Composite Vapor Detectors Research Articles

Comparison of Fisher's linear discriminant to multilayer perceptron networks in the classification of vapors using sensor array data

Two different classification methods, Fisher's linear discriminant (FLD) and a multilayer perceptron neural network (MLP), were directly compared with respect to their abilities to differentiate response patterns arising from arrays of chemical vapor detectors. The algorithms were compared in five different types of tasks that had been selected because they produced classification problems of varying character and difficulty. In one task, an array of 20 compositionally distinct carbon black–polymer composite vapor detectors was exposed to P/ P 0 = 0.0075 1-propanol and P/ P 0 = 0.0083 2-propanol, where P and P 0 are the partial pressure and standard vapor pressure, respectively, of a given analyte. The second task consisted of classification of a mixture of P/ P 0 = 0.011 1-propanol and P/ P 0 = 0.0090 2-propanol versus a mixture of P/ P 0 = 0.0090 1-propanol and P/ P 0 = 0.011 2-propanol. A third task consisted of multiple concentrations of three hydrocarbons, and a fourth task involved clustering two hydrocarbons in the presence of a variable background composition. An additional dataset was generated by exposing an array of five thin-film metal-oxide sensors to the headspace of seven different coffee blends. In each case, the MLP and FLD techniques were compared using the 5-sensor subset of the 20 available sensors that proved optimal for that dataset. The FLD and MLP algorithms yielded comparable performance on straightforward classification tasks, whereas the MLP technique yielded better performance on tasks that involved non-linear classification boundaries. In addition, for the four datasets produced by the carbon black–polymer composite detector array, the performance of each possible 5-sensor subset was evaluated using both signal processing approaches. The performance of the best 5-sensor subset selected with MLP was found to be slightly better than the performance of the FLD-selected subsets, and the performance of the median 5-sensor subset using MLP was nearer to that of the optimal subset than the median sensor array selected by FLD. In one case, the optimal test set performance distribution was found to be significantly better with MLP than with FLD: MLP had a clear advantage (86% versus 57% correct classification rate) when applied to the “coffees” dataset, and this trend is likely applicable to other multi-cluster classification tasks that consisted of non-Gaussian shaped data in lower-dimensional spaces.

Comparison of analytical methods and calibration methods for correction of detector response drift in arrays of carbon black-polymer composite vapor detectors

The responses of 15 carbon black-polymer composite chemiresistors have been analyzed during exposure to eight different analytes ( n-hexane, tetrahydrofuran, ethanol, ethyl acetate, cyclohexane, n-heptane, n-octane, and isooctane) in random order at low concentration (0.5% of the vapor pressure of analyte at room temperature) over 4 months (8000 total analyte exposures) of data collection. Data were collected for periods during which the array was continuously exposed periodically to analytes and after long periods during which no analyte exposures had been performed. All but the most difficult separation tasks (for example, discrimination between low concentrations of straight-chain hydrocarbons) could be performed robustly over the entire 4 month time period based only on the use of a decision boundary formulated from an initial training set of 200 exposures, indicating the sensor drift had minimal effect on system performance in such classification tasks. For the remaining classification tasks, modeling the dynamics of sensor drift either through a linear regression or Fourier transform decomposition of the individual relative differential resistance responses versus time of each sensor yielded little improvement in classification performance, indicating that external events were largely responsible for changes in sensor response versus time. Six analytes that were not treated as unknowns for a binary separation task were individually treated as calibrants whose response was intermittantly used to renormalize the response of the sensor array. A simple linear sensor-by-sensor calibration scheme proved effective at restoring the classification performance of difficult binary separation tasks to the performance that was observed in the initial training set period. Calibrants that were mutually similar to the analytes being differentiated tended to be more effective than calibrants that were very chemically different from the analytes of interest. Evaluation of various calibration protocols indicated that an optimal tradeoff existed between the number of calibration exposures and the frequency of calibration periods. Condition-based calibration, in which calibration was only performed when the classification model exhibited a decline in classification performance below a predetermined threshold value, was observed to be superior to a time-based calibration approach or to interval-based, cyclic calibration protocols for this set of analytes exposed under the chosen analysis conditions.

