土壤是人类赖以生存的物质基础, 它的特性与人们的生产生活密切相关, 传统土壤重金属检测方法如原子吸收光谱法、 电感耦合等离子体质谱法分析能力较弱且价格昂贵, 因此亟需一种开发低成本易操作的土壤多种元素同时定量分析方法。 激光诱导击穿光谱(LIBS)技术因其分析快速、 多元素同时分析等优点被广泛关注, 但由于其体积较大不便于携带并且大多应用于实验室分析, 为了满足现场检测的需求, 研制了一台分体式现场LIBS检测仪, 其设计是将仪器分为探测头和机箱两部分并通过包塑管进行连接, 采用微型二极管泵浦激光器, 脉冲能量可达100 mJ, 波长1 064 nm, 重复频率为1~10 Hz; 此外采用多通道高分辨率光谱仪, 提高LIBS的分析性能; 为了减小辐射背景干扰, 利用FPGA实现μs级延迟时间功能。 对其应用在11种土壤获取光谱数据, 实验设置脉冲能量为100 mJ, 延迟时间设为1 μs, 积分时间2 ms, 每个样品采集10个不同位置的光谱, 每个位置测量20次, 共采集200个光谱数据, 为减少噪声干扰, 对每个样品的光谱数据进行均值预处理后再进行Beads算法基线校正, 利用PCA主成分分析得到贡献率最大的3个主成分分量, 再通过Kmeans++算法对不同地区不同类型的11种土壤进行聚类分析, 将相同类别的土壤代入偏最小二乘(PLSR)算法, 每个元素选取两个特征谱线以及上下各10个点来增强光谱信息, 选择一种样品作为预测对Cu、 Cr、 Ni、 Co、 Cd五种土壤重金属元素进行定量分析。 结果表明, 与未进行聚类分析相比, 此方法可明显提高元素的拟合相关系数, 五种重金属元素的相关系数分别从0.953、 0.992、 0.989、 0.982、 0.99提高至0.999、 0.998、 0.999 5、 0.996 5、 0.993, 相关系数均达到0.99及以上满足LIBS线性分析要求, 其预测结果与实际含量之间的平均相对误差分别从83.45%、 16.03%、 22.94%、 43.91%、 125.768%提高至1.14%、 0.99%、 0.895%、 1.879%、 1.862%, 可以发现经过聚类分析后, 其预测误差大大降低, 均在5%以内, 具有较好的分析性能, 五种元素的相关系数和预测误差相比于直接进行PLSR方法均有提升。 PCA与Kmeans++结合的方法能够更准确的进行聚类, 在降维后进行聚类可以减少噪声和冗余信息的影响, 加快计算速度, 减少异常点对聚类效果的影响提高鲁棒性。

Soil is the material basis of human survival; its characteristics closely relate to people's production and life. Traditional soil heavy metal detection methods such as atomic absorption spectroscopy and inductively coupled plasma mass spectrometry analysis are weak and expensive, so the development of low-cost operating soil elements quantitative analysis method at the same time. Laser-induced breakdown spectroscopy (LIBS) technology has been widely used because of its rapid and multi-element simultaneous analysis. However, because it is not easy to carry, a split-type field LIBS detector was developed to meet the field testing needs. Its design is to divide the instrument into two parts, probe head, and chassis, and connect it through a plastic pipe. Using a miniature diode pump laser, the pulse energy is up to 100 mJ, with a wavelength of 1 064 nm. The repetition frequency is 1~10 Hz. In addition, using a multichannel high-resolution spectrometer improves LIBS's analytical performance. FPGA is used to realize the us-level delay time function to reduce radiation background interference. To obtain spectral data in 11 soils, The pulse energy was 100 mJ, The delay time was set to 1us, Integration time of 2 ms, Spectra from 10 different positions were collected for each sample, Each position was measured 20 times, A total of 200 spectral data were collected, To reduce the noise interference, The spectral data for each sample were mean-preprocessed after the Beads algorithm baseline correction, The three principal component components with the largest contribution rate were obtained using PCA principal component analysis, In the clustering analysis of 11 different types of soils in different regions by the Kmeans++ algorithm, of the same category of soil into the partial least squares (PLSR) algorithm, Each element selects two characteristic lines and 10 points to enhance the spectral information, One sample was selected as a prediction for quantitative analysis of five soil heavy metal elements, Cu, Cr, Ni, Co, and Cd. the results show that, In contrast to that where no cluster analysis was performed, This method can significantly improve the fitting correlation coefficient of the elements, The correlation coefficients of the five heavy metal elements increased from 0.953, 0.992, 0.989, 0.982, 0.99 to 0.999, 0.998, 0.999 5, 0.996 5, 0.993, respectively, The correlation coefficient of 0.99 and above all meet the requirements of LIBS linear analysis, The average relative error between the prediction results and the actual content increased from 83.45%, 16.03%, 22.94%, 43.91%, 125.768% to 1.14%, 0.99%, 0.895%, 1.879%, 1.862%, respectively, It can be found that after the cluster analysis, Its prediction error is greatly reduced, All were within 5%, With a relatively good analytical performance, The correlation coefficient and prediction error of the five elements are improved compared with the direct PLSR method. Combining PCA and Kmeans++ can be more accurate clustering after dimension reduction, reduce the influence of noise and redundant information, speed up the calculation, reduce the influence of abnormal points on the clustering effect, and improve the robustness.