激光诱导击穿光谱数据特征自动提取方法研究

CORC > 沈阳自动化研究所 > 中国科学院沈阳自动化研究所 > 信息服务与智能控制技术研究室

题名	激光诱导击穿光谱数据特征自动提取方法研究
作者	孔海洋
学位类别	博士
答辩日期	2015-05-28
授予单位	中国科学院沈阳自动化研究所
授予地点	中国科学院沈阳自动化研究所
导师	胡静涛 ; 孙兰香
关键词	激光诱导击穿光谱特征选择遗传算法主成分分析人工神经网络偏最小二乘法
其他题名	Automatic Methods to Extract Spectral Features of Laser-Induced Breakdown Spectroscopy
学位专业	机械电子工程
中文摘要	自20世纪中叶诞生，尤其是80年代被以LIBS简称以来，激光诱导击穿光谱技术以其具有的种种优良特性，如可实时在线分析，对样品没有或只有较少的损伤，等等，获得了科研人员的广泛关注，被视为光谱学领域的“明日之星”。进入21世纪以来，科研人员将激光诱导击穿光谱的研究重点，从对激光诱导等离子体本身的特性的基础理论研究和对激光诱导击穿光谱应用范围的探索研究上，转移到了制约激光诱导击穿光谱进一步发展和推广较低的精确度和准确度上，致力于提升激光诱导击穿光谱的硬件设备，开发具有高精准度的分析算法，提高激光诱导击穿光谱的分析结果。在这一过程中，软件算法上，在其他光谱学领域中广泛应用并取得很好结果的化学计量学方法，如主成分分析、人工神经网络和偏最小二乘回归等，被引入激光诱导击穿光谱领域，科研人员籍此开发了许多定性定量分析模型，对激光光谱进行分析获得了很好的结果。硬件设备上，激光器的发展使得由此诱导产生的等离子特性更为稳定，为激光光谱的数据稳定性提供了保障。而光谱仪的更新换代使得由此获得的光谱数据的分辨率大幅提升，在为光谱分析提供便利的同时，也使得光谱数据量越来越大，给后续的保存和处理都带来了一定的挑战。与此同时，科研人员意识到在软件模型的建立中，对光谱数据进行特征提取十分必要。对于传统的定标法或内标法，需要从原始光谱中选择恰当的特征谱线；对于诸如人工神经网络和偏最小二乘回归的化学计量学方法，则需要从原始光谱中提取光谱特征。在提取有效信息的同时，去除其中的噪声和干扰，提高原始数据的信噪比，在此过程中，大幅压缩数据量，降低由此建立的分析模型的复杂度，减小模型建立过程中发生过拟合的可能性。本文在这个背景和前提下，对激光诱导击穿光谱的光谱数据特征自动提取方法进行了研究，具体的研究内容和由此获得的成果主要包括：(1) 对比了三种不同的LIBS光谱特征提取方法，即全部原始光谱、全部光谱的峰值和手动选择的包含丰富信息的光谱谱段，并比较了以这三种特征提取方法提取的光谱特征作为数据来源，对其进行主成分分析提取主成分作为输入，来训练人工神经网络而对光谱数据进行分类研究的结果。研究发现，原始光谱的结果最差，使用提取出的峰值结果较好，而以手动选择的光谱谱段作为输入时，由于噪声和干扰被大量丢弃，有效信息却没有太大损耗，信噪比获得提升，所以能够获得最好的分类结果。(2) 提出了一种面向分类应用的LIBS光谱特征自动提取方法。该方法使用遗传算法从原始光谱选择谱段，以主成分分析将对应于遗传算法每个个体的光谱谱段进行数据压缩并提取其主成分，并将其作为输入建立人工神经网络模型对光谱数据进行定性分类，以分类结果的优劣度作为遗传算法的适应度函数，如此迭代循环直至选出最优谱段。其中详细讨论了从原始光谱中选择光谱谱段的三种方式，即定宽谱段、变宽谱段和光谱子段组合。定宽谱段优化简单，对光谱数据的处理也简单，可以此选择具有较窄带宽的光谱仪而对光谱仪的成本进行压缩；变宽谱段选出的数据量较大，虽然结果很好，但优点并不突出，只适合定宽谱段不能满足要求的情况；而紧密贴合激光诱导击穿光谱本身特征的光谱子段组合方式则能够以最小的数据量获得最好的结果，对光谱数据压缩有重要参考意义。(3) 针对内标法提出了一种面向单变量定量分析的从原始LIBS光谱中自动选择分析线和参考线的方法。内标法是激光诱导击穿光谱定量分析中的经典方法，但在具体的操作中，需要使用者为其选择合适的分析线和参考线，而原始光谱数据量庞杂，不同元素往往有较多的特征谱线，如果研究者手动选择，既需要丰富的先验知识，还需要耗费大量的时间和精力。本文基于遗传算法提出的方法，能够无需人工参与全自动地从原始光谱中选择最优分析线和参考线，为内标法的应用提供便利，获得了很好的分析结果。而且，在与其他研究者手动选择的谱线进行的对比研究中也发现，由此方法自动选择的谱线获得的结果要比研究者精心选择的谱线更好。 (4) 提出了一种面向多元定量分析的LIBS光谱数据特征自动提取方法。该特征提取方法使用遗传算法从原始光谱中优化选择光谱谱段，以遗传算法每个个体所对应的光谱谱段建立PLS定量分析模型，并以定量分析结果的衡量指标作为遗传算法的适应度函数，循环迭代优选最优谱段。该方法也讨论了定宽谱段、变宽谱段和光谱子段组合三种方式，发现以经过优选后的数据建立PLS模型对原始光谱数据进行定量分析，可以得到相比使用全部光谱更好的结果。而这三种特征提取方式中，又以按照LIBS光谱本身特性而设计的光谱子段组合的方式最能提取原始光谱的特征信息，而获得最好的定量分析结果。
索取号	TP391.41/K48/2015
英文摘要	Laser-Induced Breakdown Spectroscopy (LIBS), which has various distinguishing features, such as, it can implement real-time analysis of samples in situ, with no or little damage to samples, has been paid wide attention by an increasing number of spectroscopist since it originally emerged in the 1960s and was referred to LIBS from the 1980s especially and it has been considered as a rising future star of spectroscopy.Since the 21st century, spectroscopist have transferred the focus of spectroscopy research on LIBS from the basic theoretical research of characteristics of laser-induced plasma and the exploration of the application scope of LIBS to the low precision and accuracy of LIBS analysis which have constrained the rapid development and extension of LIBS. An increasing number of researchers are committed to enhance the hardware of LIBS and develop novel analytical algorithms with high precision and accuracy to improve the analytical results of LIBS.As for developing analytical algorithms, various chemometric methods which have obtained