Abstract In this study, a novel soft classification scheme is proposed for metal scrap identification with laser-induced breakdown spectroscopy (LIBS) measurements. LIBS provides unique spectra for different metals that can be utilized for classifying metal scraps in real time. Despite its potential, LIBS-based metal classification is not yet fully implemented in practice due to the large shot-to-shot variation and non-linear relationships between spectral intensities and elemental concentrations. Particularly for recycling metal alloys of the same base, learning all candidate types is infeasible, and conventional classification approaches exhibit limited performance in classifying unknown samples of untrained types due to the variability and non-linearity in LIBS measurements. To overcome the limitations of LIBS-based metal scrap classification, the proposed scheme employs pre-cluster-based regression (PCBR) analysis. PCBR takes advantage of the joint relationships between the elemental concentration variations of the pre-clusters (p-clusters), which are pre-determined by prior information of the probability distributions. The variance of the regression of individual p-clusters can be significantly reduced compared to global regression by jointly taking into account common relationships between the elemental concentrations within a particular p-cluster. By combining the layered regression results with their estimated statistics, soft multi-label classification and extraction of the likelihood values of trained classes is possible even for samples of untrained types. For performance evaluation, a list of reference alloys from the National Institute of Standard and Technology (NIST) and Brammer databases was divided into finite sets of p-clusters based on the relationships of elemental concentrations, in particular, four p-clusters for Cu-based alloy tests with Cu and Zn concentrations. Then, PCBR were trained with LIBS-captured spectra of 35 certified reference materials for all and four individual p-clusters. The partial least squares (PLS) and random forest (RF) regression methods were employed, and the root mean square error (RMSE) of the estimation and soft classification measures was investigated. The evaluation results of same-base alloy regression revealed that the proposed PCBR reduced the RMSE of the major element concentration estimation compared to conventional regression schemes. In addition, the accuracy of the soft classification of same-base alloys by PCBR for untrained types was tangibly improved compared to that of prior approaches, such as PLS discriminant analysis and soft independent modeling of class analogy (SIMCA).

中文翻译：

---

使用基于预聚类的回归软分类方案，用于使用激光诱导击穿光谱识别同基合金

摘要 在这项研究中，提出了一种新的软分类方案，用于通过激光诱导击穿光谱 (LIBS) 测量识别金属废料。LIBS 为不同金属提供独特的光谱，可用于实时分类金属废料。尽管具有潜力，但基于 LIBS 的金属分类在实践中尚未完全实施，因为每次发射的变化很大，并且光谱强度与元素浓度之间存在非线性关系。特别是对于回收相同基础的金属合金，学习所有候选类型是不可行的，并且由于 LIBS 测量的可变性和非线性，传统的分类方法在对未经训练类型的未知样本进行分类时表现出有限的性能。为了克服基于 LIBS 的金属废料分类的局限性，提议的方案采用基于预聚类的回归 (PCBR) 分析。PCBR 利用了由概率分布的先验信息预先确定的前簇（p 簇）的元素浓度变化之间的联合关系。通过联合考虑特定 p 簇内元素浓度之间的共同关系，与全局回归相比，单个 p 簇的回归方差可以显着降低。通过将分层回归结果与其估计统计量相结合，即使对于未训练类型的样本，软多标签分类和训练类似然值的提取也是可能的。对于绩效评估，来自美国国家标准与技术研究院 (NIST) 和 Brammer 数据库的参考合金列表根据元素浓度的关系划分为有限的 p 簇集，特别是用于 Cu 基合金测试的四个 p 簇与铜和锌浓度。然后，PCBR 使用 LIBS 捕获的所有和四个 p 簇的 35 种认证参考材料的光谱进行训练。采用偏最小二乘法（PLS）和随机森林（RF）回归方法，研究了估计和软分类措施的均方根误差（RMSE）。同基合金回归的评估结果表明，与传统回归方案相比，所提出的 PCBR 降低了主要元素浓度估计的 RMSE。此外，