System Identification of continuous-time systems with quantized output data using indirect inference
Examiner: Cristian Rojas
Time: Thu 2021-05-06 14.00 - 15.00
Respondent: Frida Persson , DCS
Opponent: Linda Bui
Supervisor: Rodrigo Gonzalez Vidal
Continuous-time system identification is an important subject with applications within many fields. Many physical processes are continuous in time. Therefore, when identifying a continuous-time model, we can use our insight of the system to decide the system structure and have a direct interpretation of the parameters. Furthermore, in systems such as network control systems and sensor networks, there is a common feature that the output data is quantized, meaning we can only represent our data with a limited amount of distinct values. When performing continuous-time system identification of a system with quantized output data, we have errors from process and measurement noise and also a quantization error. This will make it more difficult to estimate the system parameters. This thesis aims to evaluate if it is possible to obtain accurate estimates of continuous-time systems with quantized output data using the indirect inference method. Indirect Inference is a simulation-based method that first estimates a misspecified auxiliary model to the observed data and in the second step, the parameters of the true system are estimated by simulations. Experiments are done both on a linear system and two non-linear Hammerstein systems with quantized output data. The indirect inference estimator is shown to have the means to yield accurate estimates on both linear systems as well as non-linear Hammerstein systems with quantized output. The method performs better than the simplified refined instrumental variable (SRIVC) method on a linear system which is a commonly used method for system identification of continuous-time systems. Furthermore, it performed significantly better compared to the Hammerstein simplified refined instrumental variable (HSRIVC) method for one of the non-linear systems and slightly better for the second one. The downside is that indirect inference is computationally expensive and time-consuming, hence not a good choice when computation time is a critical factor.