快速且准确识别矿井涌水水源对于防范煤矿水灾事故有着重大的研究意义。 利用激光诱导荧光(LIF)光谱融合智能分类算法进行矿井涌水水源识别打破了传统水化学方法耗时过长等不足, 具有灵敏度高、 响应速度快等特点； 然而目前这些已使用的算法仅能依靠分类准确率定性判别不同矿井涌水水样的种类。 把随机森林(RF)算法和竞争性自适应重加权(CARS)算法相结合, 基于激光诱导荧光的水样荧光光谱数据建立偏最小二乘回归(PLSR)模型来预测不同矿井涌水的含量, 实现水样的定量评估。 首先, 采集300组以老空水为基础混入不同含量砂岩水的矿井涌水样本, 将采集到的水样按4∶1比例随机划分成校正集和预测集, 校正集共240组用于建立回归模型, 预测集共60组用于预测不同水样, 搭建激光诱导荧光涌水光谱系统完成光谱数据的获取并生成荧光光谱图。 然后分别通过S-G卷积平滑法和Lowess平滑法对原始荧光光谱进行去噪处理, 发现处理后的荧光光谱较原始光谱更为分散, 适合光谱分析, 对比了两种去噪方法的预测精度, 选择Lowess平滑法作为最终去噪方法。 接着采用RF算法约简去噪后属性重要度较低的光谱属性, 依据最优回归模型的性能选择约简出的223个属性再用于CARS算法的二次属性精简, 根据CARS算法采样过程中交叉验证均方根误差值最小原则选择出的77个属性光谱数据建立PLSR模型。 最后与全光谱、 其他变量选择方法、 不同回归模型相比, RF-CARS算法属性精简效果最好, 较全光谱建模, 属性由2 048个减少到77个, 模型预测集判定系数R2pre由0.991 4增长到0.996 7, 预测均方根误差RMSEP由0.029 4减少到0.018 3, 预测精度得到提升, 其余评估指标也相对较好。 实验结果表明, RF-CARS结合激光诱导荧光光谱可快速、 精准预测矿井涌水, 精简出的光谱属性用来建立回归模型, 为实现矿井涌水含量的实时定量评估提供了一定的理论保障。

Quick and accurate identification of mine water inflow has important research significance for preventing coal mine flood accidents, the laser-induced fluorescence (LIF) spectroscopyis used to integrate withthe intelligent classificationalgorithm to identify the mine water inflow, it breaks the shortcomings of traditional water chemistry methods, such as long time consuming, etc., and has the characteristics of high sensitivity and fast response. However, these currently used algorithms can only rely on the classification accuracy to qualitatively discriminate the types of water samples from different mine water inflow. This paper combines the random forest algorithm with the competitive adaptive weighting algorithm (RF-CARS), the partial least squares regression (PLSR) model based on fluorescencespectrum data from the laser-induced fluorescence was used to predict the water inflow in different mines and to achieve quantitative assessment of water samples. Firstly, 300 sets of mine water inflow samples mixed with different sandstone waters based on goaf water were collected, and the collected water samples were randomly divided into the calibration set and the prediction setaccording to the ratio of 4∶1, a total of 240 sets of calibration sets were used to establish a regression model, a total of 60 sets of prediction sets were used to predict different water samples, and a laser-induced fluorescence inflow spectroscopy system was built to complete the acquisition of spectral data and generated a fluorescence spectrum. Then the original fluorescence spectrum was denoised by S-G convolution smoothing method and Lowess smoothing method, and it was found that the processed fluorescence spectrum was more dispersed than the original spectrum, which was suitable for spectral analysis, the prediction accuracy of two denoising methods were compared, the Lowess was chosen as the final denoising method. Then, the RF algorithm was used to reduce the spectral attributes with low attribute importance after denoising, according to the performance of the optimal regression model, the 223 reduced attributes were selected and then it was used for the secondary attribute reduction of the CARS algorithm. The PLSR model was established based on 77 spectral attribute data selected according to the principle of minimum cross validation root mean square error in the sampling process of CARS algorithm. Finally, we compared with the full spectrum, other variable selection methods, and different regression models, the RF-CARS algorithm had the best streamlining effect, and the total spectral modeling attribute was reduced from 2 048 to 77, the model prediction set determination coefficient R2pre increased from 0.991 4 to 0.996 7, the predicted root mean square error RMSEP decreased from 0.029 4 to 0.018 3, the prediction accuracy was improved, and the remaining evaluation indicators were relatively good. The experimental results show that the RF-CARS combined with laser induced fluorescence technology can quickly and accurately predict mine water inflow, the simplified spectral attributes are used to establish regression model, which provides a theoretical guarantee for real-time quantitative evaluation of mine water inflow.