Estimation of chemical and physical characteristics of analyte vapors through analysis of the response data of arrays of polymer-carbon black composite vapor detectors

Analysis of the signals produced by a collection of organic polymer-carbon black composite vapor detectors has been performed to assess the ability to estimate various chemical and physical properties of analyte vapors based on information contained in the response patterns of the detector array. A diverse array of composite chemiresistive vapor detectors was exposed to a series of 75 test analytes that had been selected from among five different chemical classes: alcohols, halogenated hydrocarbons, aromatics, unsubstituted hydrocarbons, and esters. The algorithmic task of interest was to use the resulting array of response data to assign one of the five chemical class labels to a test analyte, despite having left that analyte out of the model used to generate the class labels. Algorithms evaluated for this purpose included principal components analysis (PCA) and k-nearest neighbor ( k-NN) analysis employing either Euclidean or Mahalanobis distance calculations. Each data cluster that was produced by replicate exposures to an individual analyte was well resolved from all of the other 74 analyte clusters. Furthermore, with the exception of the halide cluster, the analyte response clusters could be robustly grouped into supersets such that each of the five individual chemical classes was well-separated from every other class of analytes in principal component space. Accordingly, using either of the k-nearest neighbor algorithms, in excess of 85% of the non-halide test analyte exposures were correctly assigned to their chemical classes, and halides were only routinely confused with aromatics or esters but not with alcohols or hydrocarbons. The detector array response data also was found to contain semi-quantitative information regarding physicochemical properties of the members of the test analyte series, such as the degree of unsaturation of the carbon chain, the dipole moment, the molecular weight, the number of halogen atoms, and type of aromatic ring in the test analytes. The performance in these types of tasks is relevant for applications of a semi-selective array of vapor detectors in situations when no prior knowledge of the analyte identity is available and when there is no assurance that the test analyte will have been contained in the training set database produced by a compiling a library of responses from the detector array.

Characterization of the Temporal Response Profile of Carbon Black−Polymer Composite Detectors to Volatile Organic Vapors

The relative differential resistance responses of carbon black−poly(ethylene-co-vinyl acetate) (PEVA) composite vapor detectors were evaluated in response to short rise time (<2 ms for a 17 ms pulse length) square pulses of acetone, n-hexane, methanol, 2-propanol, or toluene, in a background of synthetic air. The use of ultrathin films, along with a rapid vapor delivery system, facilitated measurement of the rapid time response available from this exemplary carbon black−polymer composite chemiresistive film for the detection of common organic vapors. Detectors formed from very thin (<200 nm) PEVA−carbon black composites produced steady-state responses within 17 ms upon exposure to methanol and produced steady-state responses within 90 ms upon exposure to toluene, acetone, and n-hexane. In accord with Fickian diffusion, the response times of the relative differential resistance of PEVA−carbon black detectors to analyte exposures were proportional to the square of the film thickness, l, in the range 510 ≤ l ≤ 5700 nm. Additionally, the relative differential resistance versus time profiles of PEVA−carbon black detectors were well fit by a simple finite difference model based on Fickian analyte diffusion, using a single analyte diffusion coefficient, for a variety of different film thicknesses and analyte concentrations.

Enhanced sensitivity to and classification of volatile carboxylic acids using arrays of linear poly(ethylenimine)-carbon black composite vapor detectors.

Vapor detectors formed from composites of conductors and insulating organic polymers have been tailored to produce increased sensitivity toward specific classes of analyte vapors. Upon exposure to acetic acid at 1% of its vapor pressure, detectors consisting of linear poly(ethylenimine) (1-PEI)-carbon black composites showed an approximately 10(3) increase in signal/noise relative to the performance of typical insulating organic polymer-carbon black composite vapor detectors. Compositional diversity in an array of such vapor detectors was obtained by varying the degree of plasticization of the 1-PEI films. The resulting vapor detector array produced sensitive detection of, and robust discrimination between, various volatile organic acids and relatively little response from nonacidic organic vapors or from water vapor. Measurements of the mass uptake, thickness change, and electrical conductivity of such composites indicate that swelling of the polymer film, and thus its normalized resistance response, is beyond that expected by mass uptake alone upon exposure to acetic acid vapor. This additional thickness increase is attributed to charge-induced polymer swelling occurring from polymer-analyte interactions. Electrical percolation also plays a significant role in producing the large increase in normalized resistance response of these composites upon exposure to acetic acid vapor.

