A low-cost infrared sensor that uses room temperature pyroelectric detectors integrated with bandpass filters to provide low-resolution spectral scans of the absorption characteristics of hazardous chemicals was developed for fixed security applications. The sensor provides fast (1 s) and continuous monitoring, detection, and identification capabilities. A unique detection and identification algorithm that uses non-linear computation techniques to account for the exponential nature of optical absorption was developed. Chemical detection and identification is achieved by matching the recorded sensor response vector to an updatable signature library that currently includes the signatures of 14 chemicals. The sensor and algorithm were tested by introducing methanol vapor at optical depths between 225 - 270 ppm-m. Using 1 s signal samples obtained during approximately 20 min. test, resulted in no false positive alarms and 3.4% of false negatives. All false negatives were shown to be due to misidentification of methanol as isopropanol, which is spectrally similar to methanol. By grouping isopropanol with methanol the rate of false negatives was reduced to 0%. Results of the same test using a 30 s signal integration time resulted in no false positive and no false negative alarms.
A 16-channel, cross-reactive remote infrared chemical sensor for detection of toxic industrial chemicals in fixed-location applications is being developed. The outputs of the 16 channels, uncooled pyroelectric detectors fitted with infrared bandpass filters, can be viewed as a coarse spectrum of the chemical(s) in the field of view. This spectrum must be unmixed, wherein the identity and optical depth of the chemical(s) are estimated by processing the spectrum with a library of known signatures for the chemical(s) of interest.
Several unmixing methods are presented, including enhancements to linear projection methods, parameterization (curve fitting) of the system response, and non-linear, iterative techniques. It is found that linear methods and simple curve parameterizations produce excessive unmixing errors. Higher-order parameterization and iterative methods provide much better estimates, with the latter being more computationally intensive. The suitability of the methods for the application at hand is discussed.