海上溢油事故不仅造成极大的石油资源浪费, 而且严重威胁生态环境。 因此, 利用荧光光谱对油膜厚度进行快速无损检测对于有效评估溢油量有重要意义。 基于激光诱导荧光(LIF)技术对海水表面0#柴油、 5#白油油膜的荧光光谱进行检测, 进而实现对油膜厚度的量化分析。 首先使用SG平滑滤波对原始光谱数据进行预处理以减少原始光谱中的背景噪声。 然后采用间隔随机蛙跳算法(IRF)结合变量子集迭代优化法(IVSO)对获取的全光谱数据进行波长选择以剔除冗余变量, 将经过二次筛选出的光谱特征波长作为偏最小二乘回归(PLS)的自变量输入数据建立油膜厚度反演模型。 该方法第一步利用IRF从全光谱数据中筛选出特征波段, 再利用IVSO对特征光谱波段组合进一步筛选出特征波长变量, 从而有效提高优选出的特征波长建立油膜厚度反演模型的预测能力和稳定性。 将IRF-IVSO与全光谱及移动窗口偏最小二乘法(MWPLS)、 间隔随机蛙跳算法(IRF)、 变量组合集群分析法(VCPA)、 变量子集迭代优化法(IVSO)四种波长优选方法进行对比, 发现IRF-IVSO筛选出0#柴油数据和5#白油数据的特征波长数量分别占全光谱数据的4.48%和19.40%。 将全光谱及上述波长优选方法筛选出的特征波长作为输入建立PLS模型进行分析讨论。 结果表明, 特征波长选择方法结合PLS所建立的不同模型预测能力和效率较全光谱有明显提高。 其中, IRF-IVSO结合PLS所建立的油膜厚度反演模型预测效果最优, 该模型可以实现对厚度分别为0.141 5～2.291 8和0.052～0.980 mm的0#柴油及5#白油油膜的有效反演, 柴油油膜测试集相关系数RP可达到0.961 1, 测试集均方根误差RMSEP为0.137 5, 白油油膜测试集相关系数RP可达到0.971 2, 测试集均方根误差RMSEP为0.079 0。 该研究表明, IRF-IVSO通过结合区间波段筛选和单一变量选择能够有效而稳定地筛选出特征波长变量, 结合PLS建立的油膜厚度反演模型能够实现可靠预测。

Offshore oil spill accidents cause a great waste of oil resources and seriously threaten the ecological environment. Therefore, it is important to use fluorescence spectroscopy to detect oil film thickness quickly and nondestructively for effective evaluation of oil spills. Based on laser-induced fluorescence (LIF) technology, the fluorescence spectra of oil film of 0# diesel oil and 5# white oil on sea water surface were detected, and then the oil film thickness was quantified. Firstly, SG was used to preprocess the original spectral data to reduce the background noise in the original spectrum. Then, interval random frog (IRF) combined with iteratively variable subset optimization (IVSO) was used to select the wavelength of the obtained full spectral data to eliminate redundant variables. The characteristic wavelength of the spectrum screened out twice was used as the independent variable input data of partial least squares regression (PLS) to establish the oil film thickness inversion model. In the first step of the method, the characteristic bands are screened from the full spectral data by IRF, and the characteristic wavelength variables are further screened by the combination of characteristic spectral bands by IVSO to effectively improve the prediction ability and stability of the oil film thickness inversion model based on the selected characteristic wavelengths. IRF-IVSO was compared with four wavelength optimization methods: full spectrum and moving window partial least squares (MWPLS), interval random frog (IRF), variables combination population analysis (VCPA) and iteratively variable subset optimization (IVSO). The characteristic wavelengths of 0# diesel oil and 5# white oil screened by IRF-IVSO accounted for 4.48% and 19.40% of the total spectral data, respectively. The full spectrum and the characteristic wavelengths screened by the above wavelength optimization method were used as input to establish a PLS model for analysis and discussion. The results show that the prediction ability and efficiency of different models established by using the feature wavelength selection method combined with PLS are significantly higher than that of the full spectrum. Among them, the oil film thickness inversion model established by IRF-IVSO combined with PLS has the best prediction effect. This model can realize effective inversion of 0# diesel oil and 5# white oil with the thickness of 0.141 5～2.291 8 and 0.052～0.980 mm, respectively, and the correlation coefficient RP of diesel oil film test set can reach 0.961 1. The RMSEP of the test set is 0.137 5, the correlation coefficient RP of the white oil film test set is 0.971 2, and the RMSEP of the test set is 0.079 0. This study shows that IRF-IVSO can effectively and stably screen characteristic wavelength variables by combining interval band screening and single variable selection, and the oil film thickness inversion model established by combining PLS can achieve reliable prediction.