3.5. Robustness to measurement noise
In order to evaluate the robustness of SPA–LDA and SIMCA with
respect to measurement noise, a Monte Carlo simulation was carried
out (Martins et al., 2003). For this purpose, artificial noise was
added to the spectra of the test samples. The noise level was set according to the pooled standard deviation calculated in Section
3.1. Ten different noise permutations were added to the spectra
of each of the 43 test samples, thus generating 430 noisy spectra.
By using the SPA–LDA model, a classification accuracy of 96%
was achieved (19 errors out of 430). The full-spectrum SIMCA
models yielded 10 Type-I and 260 Type-II errors. By using the variables
selected by SPA–LDA, SIMCA resulted in 20 Type-I and 224
Type-II errors.