快速准确的识别煤矿含水层水源对于煤矿突水预警及灾后救援意义重大, 针对传统水源识别耗时较长, 不适宜构建在线式预警系统, 提出使用激光诱导荧光技术用于煤矿水源类型识别的方法。 利用激光激发待测水样, 获取其荧光光谱, 结合模式识别对水源进行快速辨识。 实验采集了淮南矿区谢桥煤矿的两种纯水样本-老空水与砂岩水, 并根据不同混合比配成5种混合水样进行实验。 首先针对获取的水源荧光光谱中可能会存在的各种噪声及干扰信息, 采用SG、 Normalize、 Gapsegment求导、 Detrend和MSC 5种常用的光谱预处理算法对光谱数据进行处理。 其次针对荧光光谱数据量过大, 对数据进行PCA降维, 作为对比6种预处理方式(含原始光谱)主成分数皆取3, 结果显示SG预处理累计贡献度最大, 为97.26%； 其次是原始光谱, 为92.38%, Normalize与Detrend累计贡献度相差不大, 分别为88.04%和87.59%, MSC为66.41%, Gapsegment最差, 为22.65%。 最后分别对PCA降维后的数据使用线性LDA以及非线性RBF-SVM模型进行识别对比。 使用LDA进行建模, SG-PCA-LDA正确率最高, 达到了98.86%, 依据建立的LDA模型, 对验证集数据进行识别, SG-PCA-LDA的正确率依然最高, 为100%。 使用RBF-SVM进行建模, Original-PCA-RBF-SVM, SG-PCA-RBF-SVM, Normalize-PCA-RBF-SVM正确率最高, 皆为97.14%, 依据建立的RBF-SVM模型, 对验证集数据进行识别, Original-PCA-RBF-SVM和SG-PCA-RBF-SVM正确率依然最高, 为97.14%。 对比两类模型可以发现, LDA验证集正确率较建模集有一定的提升, 而RBF-SVM验证集正确率较建模集有小幅度降低, 说明LDA模型对于此煤矿水源荧光光谱数据的泛化能力较好, 且成功率较高。 结果表明, SG-PCA-LDA模型结合激光诱导荧光技术是一种较佳的应用于本地煤矿水源识别的方法, 且验证了对老空水、 砂岩水的纯水样和混合水样识别的可能性, 可以推广到煤矿其他混合水源的识别中。

The rapid and accurate identification of coal mine aquifer water source is of great significance for coal mine water inrush warning and post-disaster rescue. It takes a long time for water source identification with the traditional method, and it is not suitable to construct an online early warning system. A method of using laser induced fluorescence technology to identify the type of coal mine water source is proposed. The laser is used to excite the water sample. Then the fluorescence spectrum is obtained, with pattern recognition the water source can be rapidly identified. Two kinds of water samples-goaf water and sandstone water of Xieqiao Coal Mine in Huainan Mining Area were collected, and five mixed water samples were prepared according to different mixing ratios. Firstly, according to the various noise and interference information that may exist in the obtained water source fluorescence spectrum, the spectral data were pretreated by SG, Normalize, Gapsegment derivation, Detrend and MSC. Secondly, PCA was used to reduce the dimension of fluorescence spectral data due to a large amount of data. As a comparison of the six pretreatment methods (including the original spectrum), the number of principal components was taken by 3, and the results showed that the cumulative contribution of SG pretreatment is the largest, which was 97.26%. The second was the original spectrum, which was 92.38%. The cumulative contribution of Normalize and Detrend were not much different, which were 88.04% and 87.59%, MSC was 66.41%, and Gapsegment was the worst with 22.65%. Finally, the linear model of LDA and nonlinear model of RBF-SVM were used to identified and compared with the data of reduced dimension by PCA. Using LDA for modeling, SG-PCA-LDA had the highest accuracy rate, which reached 98.86%. According to the LDA model established, the verification set data were identified, and the accuracy rate of SG-PCA-LDA was still the highest with 100%. Using RBF-SVM for modeling, Original-PCA-RBF-SVM, SG-PCA-RBF-SVM, and Normalize-PCA-RBF-SVM had the highest accuracy rate, both of which was 97.14%. Based on the RBF-SVM model established, verification set data were identified, and the accuracy rate of Original-PCA-RBF-SVM and SG-PCA-RBF-SVM was still the highest, which is 97.14%. Tt can be found that the accuracy rate of the LDA verification set was improved which compared with the modeling set, and the accuracy rate of the RBF-SVM verification set was slightly lower than the modeling set, which showed that LDA model had better generalization ability and higher accuracy rate for fluorescence spectral data of this coal mine water. The results showed that the SG-PCA-LDA model combined with laser induced fluorescence technology is a better method for local coal mine water source identification, and it verified the possibility of identification for goaf water, sandstone water and mixed water, which can be extended to identify other mixed water sources of coal mines.