Husheng Yang

Comparison of partial least squares regression and multi-layer neural networks for quantification of nonlinear systems and application to gas phase Fourier transform infrared spectra

The performance of back-propagation artificial neural networks (NN) and partial least squares (PLS) regression for the calibration of linear and nonlinear systems has been investigated by using six types of synthetic data. Three PLS methods, conventional linear-PLS and two nonlinear-PLS methods, have been used in the study. In all but one of the synthetic data types, the band intensities varied nonlinearly with concentration. These five data types were designed to represent the effect of band shifts with increasing concentration, a nonlinear relationship between peak height and concentration, or a combination of both types of nonlinearities. The results showed that NNs perform better than PLS for all the nonlinear datasets. When a band shift is the major reason for the nonlinearity, the relative performance of NNs and PLS depends on the overlap of the absorption bands. If there is no band overlap, neither NN nor PLS can calibrate the data accurately but the results could be improved by convolving the spectral features with a Gaussian broadening function. The results indicate that a combination of peak position shift and peak height change is the most difficult nonlinearity to calibrate. NN and PLS were also used to determine the concentration of CHCl3 in pure component and mixtures of CHCl3 and CH2Cl2 using their Fourier transform infrared (FT-IR) spectra, a dataset that has been proved nonlinear in high concentrations due to the nonlinear response of the detector. The best results for the experimental data were obtained by applying one hidden layer NN to the mean-centered absorbance spectra.

Raman spectrometry and neural networks for the classification of wood types. 2. Kohonen self-organizing maps

Abstract One- and two-dimensional Kohonen self-organizing maps (SOMs) were successfully used for the unsupervised differentiation of the Fourier transform Raman spectra of hardwoods from softwoods. The SOMs were also applied to differentiate temperate woods from tropical woods, and results showed that the two types of woods could only be partly differentiated. A semi-quantitative method that is based on the Euclidean distances of the weight matrix has been developed to assist the automatic clustering of the neurons in a two-dimensional SOM.

Effects of Detector Nonlinearity on Spectra Measured on Three Commercial FT-IR Spectrometers

The effect of the nonlinear response of mercury-cadmium-telluride (MCT) detectors has been evaluated on three commercial Fourier transform infrared (FT-IR) spectrometers. The greater the photon flux and the photon flux density, and the smaller the area of the detector on which the source image is focused, the greater are the effects of the nonlinearity. The signal-to-noise ratio (SNR) of spectra measured with an MCT detector under conditions of high photon flux and, especially, high photon flux density is significantly less than the SNR calculated by using the manufacturers D* value. Detector nonlinearity usually leads to negative deviations in Beers law plots. An empirical correction algorithm has been applied to Beers law spectra acquired with the use of photoconductive MCT detectors and has been found to work well.

Evaluation of a Correction for Photometric Errors in FT-IR Spectrometry Introduced by a Nonlinear Detector Response

For strongly absorbing bands measured with a Fourier transform infrared (FT-IR) spectrometer, the effects of a nonlinear detector response must be eliminated before Beers law linearity can be achieved. An empirical method for greatly reducing the effect of detector nonlinearity on FT-IR Beers law spectra measured by using an FT-IR spectrometer equipped with a mercury-cadmium-telluride (MCT) detector is investigated. This first-order analytical function has been applied to correct nonlinear vapor-phase spectra and statistically evaluated for validity for spectral regions above the detector cutoff. In addition, a series of second-order functions has been evaluated to investigate the possibility that the transmittance scale is slightly nonlinear even after the first-order correction has been applied. Any improvement caused by the second-order functions was not statistically significant.

Open-path FT-IR spectrometry: is completely unattended operation feasible?

Most protocols used for open-path Fourier transform infrared spectrometry (OP/FT-IR) require that spectra be measured at a resolution of 1 cm(-1) and that the concentrations of the analytes be calculated by classical least squares regression (CLS). These specifications were largely developed for monitoring light molecules with easily resolvable rotational fine structure. For most volatile organic compounds in air, the rotational fine structure is not resolvable and better accuracy can be obtained when the spectrum is measured at lower resolution (typically 8 cm(-1)), provided that the algorithm used for quantification is partial least squares regression (PLS). By measuring the spectrum at low resolution, the need for a liquid-nitrogen-cooled mercury cadmium telluride detector is reduced and a pyroelectric detector operating at ambient temperature can be used instead. By using PLS rather than CLS, spectral features due to water vapor do not have to be compensated and a short-path background spectrum can be used, greatly simplifying field measurements.

Accurate wavelength measurements of a putative standard for near-infrared diffuse reflection spectrometry.

The diffuse reflection (DR) spectrum of a sample consisting of a mixture of rare earth oxides and talc was measured at 2 cm−1 resolution, using five different accessories installed on five different Fourier transform near-infrared (FT-NIR) spectrometers from four manufacturers. Peak positions for 37 peaks were determined using two peak-picking algorithms: center-of-mass and polynomial fitting. The wavenumber of the band center reported by either of these techniques was sensitive to the slope of the baseline, and so the baseline of the spectra was corrected using either a polynomial fit or conversion to the second derivative. Significantly different results were obtained with one combination of spectrometer and accessory than the others. Apparently, the beam path through the interferometer and DR accessory was different for this accessory than for any of the other measurements, causing a severe degradation of the resolution. Spectra measured on this instrument were removed as outliers. For measurements made on FT-NIR spectrometers, it is shown that it is important to check the resolution at which the spectrum has been measured using lines in the vibration-rotation spectrum of atmospheric water vapor and to specify the peak-picking and baseline-correction algorithms that are used to process the measured spectra. The variance between the results given by the four different methods of peak-picking and baseline correction was substantially larger than the variance between the remaining five measurements. Certain bands were found to be more suitable than others for use as wavelength standards. A band at 5943.13 cm−1 (1682.62 nm) was found to be the most stable band between the four methods and the six measurements. A band at 5177.04 cm−1 (1931.61 nm) has the highest precision between different measurements when polynomial baseline correction and polynomial peak-picking algorithms are used.

