In most papers on parameter estimation schemes the order of the process under study is assumed to be known a priori. In many practical situations, however, this information is lacking. Consequently simple methods for the determination of the order are necessary. In this paper a description is given of five tests based on respectively: (1) the behaviour of the error function, (2) the whiteness of the residuals (correlation function), (3) the statistical independency of loss functions, (4) the behaviour of the determinant, and (5) the pole-zero cancellation effect. Some of these methods, the pole-zero cancellation, test of whiteness, and F-test are used as known, but the others are revised in such a way that a better discrimination of the order is obtained. The methods based on the behaviour of the error function and the behaviour of the determinant respectively are extended in such a way that the orders of both process- and noise dynamics are estimated separately. An extensive set of simulations of the Astrom-process is presented in order to elucidate the comparison of the different methods. It turns out that an acceptable determination of the order is possible even with a signal-to-noise ratio at the process output of - 15dB.
[1]
I. Gustavsson.
Maximum Likelihood Identification of Dynamics of the Ågesta Reactor and Comparison with results of Spectral Analysis
,
1969
.
[2]
Walerian Kipiniak,et al.
Optimal Estimation, Identification, and Control
,
1964
.
[3]
R. H. Battin,et al.
Random Processes in Automatic Control.
,
1957
.
[4]
Karl Johan Åström,et al.
Lectures on the Identification Problem : The Least Squares Method
,
1968
.
[5]
C. Woodside.
Estimation of the order of linear systems
,
1971
.
[6]
J. C. Chow.
On estimating the orders of an autoregressive moving-average process with uncertain observations
,
1972
.