基于激光诱导击穿光谱技术的废旧金属分类辨识

废旧金属回收是工业中金属的重要来源之一,是发展循环经济的重要内容。废旧金属产量巨大,通常表面覆盖杂质,凹凸不平,因此对分类方法的判别能力和计算速度提出较高要求。采用激光诱导击穿光谱技术研究分析了7种废旧金属分类识别问题,包括生铝、熟铝、镁、不锈钢、锌、黄铜与红铜。为了符合现场应用条件,实验中每个样本点只激发一次建立并分析了多种分类模型,包括支持向量机(SVM)分类模型,主成分分析方法结合支持向量机(PCA-SVM)分类模型,遗传算法结合支持向量机(GA-SVM)分类模型,遗传算法选择特征光谱结合主成分分析方法和支持向量机(GA-PCA-SVM)分类模型,以及遗传算法选择特征光谱结合主成分分析方法和人工神经网络(GA-PCA-BP)分类模型。通过遗传算法选取包含丰富特征的谱段组合与支持向量机方法相结合建立GA-SVM分类模型,490组验证样本分类准确率为93.47%。为了判断该模型的鲁棒性,对一批新样品,在自研的分选系统上以传送带匀速运行的方式进行测试,获取的750组光谱测试数据,分类准确率为88.27%,证明了该分类模型具有很好的移植性和应用性。

Classification and identification of scrap metals based on laser-induced breakdown spectroscopy

Scrap metal recycling was one of the important sources of metals in industry and an important part of developing recycling economy. The production of scrap metal was huge. Usually the surface was covered with impurity and the shape was irregularity. The classification method with strong discriminant ability and fast calculation speed was required. Seven kinds of classification and identification problems for scrap metal, including cast aluminum, wrought aluminum, magnesium, stainless steel, zinc, brass and red copper, were studied and analyzed by means of laser-induced breakdown spectroscopy (LIBS). In order to meet the field application conditions, samples were classified based on optical emission following a single laser pulse in the experiment. Support vector machine (SVM) model, support vector machine combined with principal component analysis (PCA-SVM) model, support vector machine combined with genetic algorithm (GA-SVM) model, support vector machine combined with genetic algorithm and principal component analysis (GA-PCA-SVM) model and back-propagation neural network combined with genetic algorithm and principal component analysis (GA-PCA-BP) model were established and compared. The characteristic spectral segments with ample feature information and less interference were extracted by genetic algorithm. A novel method combining genetic algorithm and support vector machine was proposed, resulting the percentage of 490 validation samples correctly identified was 93.47%. In order to assess the robustness and versatility of the model, a group of new samples were tested on a self-developed sorting system with a conveyor belt running at a uniform speed. The 750 test samples could be classified with 88.27% correct assignments, which proved that the GA-SVM model had good portability and applicability.