Classification performance of carbon black-polymer composite vapor detector arrays as a function of array size and detector composition

The vapor classification performance of arrays of conducting polymer composite vapor detectors has been evaluated as a function of the number and type of detectors in an array. Quantitative performance comparisons were facilitated by challenging a collection of detector arrays with vapor discrimination tasks that were sufficiently difficult that at least some of the arrays did not exhibit perfect classification ability for all of the tasks of interest. Specific discrimination tasks involved differentiating between low concentration (<1% of the vapor pressure) exposures to 1-propanol versus 2-propanol, low concentration exposures to n-hexane versus n-heptane, and differentiating between compositionally similar mixtures of closely related analytes, such as 9.37 ppm m-xylene with 10.2 ppm p-xylene versus 7.67 ppm m-xylene with 12.4 ppm p-xylene. A decision boundary was developed using a cross-validated Fisher linear discriminant algorithm on a training set of analyte presentations and the resulting chemometric model was then used to classify a subsequent collection of test analyte presentations to the array being evaluated. In other cases, classification performance was evaluated using the Fisher linear discriminant and a leave-one-out (LOO) cross-validation procedure. For nearly all of the discrimination tasks investigated in this work, classification performance either increased or did not significantly decrease as the number of chemically different detectors in the array increased. Any given subset of the full array of detectors, selected because it yielded the best classification performance at a given array size for one particular task, was invariably outperformed by a different subset of detectors, and by the entire array of 20 chemically diverse detectors when used in at least one other vapor discrimination task. Arrays of detectors were nevertheless identified that yielded robust discrimination performance between compositionally close mixtures of 1-propanol and 2-propanol, n-hexane and n-heptane, and m-xylene and p-xylene, attesting to the excellent analyte classification performance that can be obtained through the use of such semi-selective vapor detector arrays.

Detection and classification characteristics of arrays of carbon black/organic polymer composite chemiresistive vapor detectors for the nerve agent simulants dimethylmethylphosphonate and diisopropylmethylphosponate.

Arrays of conducting polymer composite vapor detectors have been evaluated for performance in the presence of the nerve agent simulants dimethylmethylphosphonate (DMMP) and diisopropylmethylphosponate (DIMP). Limits of detection for DMMP on unoptimized carbon black/ organic polymer composite vapor detectors in laboratory air were estimated to be 0.047-0.24 mg m(-3). These values are lower than the EC50 value (where EC50 is the airborne concentration sufficient to induce severe effects in 50% of those exposed for 30 min) for the nerve agents sarin (methylphosphonofluoridic acid, 1-methylethyl ester) and soman (methylphosphonofluoridic acid, 1,2,2-trimethylpropyl ester), which has been established as approximately 0.8 mg m(-3). Arrays of these vapor detectors were easily able to resolve signatures due to exposures to DMMP from those due to DIMP or due to a variety of other test analytes (including water, methanol, benzene, toluene, diesel fuel, lighter fluid, vinegar, and tetrahydrofuran) in a laboratory air background. In addition, DMMP at 27 mg m(-3) could be detected and differentiated from the signatures of the other test analytes in the presence of backgrounds of potential interferences, including water, methanol, benzene, toluene, diesel fuel, lighter fluid, vinegar, and tetrahydrofuran, even when these interferents were present in much higher concentrations than that of the DMMP or DIMP being detected.

Comparison of odor detection thresholds and odor discriminablities of a conducting polymer composite electronic nose versus mammalian olfaction

