Bottled Water Classification Using Spectroscopy and Deep Belief Network

Neural networks and their variations have been widely applied to classification problems. Among them, the classification of mixtures using neural networks has been investigated by experts from both signal processing and chemical fields. However, the extracted features of beverages such as water, soft drinks, and alcohols, are usually highly collinear. Therefore, traditional methods or simple artificial neural networks may not yield the most accurate results. In this article, several steps were introduced to improve the classification accuracy. Firstly, proper system design and measurement methods were used to mitigate errors. Secondly, deep belief network was applied to adapt to the non-linearity and collinear of features. Finally, suitable learning method was selected. As a result, our system could classify samples from various bottled water branches with nearly 100% accuracy.