TABLE OF CONTENTS Once the model setup and implementation are checked out, we are ready to apply the model identification and structure reduction procedures first introduced in Chapter 12. The identification algorithm is run until a fully converged solution is achieved. The Secant algorithm (Sec. 12.2.3) can require 500 1000 iterations for more complex model structures to reach convergence, especially if there are several insensitive or highly correlated parameters. The convergence process is nonlinear and might appear to advance slowly for many iterations, but it is very important to continue running the optimization until it reaches a fully converged minimum average cost function J ave. The identification algorithm is then restarted from this point, and it often achieves improved convergence away from this local minimum. A global minimum is reached if the solution then returns to the same solution after a restart, thus indicating that a satisfactory identification result is achieved for this model structure.
The guidelines for desirable cost functions are given in Eqs. (12.23) and (12.24). The model and data frequency-response plots should also be checked to verify that the key dynamic characteristics are well represented. If large mismatches occur in regions corresponding to poor coherence, the data might be unreliable, and the associated frequency-response range should be trimmed back (Sec. 13.5.2). Individual frequency-response pairs T l with marginal coherence and large costs [in excess of Eq. (12.24)] should sometimes be dropped altogether and the corresponding derivatives removed from the model structure (Sec. 13.5.3).