Response data from an array of conducting polymer composite vapor detectors that form an electronic nose were collected for the purpose of comparing selected, quantitatively measurable, phenomena in odor detection and classification to the olfactory characteristics of monkeys and humans. Odor detection thresholds and discriminability between structurally similar pairs of odorants were the two primary quantities evaluated for this comparison. Comparisons were only made for volatile organic vapors as opposed to aroma active odorant vapors. Electronic nose detection thresholds for a homologous series of n-alkane and 1-alcohol odorants were determined and the results were compared to literature values for the mean olfactory detection thresholds observed in psychophysical experiments on humans exposed to these same vapors. The trends in odor detection thresholds of the electronic nose towards the tested analytes were very similar to those exhibited by humans. The discrimination performance of the electronic nose for distinguishing between pairs of odorants within incrementally varying series of esters, carboxylic acids and alcohols were also compared to the published data of Laska and co-workers on the psychophysical performance of humans and monkeys for these same odorant pairs. Similar trends were generally observed between the humans, monkeys, and the electronic nose in that discrimination performance increased as the compounds of an odorant pair became more structurally dissimilar. With use of the Fisher linear discriminant algorithm for classification of these test pairs of odorants, the electronic nose exhibited significantly better discriminability than humans or monkeys for the odorant pairs evaluated in this work under the test conditions for which the discriminability was evaluated.

Comparison of the performance of different discriminant algorithms in analyte discrimination tasks using an array of carbon black--polymer composite vapor detectors.

An array of 20 compositionally different carbon black--polymer composite chemiresistor vapor detectors was challenged under laboratory conditions to discriminate between a pair of extremely similar pure analytes (H2O and D2O), compositionally similar mixtures of pairs of compounds, and low concentrations of vapors of similar chemicals. Several discriminant algorithms were utilized, including k nearest neighbors (kNN, with k = 1), linear discriminant analysis (LDA, or Fisher's linear discriminant), quadratic discriminant analysis (QDA), regularized discriminant analysis (RDA, a hybrid of LDA and QDA), partial least squares, and soft independent modeling of class analogy (SIMCA). H2O and D2O were perfectly classified by most of the discriminants when a separate training and test set was used. As expected, discrimination performance decreased as the analyte concentration decreased, and performance decreased as the composition of the analyte mixtures became more similar. RDA was the overall best-performing discriminant, and LDA was the best-performing discriminant that did not require several cross-validations for optimization.

Relationships among resonant frequency changes on a coated quartz crystal microbalance, thickness changes, and resistance responses of polymer-carbon black composite chemiresistors

The relationships among frequency changes on a film-coated quartz crystal microbalance, thickness changes, and dc resistance changes have been investigated for carbon black-insulating polymer composite vapor detectors. Quartz crystal microbalance (QCM) measurements and ellipsometry measurements have been performed simultaneously on polymer films that do not contain carbon black filler to relate the QCM frequency change and the ellipsometrically determined thickness change to the analyte concentration in the vapor phase. In addition, quartz crystal microbalance measurements and dc resistance measurements on carbon black composites of these same polymers have been performed simultaneously to relate the QCM frequency change and dc electrical resistance response to the analyte concentration in the vapor phase. The data indicate that the dc resistance change is directly relatable to the thickness change of the polymers and that a variety of analytes that produce a given thickness change produce a constant resistance change for each member of the test set of polymers investigated in this work.

An investigation of the concentration dependence and response to analyte mixtures of carbon black/insulating organic polymer composite vapor detectors.

The responses relative to an air background of carbon black/polymer composite vapor detectors have been determined as a function of the concentration of a homologous series of alcohols (n-CnH2n+1OH, 1 < or = n < or = 8), a homologous series of alkanes (n-CnH2n+2, 5 < or = n < or = 10 and n = 12, 14), and a set of diverse solvent vapors. In all cases, the steady-state relative differential resistance responses, delta R/Rb, of the carbon black/polymer composite vapor detectors were well-described by a linear relationship with respect to the analyte partial pressure, at least over the tested concentration range (P/P degree = 0.005-0.03, where P degree is the vapor pressure of the analyte). When two vapors in air were simultaneously presented to the detectors, the delta R/Rb response, relative to an air background, was the sum of the delta R/Rb values obtained when each analyte was exposed separately to the carbon black/polymer composite detectors under study. Similarly, when an analyte was exposed to the detectors on top of a background level of another analyte, the delta R/Rb values of the array of detectors were very close to those obtained when the test analyte was exposed to the detectors only in the presence of background air. The initial training requirements from the array response output data of such detectors are minimized because the delta R/Rb response pattern produced by the analyte of concern can be associated uniquely with that odor, under the conditions explored in this work.