Improved Data Processing by Application of Brault's Method to Ultra-Rapid-Scan FT-IR Spectrometry:

In Fourier transform infrared spectrometry, the use of analog-to-digital converters clocked at a constant rate leads to interferograms the data points of which are evenly spaced in time. An interpolation step is generally required to resample these data to equal intervals of retardation. Three methods of resampling the data were compared: Fourier interpolation (i.e., zero-filling), cubic interpolation, and digital filter interpolation as described by Brault. These approaches were investigated using both simulated data and interferograms acquired from an ultra-rapid scanning Fourier transform infrared spectrometer. In both cases, the optical path difference varied sinusoidally with time and interferograms were sampled at equal time intervals. It was shown that the data processing time can be reduced significantly, relative to previous methods, through the application of Braults method, which can be considered as three separate steps involving a first-, second-, and third-order correction. If only the first-order correction for the local phase relationship between the sampling grids is applied, the noise level of resulting spectra is similar to that obtained by cubic fitting, but the data processing time is reduced significantly. If the second-order correction for the variation of velocity is also used, the data processing time is approximately the same as the first-order correction, while the signal-to-noise ratio of the spectra is increased significantly. The implementation of the third-order correction for channel delay mismatch appears to be unnecessary for this instrument because the detector channel delays are small.

Effect of resolution on the wavelength determination of a putative standard to be used for near infrared diffuse reflection spectra measured on grating spectrometers

Diffuse reflection (DR) Fourier transform near infrared spectra of a powdered mixture of Er2O3, Dy2O3, Ho2O3 and talc were measured at a constant resolution of at least 2 cm−1 on four different combinations of spectrometers and sampling accessories. The wavenumber scale of each of these spectra was corrected with lines in the vibration–rotation spectrum of water vapour so that the accuracy was better than 0.02 cm−1. DR spectra of the powdered sample were then calculated at constant-wavelength resolution to simulate spectra that would have been measured on a grating monochromator. The precision and accuracy of the bands in the spectrum of this sample were then estimated. It is believed that the accuracy of the reported positions of the bands in the spectrum of this putative wavelength standard is about ten times better than the corresponding values in the DR spectrum of a mixture of Er2O3, Dy2O3 and Ho2O3 that had previously been reported by the US National Institute of Standards and Technology.

Effect of resolution on the wavenumber determination of a putative standard to be used for near infrared diffuse reflection spectra measured on Fourier transform near infrared spectrometers

A mixture of powdered Er2O3, Dy2O3, Ho2O3 and talc has been proposed as a wavenumber standard for near infrared (NIR) diffuse reflection (DR) measurements. Previous measurements were made at a resolution of 2 cm−1 which is significantly less than the full-width at half-height of the peaks in the spectrum. Since many NIR DR spectra are measured at a much lower resolution, some of the peaks are observed to fuse and/or shift. In this paper, the behaviour of the band positions is studied as a function of resolution. The factors that allow some of the peaks to be useful as wavenumber standards for NIR DR measurements while others should not be used are discussed. The instrument should first be calibrated with certain lines in the vibration–rotation spectrum of water vapour. Because most of the lines in the water spectrum are located less than 1 cm−1 from their nearest neighbours, only two of the 3768 lines in the region between 5750 and 7990 cm−1 can be used for calibration and only then if the spectrum is converted to its second derivative and interpolated so that there are at least five data points per resolution element and the wavenumber of maximum absorption is found by fitting the line to a quartic function. For the spectrum of the putative wavenumber standard measured at a resolution of 3–10 cm−1, the band centre is best measured after baseline correction of the spectrum with a polynomial function, while lower resolution spectra are better converted to the second derivative. As might be expected, when choosing peaks as wavenumber standards for instruments operating at different resolutions, one should avoid bands that either become fused at resolutions close to the value being used or whose position varies rapidly with resolution. Peaks should be selected that, besides being in the wavenumber range of interest, can be measured with both high precision and high accuracy.

Low-cost OP-FTIR spectrometer for workplace monitoring

OPTRA has developed a low-cost, extremely compact, rugged open-path Fourier transform infrared (OP-FTIR) spectrometer for workplace air quality monitoring. This research was funded under a United States Air Force ABIR Phase II contract. The goal of the program has been to identify and alleviate all aspects of currently available OP-FTIR systems which result in high-cost and complex user requirements. This low-resolution ssytem (Δσ = 8 cm-1) employs an uncooled DLATGS detector and a novel encoder-based reference metrology. Other design economies include a plastic injection-molded retroreflector array to return the open-path beam. This effort has included the development of a set of algorithms based on artificial neural networks (ANNs) and partial least squares (PLS) by the University of Idaho; these algorithms are specifically tailored to low-resolution systems applied to multi-component analysis of large, organic molecules characterized by broad infrared resonance bands. The algorithms, coupled with our OP-FTIR, are designed to autonomously identify and quantify a list of 105 common industrial organic molecules in the presence of varying humidity levels. Our system includes two PCI boards which host all OP-FTIR processing and servo electronics; the boards reside in a small suit-case PC along with a user-friendly Graphical User Interface.