Diz 58 en
Distributed Identification of Nonlinear Systems using Regularization
Author: Radek Beňo
Thisthesisdealswithanewmethodofidentifyingnonlinearsystemswhichcon- sist of subsystems called components. The identification problem is here under- stood as a process of parameters’ calibration of nonlinear systems with fixed struc- ture. Thewholeworkdealswiththecalibrationofparametersofnonlinearsystems in steady states. One of the greatest contribution of the work is the component regularization methodology, which mainly brings better numerical stability of the solution. The presented algorithm is distributed and decomposes the original problem accordingtotheprimaldecompositiontoaseriesofsimplersubproblems,inwhich oneparticularsteadystateofthesystemissolved,i.e. fittedtothedata,accordingto a given global parameter vector. These subproblems can be solved independently of each other, a global optimizer collects these individual contributions and itera- tively changes the parameter’s values according to an optimization criterion. Regularized components support the calibration process in particular by cor- rect definition of the domain of the model’s validity, i.e. the area where the model isnumericallywellconditioned, andnumericalstabilityisfurtherstrengthenedby introduction of additional variables that constrain the input, output and internal signals of components. These additional constraints limit the propagation of non- physical signal values across the entire system model. A Mean-Value Model has been chosen as a type of system model over which distributed optimization works, which makes possible not only to model the basic physical phenomena of the system, but also to use it well within the framework of the further system control design. The presented method is demonstrated in this thesis on a specific example of the calibration of the non-linear model of a Diesel internal combustion engine.
- Radek Beňo, mailto:benorade@fel.cvut.cz