液态钢成分在线检测的LIBS信号处理技术研究
文献类型：学位论文
| 作者 | 张博
|
| 学位类别 | 博士 |
| 答辩日期 | 2014-11-28 |
| 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 | 于海斌 ; 孙兰香 |
| 关键词 | 激光诱导击穿光谱 液态钢 小波分析 重叠峰解析 |
| 其他题名 | Research on LIBS Signal Processing for Online Component Detection of Molten Steel |
| 学位专业 | 机械电子工程 |
| 中文摘要 | 钢铁行业是我国经济建设重要的支柱产业，并作为流程工业的典型，在冶炼过程中对钢铁产品质量的及时检测与各元素成分组成的控制具有重要意义。传统的检测方法大多需要对样品的预处理操作，耗时较长，导致了对生产过程中的产品信息反馈不及时、产品质量控制不充分。激光诱导击穿光谱技术是一种热门的元素分析技术，与传统的检测方法相比，具有无需制备样品、可实现在线、快速、远程分析的特点，在钢铁冶炼分析上的应用前景广阔。本文以钢铁冶炼过程中的液态钢的在线成分分析为应用背景，以提高激光诱导击穿光谱的定量分析性能为目标，展开了对激光诱导击穿光谱的理论和应用研究。本文研究内容包括：LIBS系统噪声分析及小波阈值降噪的研究、小波阈值降噪的双阈值优化及灵敏度进一步提升的研究、基于连续小波变换的光谱信号的谱峰信号识别方法以及基于曲线拟合方法解析重叠峰信号的初始值选择策略。具体的研究内容和创新性成果概括如下：(1) 针对液态钢的冶炼环境恶劣，LIBS系统噪声干扰大的影响，分析了LIBS检测系统的噪声来源，利用小波阈值降噪对噪声进行移除，提出了一种具有分解层数优化的小波阈值降噪方法。该方法通过考察白噪声在不同分解层数上的方差变化，推导出其在不同分解层数上熵变化趋势。然后，结合含噪信号的小波系数能量计算其在不同分解层数上的熵值，并利用白噪声的熵值进行比较，通过考察这两种熵值的变化最终确定最佳的分解层数。本方法相比于传统确定最佳分解层数的方法具有更好的预见性，计算结果可靠准确，对降低噪声对定量分析的影响起着重要作用，提高了LIBS信号的重复性。此外，该方法不仅可用于确定小波阈值降噪的最佳分解层数，而且还可以确定其他类型的小波降噪，比如模极大值方法，亦可推广至其他的光谱分析技术上。(2) 为了进一步提升液态钢LIBS信号检测的灵敏度，提出了一种双阈值优化的小波阈值降噪方法。该方法首先考察了阈值函数与阈值之间的关系，通过建立新的上阈值校正模型利用灰色系统理论进行校正；然后，根据已经完成校正的新的上阈值建立新的下阈值模型，并利用模糊贴近度完成新的下阈值模型的刻画，并最终完成双阈值校正。利用本方法提出的双阈值进行小波阈值降噪不仅可以提高信噪比，而且可以较好的保留边缘信息，更为重要的是降低了最低检出限，提升了定量分析的灵敏度。(3) 针对液态钢LIBS信号的特征谱峰数量庞大，传统方法的谱峰识别准确度不高的问题，提出了运用连续小波变换识别液态钢LIBS信号谱峰的方法。该方法通过寻找在各个尺度下的小波系数的极大值来建立小波脊线，并最终利用小波脊线来识别光谱的谱峰信号。通过与传统的识别谱峰方法相比较，本方法具有更好的灵敏性与较低的漏检率； (4) 针对液态钢中的多种组分响应混杂在一起形成重叠峰的现象，研究了液态钢LIBS信号的重叠峰解析的常用方法，并基于曲线拟合来解析重叠峰的方法提出了初始值选择策略。为了解决曲线拟合解析重叠峰的多解性问题，首先利用连续小波变换计算近似导数来确定子峰的数目与位置的信息，但这两个参数并不进行最优化算法的迭代，降低了拟合参数的个数；然后，基于洛伦兹函数为模型，通过考察该模型的分数阶导数的特征点与导数阶数之间的关系，建立了子峰峰高与半峰宽的参数估计器，并利用该参数估计器计算得到的值作为最优化算法的初始值进行迭代，并将极大地缩小最优化算法的搜索范围。通过曲线拟合方法解析重叠峰，可以将隐含在重叠峰内的子峰进行重建，在定量分析上可以降低误差并且提高灵敏度。 |
| 索取号 | O433.54/Z31/2014 |
| 英文摘要 | The steel industry is one of pillar industries in national economy. As one of important flow industry, the quality detection in time of steel product and the composition control have great significance during melting. Most traditional detection methods need to pre-process the samples and consume more time, which results in delayed feedback of product information and inadequate control of product quality during production. Laser-induced breakdown spectroscopy (LIBS) has been shown to be a promising technique for element analysis. Compared with the traditional detection methods，there are many advantages to modern iron and steel smelting analysis demand，such as non-preparative sample，online，faster and remote analysis. This thesis is based on the background of online component detection of molten steel in the process of iron and steel smelting, aiming to improve the performance of LIBS quantitative analysis, studying the theory and application research of LIBS. The content of this thesis is as the follows: research on reducing the influence of measuring repeatability by LIBS system noise, research on improving the detection sensitivity of noisy LIBS signal, LIBS signal peak identification method based on continuous wavelet transform, the strategy of initial value selection for resolving overlapped peaks based on curve fitting. (1) For the impact of poor smelting conditions of molten steel and the heavy disturbance of LIBS system noise, the noise sources of LIBS system are analyzed, and the noise is removed by wavelet threshold denoising, based on entropy analysis a method of choosing the optimal decomposition level for wavelet threshold denoising is presented. This method examines the change of the white noise variance on each