Both the power and the challenge of hyperspectral technologies is the very large amount of data produced by spectral
cameras. While off-line methodologies allow the collection of gigabytes of data, extended data analysis sessions are
required to convert the data into useful information. In contrast, real-time monitoring, such as on-line process control,
requires that compression of spectral data and analysis occur at a sustained full camera data rate. Efficient, high-speed
practical methods for calibration and prediction are therefore sought to optimize the value of hyperspectral imaging.
A novel method of matched filtering known as science based multivariate calibration (SBC) was developed for
hyperspectral calibration. Classical (MLR) and inverse (PLS, PCR) methods are combined by spectroscopically
measuring the spectral "signal" and by statistically estimating the spectral "noise." The accuracy of the inverse model is
thus combined with the easy interpretability of the classical model. The SBC method is optimized for hyperspectral data
in the Hyper-CalTM software used for the present work. The prediction algorithms can then be downloaded into a
dedicated FPGA based High-Speed Prediction EngineTM module. Spectral pretreatments and calibration coefficients are
stored on interchangeable SD memory cards, and predicted compositions are produced on a USB interface at real-time
camera output rates. Applications include minerals, pharmaceuticals, food processing and remote sensing.
This paper will discuss recent results obtained when applying a photoconductive linear MCT array in a demonstration spectrometer designed for the NIR wavelength range from 1300 to 2500 nm. A new 128x1 element MCT sensor was developed specifically for spectroscopy, i.e. with "tall", rectangular pixels in order to optimize both wavelength resolution and optical throughput. Also new read-out electronics was developed using multilayer LTCC (Low Temperature Co-Fired Ceramics) techniques, which is integrated into the package and realizes synchronous ("lock in") detection for each of the 128 channels. Advantages of this current-detection scheme include compatibility with chopped light sources (insensitivity to ambient stray light) and elimination of read-out noise (affecting charge-detection amplifiers). The first test results reported here confirm spectrometer operation and present encouraging performance, even though the system is not yet optimized. The spectrometer is very fast, with minimum integration time of 1.2 ms, while photometric noise will reduce with longer integration times. There is no fundamental limit in the maximal length of the integration time. Testing with integration times of 1.2, 12, 120 and 1200 ms resulted in absorbance noise levels of approximately 2500, 330, 94 and 49 μA units. Demonstration spectra were measured from lactose and copying paper samples. Thanks to high speed and parallel spectral recording of 128 wavelengths, MCT array technology appears highly potential for developing powerful on-line spectrometers for process analytical applications not only in the near infrared (NIR) but also for the lower mid-IR wavelengths, up to approximately 6 μm.
Recent progress in spectroscopy and chemometrics have brought in-vitro blood glucose analysis into clinical reach. Parallel to these efforts noninvasive experiments by NIR- spectroscopy have also been proposed and carried out. A specially designed and optimized accessory for diffuse reflectance measurements in the spectral range of 9000 - 5000 cm-1 was used. The spectral data and reference concentration values were obtained using oral glucose tolerance tests. Calibration results are provided for log(1/R) and single beam spectra. In addition, the effects of smoothing and the use of derivative filtering were evaluated. The best results were achieved by multivariate PLS-modeling with raw data from single beam reflectance spectra.
A spectral analysis of human EDTA plasma from 124 different patients was undertaken using an FT-NIR spectrometer. Optimized intervals of the absorbance spectra were considered for the PLS-calibrations with reference concentrations from standard clinical methods determined in triplicate. The following relative prediction errors were achieved: total protein 1.4%, cholesterol 3.8%, triglycerides 8.6%, glucose 9.1%, and urea 18.6%.
The value of diffuse reflectance spectroscopy for infrared analytical methods has some limitations caused by the low throughput efficiency of commercial accessories. We have redesigned a home-built diffuse reflectance accessory with an on-axis ellipsoidal collecting mirror which can also have benefits for the analysis of bulky samples. For illumination, use is made of the Circle cell optics which fits by adaption into the cone with a half angle of 30 degree(s) not considered for detection. Special features of the accessory are discussed, especially in view of our applications to the study of human tissue.