achievements in other fields of spectroscopy were introduced to LIBS, such as Principal Component Analysis (PCA), Artificial Neural Networks (ANN) and Partial Least Squares (PLS). Researchers of LIBS have developed a number of qualitative and quantitative analytical models based on these chemometric methods to analyze LIBS and achieved good results. When it comes to hardware, the development of lasers, which makes the characteristics of laser-induced plasma more and more stable, guarantees the stability of spectral data of LIBS. Besides, the spectrometers are significantly improved and the resolution of the spectral data is correspondingly becoming higher and higher. Although this provides convenience for spectral analysis, it will be challenging to save and process the increasing amount of sophisticated spectral data.In the mean time, many spectroscopist have realized that it is necessary to select wavelengths and extract characteristics from the original spectra while developing the analytical models. It is vital important to select appropriate characteristic lines from the original spectra when using conventional calibration methods and internal standard methods and it is also necessary to select wavelengths and extract spectral features when using the chemometric methods such as ANN and PLS. The procedure of extracting features and selecting wavelengths will remove the noises and interferences among the original spectra while extracting useful information and compress the amount and improve the Signal to Noise Ratio (SNR) of the spectral data substantially, so the complexity and the probability of over-fitting of the corresponding analytical model established based on the extracted features will be reduced significantly. We have studied the automatic methods to select wavelengths and extract features from the original spectra of LIBS based on the context and the specific contents and obtained results can be concluded as follows:(1) Three different methods to extract spectral features, the whole spectra, all spectral peaks and the intensive spectral partitions, were compared to select wavelengths from the original spectra and the data of extracted features was utilized as the data resources to implement PCA to extract the Principal Components (PCs) which were used as the input of ANN to train a classification model to classify the spectra of steels. The results show that, the result obtained using the original whole spectra is the worst and the result obtained using all the peaks of spectra is better than that obtained by the whole spectra, and the best classification result can be achieved when using the intensive spectral partitions manually selected since a large amount of noises and interferences was discarded while the efficient information remained stable in the process of extracting spectral features.