Abstract In recent years, pellet has received increasing attention among other biofuels due to its low storage costs and high combustion efficiency. The traceability of pellet quality along the entire supply chain is a critical issue, since fraudulent behaviours, such as the replacement with lower quality pellet, may both cause an economic damage and harm consumers’ health. Traditionally, pellet quality is evaluated through laboratory analysis, which is costly and time-consuming. To overcome these limitations, in this work we define a methodology for quick and low-cost evaluation of pellet quality, which may be used along the entire supply chain. The proposed technique is based on the classification of pellet spectra through machine learning techniques. Spectra are obtained by means of a near-infrared (NIR) spectrophotometer, which is a relatively cheap instrument of small dimensions (even portable) that is suitable for on-site analysis at any phase of the supply chain. We propose two different approaches, namely an automatic classification of pellet, which does not require laboratory analysis, and a semi-automatic approach, that increases the overall accuracy but requires laboratory analysis for uncertainly classified samples. We validate the methodology by performing several experiments on real-world data, by training different machine learning algorithms and evaluating the impact of several transformations introduced to reduce the scattering effect, which is a well-known issue related to NIR data.

中文翻译：

---

通过机器学习技术和近红外光谱预测颗粒质量

摘要 近年来，颗粒由于其低储存成本和高燃烧效率而在其他生物燃料中受到越来越多的关注。整个供应链中颗粒质量的可追溯性是一个关键问题，因为欺诈行为，例如更换质量较低的颗粒，既可能造成经济损失，也可能损害消费者的健康。传统上，颗粒质量是通过实验室分析来评估的，这既昂贵又耗时。为了克服这些限制，在这项工作中，我们定义了一种快速、低成本评估颗粒质量的方法，可用于整个供应链。所提出的技术基于通过机器学习技术对颗粒光谱进行分类。光谱是通过近红外 (NIR) 分光光度计获得的，这是一种相对便宜的小尺寸仪器（甚至便携式），适用于供应链任何​​阶段的现场分析。我们提出了两种不同的方法，即颗粒的自动分类，不需要实验室分析，以及半自动方法，提高整体准确性，但需要对不确定分类的样本进行实验室分析。我们通过对真实世界数据进行多次实验、训练不同的机器学习算法并评估为减少散射效应而引入的几种转换的影响来验证该方法，散射效应是与 NIR 数据相关的一个众所周知的问题。即颗粒的自动分类，不需要实验室分析，以及半自动方法，提高整体准确性，但需要对不确定分类的样品进行实验室分析。我们通过对真实世界数据进行多次实验、训练不同的机器学习算法并评估为减少散射效应而引入的几种转换的影响来验证该方法，散射效应是与 NIR 数据相关的一个众所周知的问题。即颗粒的自动分类，不需要实验室分析，以及半自动方法，提高整体准确性，但需要对不确定分类的样品进行实验室分析。我们通过对真实世界数据进行多次实验、训练不同的机器学习算法并评估为减少散射效应而引入的几种转换的影响来验证该方法，散射效应是与 NIR 数据相关的一个众所周知的问题。