Ronald E. Shaffer

A comparison study of chemical sensor array pattern recognition algorithms

Probabilistic neural networks (PNN), learning vector quantization (LVQ) neural networks, back-propagation artificial neural networks (BP-ANN), soft independent modeling of class analogy (SIMCA), Bayesian linear discriminant analysis (BLDA), Mahalanobis linear discriminant analysis (MLDA), and the nearest-neighbor (NN) pattern recognition algorithms are compared for their ability to classify chemical sensor array data. Comparisons are made based on five qualitative criteria (speed, training difficulty, memory requirements, robustness to outliers, and the ability to produce a measure of uncertainty) and one quantitative criterion (classification accuracy). Four sample data sets from our laboratory, involving simulated data and polymer-coated surface acoustic wave chemical sensor array data, are used to estimate classification accuracies for each method. Among the seven algorithms in this study and the four data sets, the neural network based algorithms (LVQ, PNN, and BP-ANN) have the highest classification accuracies. When considering the qualitative criteria, the LVQ and PNN approaches fare well compared to BP-ANN due to their simpler training methods. The PNN is recommended for applications where a confidence measure and fast training are critical, while speed and memory requirements are not. LVQ is suggested for all other applications of chemical sensor array pattern recognition.

The “NRL-SAWRHINO”: a nose for toxic gases

Abstract At the Naval Research Laboratory (NRL), surface acoustic wave (SAW) chemical sensor systems have been in development since 1981. The primary focus has been the detection and identification of chemical agents and other toxic gases or vapors. In the recently developed “NRL-SAWRHINO” system (Rhino, Gr. Nose), a self-contained unit has been developed capable of autonomous field operation. An automated dual gas sampling system is included, for immediate and periodic detection capability. The latter, utilizes a trap-and-purge miniature gas chromatographic column, which serves to collect, concentrate, and separate vapor or gas mixtures prior to SAW analysis. The SAWRHINO includes all the necessary electronic and microprocessor control, SAW sensor temperature control, onboard neural net pattern recognition capability, and visual/audible alarm features for field deployment. The SAWRHINO has been trained to detect and identify a range of nerve and blister agents, and related simulants, and to discriminate against a wide range of interferent vapors and gases.

Speciation of chromium in simulated soil samples using X-ray absorption spectroscopy and multivariate calibration

Abstract X-ray absorption near-edge structure (XANES) spectroscopy has been applied to the speciation of 24 simulated soil samples spiked with known concentrations (8–1015xa0ppm) of hexavalent and trivalent chromium. For XANES spectroscopy, multivariate calibration models, such as partial least-squares (PLS) regression provide greater capabilities than traditional univariate calibration models. Using normalized XANES spectra (5961.8–6114.8xa0eV), a one-factor PLS model was able to determine the percentage of hexavalent chromium with a root-mean-squared-cross-validation (RMSECV) error of 6.1%. The corresponding univariate calibration model based upon the hexavalent chromium pre-edge feature had a RMSECV of 9.1%. PLS was also used to determine total chromium content in the simulated soil samples. Using unnormalized XANES spectra (5980.3–6099.8xa0eV), a two-factor PLS model was able to predict the concentration of chromium with an RMSECV error of 12.1xa0ppm. This work suggests that a single XANES measurement can be used for both the quantitative analysis and speciation of soil samples contaminated with greater than 10xa0ppm chromium.

