| Περίληψη: | Since the development of the laser, the scientific landscape has altered dramatically, resulting in novel experimental methodologies and a variety of technological spinoffs. One field-deployable analytical approach is Laser-Induced Breakdown Spectroscopy (LIBS), sometimes termed Laser-Induced Plasma Spectroscopy (LIPS). LIBS employs a high-energy pulsed laser to produce a plasma that vaporizes a sample. Spectral characteristics produced by excited species, i.e., atoms, ions and molecules, are utilized to gather quantitative and qualitative analytical information about the sample. Due to the fact that the optical emission from the plasma comprises the spectral signatures of all the elements present in the sample material, the elemental composition of the investigated sample may be rapidly determined by observing its LIBS spectra. LIBS offers the remarkable capability to perform multielement real-time analysis, which is not achievable with other traditional techniques. However, due to its relatively low sensitivity, the detection of trace elements remains difficult, which is a key disadvantage of this approach.
The first record of a laser produced plasma was reported almost immediately after the invention of the laser, while within the last three decades a considerable number of LIBS-related applications have been proposed and realized. Moreover, statistical approaches for analyzing LIBS spectra are being improved, and commercialized LIBS instruments are available, while theoretical and computational models of plasma formation and expansion have been thoroughly examined through experiments. In the past decade, chemometric and machine learning tools for analyzing LIBS spectroscopic data have reignited scientific interest in LIBS-related applications, because of the huge datasets with thousands of variables provided in extremely quick acquisition times, compared to other spectroscopic techniques. An emerging and challenging application is the analysis of foodstuff, mainly as a quality assurance method.
In this thesis the application of LIBS, assisted by machine learning, to the analysis of olive oils is investigated. The classification of different olive oils is performed based either on their geographical or cultivar origins. Different machine learning algorithms are tested and the analysis of the LIBS data provides insight into the spectral features that are most important for the successful classification. Moreover, a direct comparison of LIBS with absorption spectroscopy is performed, and the subsequent fusion of the different spectroscopic data is performed to enhance the classification accuracy. In that spirit, this work demonstrates the enhanced capabilities of LIBS for the analysis of foodstuff, as a tool proposed for the quality assessment of olive oil products.
Chapter 1 of the present thesis provides an overview of the current state-of-the-art concerning the application of Laser Induced Breakdown Spectroscopy for the analysis of foodstuff, in general. Special emphasis is given to several foods of interest, such as olive oil, honey and milk. The principles of LIBS technique are discussed and described in detail. A summary of the machine learning methods used in LIBS analyses is given, with emphasis on the validation of the predictive models, as well. Then, specific food science applications of LIBS are presented thoroughly.
In Chapter 2 LIBS spectra from a total of 139 extra virgin and virgin olive oil samples (EVOOs and VOOs) are classified based on the samples’ geographical origins. Different machine learning algorithms are employed. These are Linear Discriminant Analysis (LDA), Extremely Randomized Classification Trees (ERTC) and eXtreme Gradient Boosting (XGBoost) and their classification performance is assessed. Additionally, the spectral features’ importance on the classification was calculated and the most important ones were identified.
In Chapter 3 a comparative study between LIBS and UV-Vis absorption spectroscopy is presented and performed, regarding the classification of olive oils based on their geographical origins. Both LIBS and absorption spectra were initially preprocessed by means of Principal Component Analysis (PCA) and were subsequently used for the construction of predictive models, employing LDA and Support Vector Machines (SVM).
Following, in Chapter 4, extra virgin olive oils are discriminated based on their cultivar origin. In continuation to Chapter 3, LIBS and absorption spectra of the samples are classified by employing LDA and Gradient Boosting algorithms and the subsequent fusion of the two different origins spectroscopic data, i.e., the emission and the absorption spectra, is proposed as an efficient strategy for predicting the cultivar origin of olive oils.
|
|---|
