kalman filter. A key objective is to make experimental data maximally informative by using additional system inputs. The method is first demonstrated on numerical data where it was found to improve accuracy and identify underlying trends. Investigations were then conducted on full scale data from the research vessel Polarstern. Model training led to the development of a sliding predictive model using an optimized linear regression method. The model was found to accurately re-create the training data set and was used to make predictions based on future system inputs. Since both the model prediction and the system identification estimates contain different uncertainties the Kalman filter was used to combine both estimates in an optimal way. The Kalman filter estimates were observed to produce balanced and consistent results. The Kalman estimates were also not overly or consistently biased by the SSI estimates or the model predictions