Multi-criteria fire detection systems using a probabilistic neural network

Abstract The Navy program, Damage Control Automation for Reduced Manning (DC-ARM), is focused on enhancing automation of ship functions and damage control systems. A key element to this objective is the improvement of current fire detection systems. As in many applications, it is desired to increase detection sensitivity and, more importantly, increase the reliability of the detection system through improved nuisance alarm immunity. Improved reliability is needed such that fire detection systems can automatically control fire suppression systems. The use of multi-criteria-based detection technology continues to offer the most promising means to achieve both improved sensitivity to real fires and reduced susceptibility to nuisance alarm sources. A multi-signature early warning fire detection system is being developed to provide reliable warning of actual fire conditions in less time with fewer nuisance alarms than can be achieved with commercially available smoke detection systems. In this study, a large database consisting of the responses of 20 sensors to several different types of fires and nuisance sources was generated and analyzed using a variety of multivariate methods. Three data matrices were developed at discrete times corresponding to the different alarm levels of a conventional photoelectric smoke detector. The alarm times represent 0.82%, 1.63% and 11% obscurations per meter. The datasets were organized into three classes representing the sensor responses for baseline (nonfire), fires and nuisance sources. A robust data analysis strategy for use with a sensor array consisting of four to five sensors for early fire detection and nuisance source rejection was developed using a probabilistic neural network (PNN) that was developed at the Naval Research Laboratory for chemical sensor arrays. The analysis algorithms described in this paper evaluate discrete samples and develop classification models that examine individual chemical signatures at discrete points.

Multiway analysis of preconcentrator-sampled surface acoustic wave chemical sensor array data

New data processing methods for preconcentrator-sampled surface acoustic wave (SAW) sensor arrays are described. The preconcentrator-sampling procedure is used to collect and concentrate analyte vapors on a porous solid sorbent. Subsequent thermal desorption provides a crude chromatographic separation of the collected vapors prior to exposure to the SAW array. This article describes experiments to test the effects of incorporating retention information into the pattern-recognition procedures and to explore the feasibility of multiway classification methods. Linear discriminant analysis (LDA) and nearest-neighbor (NN) pattern-recognition models are built to discriminate between SAW sensor array data for four toxic organophosphorus chemical agent vapors and one agent simulant collected under a wide variety of conditions. Classification results are obtained for three types of patterns: (a) first-order patterns; (b) first-order patterns augmented with the time of the largest peak; and (c) second-order patterns with the use of the SAW frequency for each sensor over a broad time window. Classification models for the second-order patterns are also developed with the use of unfolded and multiway partial least-squares discriminants (uPLSD and mPLSD) and NN and LDA of the scores from unfolded and multiway principal-component analysis (uPCA and mPCA). It is determined that classification performance improves when information about the desorption time is included. Treating the preconcentrator-sampled SAW sensor array as a second-order analytical instrument and using a classification model based upon either uPLSD, uPCA-LDA, or NN results in the correct identification of 100% of the patterns in the prediction set. With the second-order patterns, the other pattern-recognition algorithms only do slightly worse.

Signal processing strategies for passive FT-IR sensors

Computer-generated synthetic single-beam spectra and interferograms are used to study signal processing strategies for passive Fourier transform IR (FTIR) sensor. Synthetic data are generated for one-, two-, and four- component mixtures of organic vapors in two passive FTIR remote sensing scenarios. The single-beam spectra are processed using Savitsky-Golay smoothing, first derivative, and second derivative filters of various orders and widths. Interferogram data are processed by Fourier filtering using Gaussian-shaped bandpass digital filters. Pattern recognition of the target analyte spectral signature is performed using soft independent modeling of class analogy. Quantitative models for the target gas integrated concentration-path length product are built using partial least-squares regression and locally weighted regression. Pattern recognition and calibration models of the filtered spectra and interferograms produced similar results. Chemical detection is possible for complex mixtures if the temperature difference between the source and analyte cloud is sufficiently large. Quantitative analysis is possible if the temperature of the analyte cloud is stable or known and is sufficiently different from the background temperature.