(2) An automatic method of GA-PCA-ANN based on Genetic Algorithm (GA) to extract spectral features of LIBS for classificational applications was proposed. The method used GA to select and optimize spectral partitions from the original spectra and used PCA to extract PCs of the selected partitions and then an ANN classification model was established with the PCs as the input to qualify the spectral data. The evaluation of classificational results was used as the fitness function of GA and the optimal partitions will be selected out in this loop procedure. Three methods to select spectral partitions from the original spectra, selecting fixed-width partition, non-fixed-width partition and selecting sub-partitions to combine a combination, were discussed and compared. The results show that, it is the least complicated to optimize fixed-width partitions from the original spectra and the procession of data is also simple, and therefore we can reduce the cost of spectrometers using this method. A better result can be obtained by selecting non-fixed-width partitions with the largest amount of spectral data selected, however, the advantage is not prominent and it is only suitable when the fixed-width partition cannot meet the requirements. Selecting sub-partitions to combine a combination can achieve the best result with the minimal amount of data since this method was developed considering the characteristics of LIBS considerably.(3) An automatic method to select analytical lines and reference lines for internal standard methods based on Genetic Algorithm was proposed. Internal standard method is a classical analytical method to quantify LIBS and it needs the researchers to select appropriate analytical lines and reference lines in the specific operations. However, it will require the researchers to have rich prior knowledge and consume a lot of time and effort to select proper lines since the original spectral data is extremely complex and tremendous and there are numerous characteristic lines for every element. We developed an automatic method to select the optimum analytical lines and reference lines from the original spectra based on GA and the method was used to select lines for internal standard methods and a good result was achieved correspondingly. Further, in the comparative study, we have found that the result obtained using the lines automatically selected by our method is better than that obtained using the lines manually selected by other spectroscopist.(4) An automatic method of GA-PLS based on GA to extract spectral features of LIBS for quantitative analysis by PLS was presented. The spectra corresponding to every individual of GA were used to establish PLS models to quantify spectra and the evaluation of the quantitative results was used as the fitness function of GA and the procedure was circulated to select optimal partition. This method selects fixed-width partition, non-fixed-partition and sub-partitions to combine a combination using GA to extract spectral features respectively. Then the extracted features were used to establish PLS models to quantify the spectral data. The PLS models by the selected spectral data with less data can achieve better results than the model established by the original spectra. Besides, the PLS model established with data selected by the method of selecting sub-partitions to combine a combination can obtain the best quantitative result since this method was designed according to the characteristics of LIBS.
语种	中文
产权排序	1
页码	115页
内容类型	学位论文
源URL	[http://ir.sia.ac.cn/handle/173321/16753]
专题	沈阳自动化研究所_信息服务与智能控制技术研究室
推荐引用方式 GB/T 7714	孔海洋. 激光诱导击穿光谱数据特征自动提取方法研究[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2015.