decomposition level, derives the entropy variation tendency on each decomposition level. And then calculates noisy LIBS signal entropy under each decomposition level combining with the wavelet coefficient energy, uses to compare the white noise entropy. By examining the change of two kinds of entropy, the optimal decomposition level is decided at last. Compared to traditional method of selection the optimal decomposition level, this method has better predictability with precisely reliable calculating results and noise suppression plays an important role in quantitative analysis and improving the repeatability of measurement of LIBS signal. (2) In order to further enhance LIBS signal detection sensitivity of molten steel, a wavelet threshold denoising method with double thresholds correction is presented. By examining the relation between the threshold and the threshold function, a correction method of double thresholds is presented. Firstly, establishes the new upper threshold model and uses the grey system theory to correct it; Secondly, according to the completed correction new upper threshold, establishes the new lower threshold model, and uses fuzzy nearness degree to describe the new lower threshold model. With this, the double thresholds correction is completed. Wavelet threshold denoising with the double thresholds by this method not only improves signal to noise ration but also preserves edge information. And more importantly, reduces the limit of detection and enhances the sensitivity of quantitative analysis. (3) Because the number of the characteristic peak of molten steel LIBS signal is large and the accuracy of traditional methods of peak detection is poor, LIBS signal peak identification method is presented based on continuous wavelet transform. This method searches the maximum values of the wavelet coefficients under each scale for establishing wavelet ridge lines, which are used to identify LIBS signal peak. By the comparison with the traditional methods of LIBS signal peak identification, this method has better sensitivity and lower misdetection. (4) Aiming at the phenomenon of the overlapped peaks generated by the mixed response of many components in molten steel, the methods of resolving overlapped peaks are studied. Based on curve fitting a strategy of initial value selection for resolving overlapped peaks is presented. In order to solve the multi-solution problem of curve fitting, firstly, the calculation of approximate derivative by continuous wavelet transform is used to define the number and positions of sub-peaks. But in the process of optimization algorithm, the two parameters do not participate in iteration operation; Secondly, based on Lorentz function model, by examining the relation between the characteristic points of fractional derivative of model and derivative order, the parameter estimators are established and used to calculate the initial values for iteration operation, which greatly narrows the search scope of optimization algorithm. Based on curve fitting for resolving overlapped peaks, the sub-peaks hidden in the overlapped peaks are reconstructed, which can be decreased the errors and increased sensitivity for quantitative analysis. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 95页 |
| 源URL | [http://ir.sia.ac.cn/handle/173321/16748]  |
| 专题 | 沈阳自动化研究所_工业控制网络与系统研究室 |
| 推荐引用方式 GB/T 7714 | 张博. 液态钢成分在线检测的LIBS信号处理技术研究[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. 2014. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。