Learning algorithms for identification of whisky using portable Raman spectroscopy

Abstract

Reliable identification of high-value products such as whisky is vital due to rising issues of brand substitution and quality control in the industry. We have developed a novel framework that can perform whisky analysis directly from raw spectral data with no human intervention by integrating machine learning models with a portable Raman device. We demonstrate that machine learning models can achieve over 99% accuracy in brand or product identification across twenty-eight commercial samples. To demonstrate the flexibility of this approach, we utilized the same algorithms to quantify ethanol concentrations, as well as measuring methanol levels in spiked whisky samples. To demonstrate the potential use of these algorithms in a real-world environment we tested our algorithms on spectral measurements performed through the original whisky bottle. Through the bottle measurements are facilitated by a beam geometry hitherto not applied to whisky brand identification in conjunction with machine learning. Removing the need for decanting greatly enhances the practicality and commercial potential of this technique, enabling its use in detecting counterfeit or adulterated spirits and other high-value liquids. The techniques established in this paper aim to function as a rapid and non-destructive initial screening mechanism for detecting falsified and tampered spirits, complementing more comprehensive and stringent analytical methods.

Keywords: Brand identification; Machine learning; Raman spectroscopy; Whisky.

Graphical abstract

Fig. 1

Brand identification: (a) Deep learning model accuracies per epoch on the training set. (b) Accuracy of deep learning model on the test set and training time. The blue and red dots represent the test accuracy and training time results, respectively. Each dot represents an increase in accuracy and training time as the epochs increase by 100, 200, 500, 1000, and 2000. Shaded grey areas show where the model has > 96% accuracy. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 2

Ethanol content prediction: (a) Using PCA + HPM with 2000 epochs (R_P2 = 0.998 and RMSE_P = 0.25%). (b) Using PCA + FCN with 200 epochs applied to a dataset with previously unseen whisky brands (red dot) and gin samples (green x) (R_U2 = 0.863 and RMSE_U = 2.47% for the test set). Methanol content prediction: (c) Using PCA + HPM with 1000 epochs. The spectral data of Talisker, Cragganmore, 40% ethanol/water, Caol Ila, and Cynelish spiked with methanol were divided into training (60%), validation (20%), and test (20%) sets. (d) Using PCA + HPM with 1000 epochs. The spectral data of Talisker, Cragganmore, and 40% ethanol/water spiked with methanol were split into training (80%) and validation (20%) sets. The spectra of Caol Ila and Cynelish spiked with methanol were only used as a test set. The blue and red dots represent the results of the training and test sets, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 3

Brand identification accuracy of conventional machine learning models. The blue and red dots represent the results of the training and test sets, respectively, with PCA. The dashed lines show the respective accuracy levels without PCA. The red and blue dashed lines overlap each other on the LDA and ANN graphs due to the identical accuracy they achieved on both the training and test sets. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 4

Methanol content prediction using a PLSR model (PLSR7). The spectral data of Talisker, Cragganmore, and 40% ethanol/water spiked with methanol were used as a training set. The spectra of Caol Ila and Cynelish spiked with methanol were only used as a test set. The blue and red dots represent the prediction results on the training and unseen test sets, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)