This work was presented at the Optical Society of America meeting on Fourier Transform Spectroscopy in Santa Barbara, CA on June 25, 1999.

A new vision of Fourier transform-infrared (FT-IR) signal processing is presented. The elements of the signal processing, taken individually, are not novel; the novel combination produces unexpected and useful results. First, the transfer functions of the spectrometer electronics including the detector elements, preamps and analog-to-digital converters (ADCs) are measured and recorded. Then, simultaneously-sampled pairs of data points are recorded from both the infrared and reference laser channels. The interferometer mirror may be scanned with any combination of velocities from constant to sinusoidal to random. The data pairs are recorded at evenly spaced intervals of time without regard to the reference laser modulation frequency or phase. The data are then post-processed by a variety of means that produce equivalent results. The first step in the data processing is to remove the effects, inasmuch as possible, of the detector, preamp, ADCs and noise to generate accurate representations of the pure optical signals. This step removes the frequency-dependent delays (caused by the electro-optic and electronic components) so that the phase and amplitude relationships of the original optical signals are restored. The next step is to convert the data, from points spaced by equal intervals of time to points spaced by equal intervals of retardation, via a difficult interpolation. Several approaches are described in the literature, but these are computationally demanding.^{1,2,3,4,5,6,7} The interpolation step is extremely powerful because, in combination with purification of the optical signals, it converts data sampled with any combination of mirror velocities to interferometric data which is independent of mirror velocity.

There are three convergent historical roots of this work. The first was the desire to use sigma-delta analog-to-digital converters in FT-IR spectrometry.^1,2 The second was the desire to correct for velocity errors in conventional FT-IR spectrometers, particularly ones which are operated in harsh field conditions.^8,9 The third was the fact that almost all very rapid-scan FT-IR spectrometers produce a sinusoidal profile of retardation vs. time.⁶ The common goal of these three areas is to correct the effects of mirror velocity variation. The signal processing approaches described here are successful in reaching the goal in all three cases.

Two new methods of interpolation are examined and compared to previous methods. The first new process is the use of wavelet transforms to extract the frequency and phase of the laser signal. Together with the use of wavelet and inverse wavelet transforms to interpolate the infrared signal, this is equivalent to the second approach. The second approach uses the inverse of Kawata's method¹⁰ to extract the phases of the laser and infrared signals, thereby converting them to zero-based representations. An amplitude scale factor for the infrared signal must be preserved, but this is a fairly simple operation. With the zero-based representations, interpolation, to account for the relationship between time and retardation, is very straightforward. Kawata's method can then be used to reconstruct an amplitude representation of the infrared signal, but with data points evenly spaced in retardation. These two methods produce results equivalent to previous approaches, but offer much better computational efficiency. It is thought that, together with the use of new digital signal processors, these approaches will be capable of real-time processing of more than one thousand interferograms per second.⁶ In particular, Texas Instruments has scheduled for 1998 (and rescheduled for 1999) the release of the TMS320C67 processor which is capable of more than 1 billion floating point operations per second (FLOPS). Boards combining multiple processors of this type will be suitable for real-time signal processing for a variety of applications such as fast process control, quality control and spectral imaging.

The transfer functions of the laser and infrared channels are measured by excitation with fast solid state emitters. In the proof-of-feasibility stage, light-emitting-diodes (LEDs) have been used^11,12 It has been found that inexpensive (< US$2 in 1999) LEDs can reach frequencies higher than 2 MHz. Much higher frequencies may also be available with these inexpensive components. A variety of faster and more expensive solid state emitters are also available and might be appropriate for spectrometer construction. For example, superluminescent diodes used with optical fibers for communication rates in the range of hundreds of MHz, and laser diodes, for the same purpose, with frequencies well into the GHz range, are available for less than US$500. While semiconductor emitters are notoriously unstable with temperature, it is thought that the wavelength drift will have minimal effect on measurement of the phase and magnitude response of signal channels. This is partly because the effects of wavelength drift with wideband detectors are small. Further, the temperature drifts can be made small and slow by the use of insulation, thermal mass, and active control. The crucial issue is that the emitters be stable in their temporal response, and preferably in their total emitted energy. Since their temporal response is on the order of nanoseconds to microseconds, small changes of speed over temperature will have minimal effect on the measurement of infrared and laser channel transfer functions with time constants in the milliseconds to tens of microseconds range. There are several approaches which might be used to measure the transfer functions. In general, the transfer functions of the detector channels would be measured at mirror turnaround points during rapid-scan operation, or during mirror settling time in step-scan operation. The simplest approach to extracting transfer functions is the use of sinusoidal excitation to drive the emitters, and consequently the detector channels. The phase and magnitude of the detector response can be readily measured by recording and demodulating with the same DSP core that operates the spectrometer. The advantages of this approach are conceptual simplicity and consequently straightforward implementation. Other approaches might include pulse or pseudorandom modulation.

