Heterogeneous ensembles and time series classification techniques for the non-invasive authentication of spirits

Large, James (2022) Heterogeneous ensembles and time series classification techniques for the non-invasive authentication of spirits. Doctoral thesis, University of East Anglia.

[thumbnail of 2022LargeJPhD (1).pdf]
Preview
PDF
Download (19MB) | Preview

Abstract

Spirits are a prime target for fraudulent activity. Particular brands, production processes, and other factors such as age can carry high value, and leave space for mimicry. Further, the improper production of spirits, either maliciously or through negligence, can result in harmful substances being sold for consumption. Lastly, genuine spirits producers themselves must ensure the quality and standardisation of their products before sale. Authenticating spirits can be a time consuming and destructive process, requiring sealed bottles to be opened for access to the product.

It is therefore desirable to have a fast, non-invasive means of indicating the authenticity, safety, and correctness of spirits. We advance and prototype such a system based on near infrared spectroscopy, and generate datasets for the detection of correct alcohol concentrations in synthesised spirits, for the presence of methanol in genuine spirits, and for the distinction of particular genuine products in a given bottle.

The standard chemometric pipelines for the analysis of spectra involve smoothing of the signal, standardising for global intensity, possible dimensionality reduction, and some form of least squares regression. This has decades of proof behind it, and works under the assumptions of clean signal gathering, potentially the separation of sample and particular substance of interest, and the generally linear relationship of light received/blocked and the analyte’s contents. In the proposed system, at least one of these assumptions must be violated.

We therefore investigate the use of modern classification techniques to overcome these challenges. In particular, we investigate and develop ensemble methods and time series classification algorithms. Our first hypothesis is that algorithms which consider the ordered nature of the wavelength features, as opposed to treating the spectra effectively as tabular data, can better handle the structural changes brought about by different bottle and environmental characteristics. The second is that ensembling heterogeneous classifiers is the best initial technique for a new data science problem, but should in particular be helpful for the spirit authentication problem, where different classifiers may be able to correct for different defects in the data.

In initial investigations on datasets of synthesised alcohol solutions and different products, we prove the feasibility of the authentication system to make at least indicative predictions of authenticity, but find that it lacks the precision and accuracy needed for anything more than indicative results. Following this, we propose a novel heterogeneous ensembling scheme, CAWPE, and perform a large scale evaluation on public archives to prove its efficacy. We then outline improvements in the time series classification space that lead to the state of the art meta-ensemble HIVECOTE 2.0, which makes use of CAWPE. We lastly apply the developed techniques to a final dataset on methanol concentration detection. We find that the proposed system can classify methanol concentration in arbitrary spirits and bottles from ten possible values, containing as little as 0.25%, to an accuracy of 0.921. We further conclude that while heterogeneously ensembling tabular classifiers does improve the authentication of spirits from spectra, time series classification methods confer no particular advantage beyond tabular methods.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Chris White
Date Deposited: 16 Aug 2022 08:17
Last Modified: 16 Aug 2022 08:17
URI: https://ueaeprints.uea.ac.uk/id/eprint/87284
DOI:

Actions (login required)

View Item View Item