Spectrometers are the cornerstone of analytical chemistry. Recent advances in microoptics manufacturing provide lightweight and portable alternatives to traditional spectrometers. In this dissertation, we developed a spectrometer based on Fabry-Perot interferometers (FPIs). A FPI is a tunable (it can only scan one wavelength at a time) optical filter. However, compared to its traditional counterparts such as FTIR (Fourier transform infrared spectroscopy), FPIs provide lower resolution and lower signal-noiseratio (SNR). Wavelength selection can help alleviate these drawbacks. Eliminating uninformative wavelengths not only speeds up the sensing process but also helps improve accuracy by avoiding nonlinearity and noise. Traditional wavelength selection algorithms follow a training-validation process, and thus they are only optimal for the target analyte. However, for chemical identification, the identities are unknown. To address the above issue, this dissertation proposes active sensing algorithms that select wavelengths online while sensing. These algorithms are able to generate analytedependent wavelengths. We envision this algorithm deployed on a portable chemical gas platform that has low-cost sensors and limited computation resources. We develop three algorithms focusing on three different aspects of the chemical identification problems. First, we consider the problem of single chemical identification. We formulate the problem as a typical classification problem where each chemical is considered as a distinct class. We use Bayesian risk as the utility function for wavelength selection, which calculates the misclassification cost between classes (chemicals), and we select the wavelength with the maximum reduction in the risk. We evaluate this approach on both synthesized and experimental data. The results suggest that active sensing outperforms the passive method, especially in a noisy environment. Second, we consider the problem of chemical mixture identification. Since the number of potential chemical mixtures grows exponentially as the number of components increases, it is intractable to formulate all potential mixtures as classes. To circumvent combinatorial explosion, we developed a multi-modal non-negative least squares (MMNNLS) method that searches multiple near-optimal solutions as an approximation of all the solutions. We project the solutions onto spectral space, calculate the variance of the projected spectra at each wavelength, and select the next wavelength using the variance as the guidance. We validate this approach on synthesized and experimental data. The results suggest that active approaches are superior to their passive counterparts especially when the condition number of the mixture grows larger (the analytes consist of more components, or the constituent spectra are very similar to each other). Third, we consider improving the computational speed for chemical mixture identification. MM-NNLS scales poorly as the chemical mixture becomes more complex. Therefore, we develop a wavelength selection method based on Gaussian process regression (GPR). GPR aims to reconstruct the spectrum rather than solving the mixture problem, thus, its computational cost is a function of the number of wavelengths. We evaluate the approach on both synthesized and experimental data. The results again demonstrate more accurate and robust performance in contrast to passive algorithms.
Spectrometers are the cornerstone of analytical chemistry. Recent advances in microoptics manufacturing provide lightweight and portable alternatives to traditional spectrometers. In this dissertation, we developed a spectrometer based on Fabry-Perot interferometers (FPIs). A FPI is a tunable (it can only scan one wavelength at a time) optical filter. However, compared to its traditional counterparts such as FTIR (Fourier transform infrared spectroscopy), FPIs provide lower resolution and lower signal-noiseratio (SNR). Wavelength selection can help alleviate these drawbacks. Eliminating uninformative wavelengths not only speeds up the sensing process but also helps improve accuracy by avoiding nonlinearity and noise. Traditional wavelength selection algorithms follow a training-validation process, and thus they are only optimal for the target analyte. However, for chemical identification, the identities are unknown.
To address the above issue, this dissertation proposes active sensing algorithms that select wavelengths online while sensing. These algorithms are able to generate analytedependent wavelengths. We envision this algorithm deployed on a portable chemical gas platform that has low-cost sensors and limited computation resources. We develop three algorithms focusing on three different aspects of the chemical identification problems.
First, we consider the problem of single chemical identification. We formulate the problem as a typical classification problem where each chemical is considered as a distinct class. We use Bayesian risk as the utility function for wavelength selection, which calculates the misclassification cost between classes (chemicals), and we select the wavelength with the maximum reduction in the risk. We evaluate this approach on both synthesized and experimental data. The results suggest that active sensing outperforms the passive method, especially in a noisy environment.
Second, we consider the problem of chemical mixture identification. Since the number of potential chemical mixtures grows exponentially as the number of components increases, it is intractable to formulate all potential mixtures as classes. To circumvent combinatorial explosion, we developed a multi-modal non-negative least squares (MMNNLS) method that searches multiple near-optimal solutions as an approximation of all the solutions. We project the solutions onto spectral space, calculate the variance of the projected spectra at each wavelength, and select the next wavelength using the variance as the guidance. We validate this approach on synthesized and experimental data. The results suggest that active approaches are superior to their passive counterparts especially when the condition number of the mixture grows larger (the analytes consist of more components, or the constituent spectra are very similar to each other).
Third, we consider improving the computational speed for chemical mixture identification. MM-NNLS scales poorly as the chemical mixture becomes more complex. Therefore, we develop a wavelength selection method based on Gaussian process regression (GPR). GPR aims to reconstruct the spectrum rather than solving the mixture problem, thus, its computational cost is a function of the number of wavelengths. We evaluate the approach on both synthesized and experimental data. The results again demonstrate more accurate and robust performance in contrast to passive algorithms.