Multivariate mixture analysis using reduced-mass-resolution membrane introduction mass spectrometry and variable selection

The feasibility of analyzing mixtures of spectrally similar analytes with low resolution mass spectrometers coupled with membrane introduction was studied. The performance of the multivariate calibration of an isometric ethyl benzene and p-xylene mixture remained essentially unchanged as the mass resolution degraded. The calibration performance also improved slightly as the data used for calibration decreased from the full mass spectra to only 12 or 18 fragment ions judiciously chosen by variable selection. The multivariate calibration and prediction for a more complex mixture of benzene and toluene only degraded slightly as the resolution decreased, while the result for the two isomers ethyl benzene and p-xylene got progressively worse. Depending on the variable selection algorithm and the number of fragment ions used, using only a select few ions for multivariate calibration and prediction gave results that were either similar to or slightly worse than results using the entire mass spectra. This paper demonstrates that mixture analysis performed with membrane introduction coupled to future miniature, low resolution mass spectrometers is possible.

Signal processing strategies for passive FT-IR remote sensing

Computer-generated synthetic single-beam spectra and interferograms are used to study signal processing strategies for passive Fourier transform IR (FTIR) remote sensing. Four-component mixtures of ethanol, methanol, methyl ethyl ketone, and acetone vapors in two passive FTIR remote sensing scenarios are studied. Interferogram and spectral processing strategies are compared based on the ability to generate accurate pattern recognition and multivariate calibration models. Chemical detection and quantitative analysis is further assessed based on analyte band characteristics and spectral overlap. Temperature differential between the analyte cloud and the IR source is found to be the critical factor determining classification performance. Calibration model performance is further dependent upon the ability to reliably assess analyte plume temperature. Calibration model performance is shown to degrade significantly as the uncertainty in plume temperature increases or as the temperature difference decreases. These results suggest that in order to quantify passive FTIR spectra remotely either there must be either a large temperature difference or an accurate assessment of plume temperature.

UXO target detection using magnetometry and EM survey data

Digital filtering, principal component analysis (PCA), and an automated anomaly picker have been used to improve and automate target selection of unexploded ordnance (UXO). This is the first step in a three part program to develop new data analysis methods to automate target selection and improve discrimination of UXO from clutter and ordnance explosive waste (OEW) using magnetometry (Mag) and electromagnetic induction (EM) survey data. Traditionally, target detection has been accomplished by a time-consuming manual interactive data analysis approach. Experts screen the magnetometer data and select potential UXO targets based on their intuitive experience. EM data has been used in a secondary role in this process and the anomaly picking included classification and operator bias. In this program, the target detection step will use all of the data available and a separate classifier process will be used for identification and discrimination. Digital filtering is being used to enhance important features and reduce noise, while principal component analysis is being used to fuse three channels of data and reduce noise. Seven 50 meter-square data sets from two test sites were used to investigate these techniques. Features of interest are enhanced using filtering techniques. Inspection of the first- principal component suggests that data fusion of the magnetometer and EM data can be successfully accomplished. The new image consisting of circular features of varying diameters and intensities represent significant features present in all three data channels. Data with strong magnetometer and EM signals have the greatest intensity and in most cases noise is reduced. An automated anomaly picker has been designed to select targets from Mag, EM and PCA images. The method is fast and efficient as well as providing user options to control pick criteria.

Probabilistic neural networks for the discrimination of subsurface unexploded ordnance (UXO) in magnetometry surveys

The outputs from a physics-based modeler of magnetometry data have been successfully used with a probabilistic neural network (PNN) to discriminate UXO from scrap. Model outputs from one location at a site were used to train a PNN model, which could correctly discriminate UXO from scrap at a different location of the same site. Data from one site location, the Badlands Bombing Range, Bulls Eye 2 (BBR 2), was used to predict targets detected at a different location at the site, Badlands Bombing Range, Bulls Eye 1 (BBR 1) containing different types of items. The UXO detection rate obtained for this analysis was 93 percent with a false alarm rate of only 28 percent. The possibility of discriminant individual UXO types within the context of a coarser two- class problem was demonstrated. The utility of weighting the sum of squared errors in cross-validation optimization of the (sigma) parameter has been demonstrated as a method of improving the classification of UXO versus scrap.