Experimental design protocol for the pattern recognition analysis of bandpass filtered Fourier transform infrared interferograms

Signal processing techniques are described for open path Fourier transform infrared (FTIR) measurements that overcome fundamental limitations in conventional data analysis strategies. Bandpass digital filters are applied directly to FTIR interferograms to isolate spectral frequencies of interest, followed by pattern recognition analysis of segments of the filtered interferogram to provide an automated means for detecting a target analyte. In this work, four experimental variables are significant: (1) interferogram segment length, (2) interferogram segment position, (3) filter bandpass position, and (4) filter bandpass width. The limit of detection of a compound is directly related to the ability to choose optimal settings for these variables. Laboratory data collected when SF6 was present in the optical path of the spectrometer are employed to perform a full factorial experimental design study of these four variables. Analysis of variance techniques are employed to provide a statistical means of interpreting the main and interaction effects existing among the variables. Based on these results, a protocol for designing a near-optimal bandpass filter is developed. Interferogram segment starting position, filter bandpass position, and filter bandpass width interaction effects are found to be statistically significant. It is concluded that these factors must be optimized jointly. Interferogram segment length is found to have the most overall influence on detection performance, but can be studied independently from the other variables. To validate the filter generation protocol developed with the laboratory data, additional work is performed with open path SF6 data collected during a series of field trials. The results of this study demonstrate that the protocol is valid and can be extended to develop detection schemes for other compounds.