Overview

Data windowing was introduced in Chapter 7 as a standard method of spectral analysis to reduce the random error in the spectral estimates. Repeat maneuvers are first linked into concatenated records, which are then segmented into a sequence of shorter overlapping time records called windows or periodograms. The guidelines for window-size selection (Sec. 7.10) provide a range of possible window lengths T _win. The ultimate choice of window length is a balance between increased information content at low frequency (larger windows) vs better random error suppression which is important at higher frequencies (smaller windows). Hence no single window size selection will produce optimal results over the entire frequency range of interest. Furthermore, a window that is best suited to a particular input/output frequency-response pair might be much less desirable for other input/output pairs in the MIMO frequency-response estimate matrix .

To obtain the most accurate identification estimates for a particular situation, the spectral calculations (SISO and multi-input conditioning) would have to be performed repeatedly using several different window sizes. The best window size could then be selected based on the compromise just discussed and an individual assessment based on each frequency-response pair. Clearly, this would be a manually intensive exercise, and in the end it would still be a compromise. The optimization-based composite-windowing technique both eliminates the need for manual assessment and overcomes the compromises entailed with a single window-size selection. This technique merges the results obtained from the range of window sizes into a...