The general approach of shifting the complexity of spectrometer operation and signal processing from electronic hardware to electronic software has roots in the computing revolution that made FT-IR spectrometry possible and then practical. In the end, this revolution will make FT-IR spectrometry cheap. The vision of shifting signal processing from conventional electronic hardware to computer hardware and software was first expounded for step-scan in 1993^13,14 and for rapid-scan in 1994.¹⁵ This vision seeks to lower the cost and improve the performance of interferometric measurements. One ironic aspect of the trend is that the complexity of the hardware of a digital signal processor or computer central processor is really many orders of magnitude greater than the complexity of the conventional electronic circuits that it might replace. Further, the capital cost of the integrated circuit fabrication facilities is far greater than the cost of facilities for fabrication of conventional analog components. Because the marginal cost of each additional integrated circuit is very low and because the marginal cost of each additional unit of software is even lower, the tradeoffs become more attractive with each passing year. The fundamental tradeoff of shifting electronic functionality from analog components to digital signal processing is slower execution time. However, the cost of computing is dropping at a compounded rate of 30% per year as the capability increases at a compounded 30% per year. Speed is no longer an issue. In 1960, a unit of computing power equal to 1 thousand FLOPS cost about US$5 per hour. This year (1999), the same performance will cost about US $0.00001 per hour (measured with constant 1960 dollars). This is just short of 6 orders of magnitude change over four decades. As certain companies have foreseen, value has shifted significantly from hardware to software. A convergence of market forces is once again working a revolution in FT-IR spectrometry.

References

1. J. W. Brault, lecture L14 at 10th International Conference on Fourier Transform Spectroscopy, August 27 - September 1, 1995, Budapest, Hungary.
2. J. W. Brault, Applied Optics, Appl. Opt. 35, 2891 (1996).
3. J. Kauppinen, T. Karkkainen, and E. Kyro, Appl. Opt. 17, 1587 (1978).
4. J. C. Brasunas and G. M. Cushman, Appl. Opt. 36, 2206 (1997).
5. H. Weidner and R. E. Peale, Appl. Spectrosc. 51, 1106 (1997).
6. C. J. Manning, 11th International Conference on Fourier Transform Spectroscopy, August, 1997, (AIP Conference Proceedings, Vol. 430), 84 (1997).
7. A. J. Turner, R. A. Hoult, M. D. Forster, European Patent Office Priority Application Number: EP 96307360 (961009).
8. A. S. Zachor, Appl. Optics 16, 1412 (1977).
9. A. S. Zachor and S. M. Aaronson, Appl. Optics 18, 68 (1979).
10. K-i Minami, S. Kawata and S. Minami, Appl. Opt. 31, 6322 (1992).
11. J. C. Brasunas, Infrared Phys. 38, 69 (1997).
12. Christopher J. Manning and Roger J. Combs, Industrial and Environmental Sensors, November, 1998, (SPIE Proceedings, Vol. 3537A), (1999).
13. C. J. Manning, P. R. Griffiths, Appl. Spectrosc. 47, 1345 (1993).
14. C. J. Manning, G. L. Pariente, B. D. Lerner, J. H. Perkins, R. S. Jackson and P. R. Griffiths, "Multiple Modulation Step-Scanning Fourier Transform Infrared Spectrometry" in Computer Enhanced Analytical Spectroscopy, Vol. 6 (S. D. Brown, ed.), Wiley, New York (1996) .
15. P. R. Griffiths and C. J. Manning, FT-IR Spectrometer Evaluation Techniques, Report ERDEC-CR-182, Edgewood RD & E Center, U.S. Army Biol. Defense Command (1995).