Automatic signal enhancements for spectroscopic measurements

Abstract

Material analysis is a diverse scientific field often requiring a high degree of expert knowledge to achieve satisfactory results. Therefore automation of process steps is essential to allow untrained users access to such methods and reduce manual workload for experts. This requires robust corrective measures that can be applied to measured data containing severe distortions with the goal to enable automatic identification of compounds in so called real world spectra, meaning spectra that were recorded under unknown and sometimes adverse conditions outside a laboratory environment. Extracting information from chemical substances in order to classify or even identify their components is a critical task for almost every application that is aimed to detect and/or identify unknown or hidden compounds. Methods to achieve that goal employ spectroscopic techniques that vary in instrumentation and underlying physical principals. The analysis of the collected data is often non-trivial and interpretation requires experts with years of training to extract reliable information from measurements. This thesis introduces new techniques that reliably reduce distortions in measurements without the need for fixed models, training sets or repeated measuring. The development of these new techniques was motivated and accompanied by discussions with practical physicists and chemists who expressed their discontent with available methods. The focus are measurements that are acquired outside of controlled laboratory environments requiring methods to be highly robust against a wide array of possible distortions without complex control mechanisms that require adjusted to every measurement. The work introduces advancements in noise suppression via wavelet transform that focus on the automatic reduction of noise of varying intensity within a single measurement. Combined with a new baseline estimation technique based on adaptive regression that is highly robust against distortions and requires only minimal information about expected signal characteristics this allows automatic processing and reliable data extraction in scenarios where previously existing methods failed to achieve satisfactory results.