矿井水害对煤矿安全生产存在巨大威胁, 所以快速识别矿井突水水源, 对煤矿水灾预警及灾后救援工作开展都有重大意义。 激光诱导荧光(LIF)技术具有快速、 高效、 灵敏度高等特点, 克服了传统水化学方法识别时间长的缺点。 循环神经网络(RNN)在解决长序列训练过程中产生的梯度消失、 梯度爆炸等问题上存在明显不足, 而特殊变体RNN即长短期记忆(LSTM)神经网络很好地弥补了RNN的短板及缺陷。 提出了将LIF技术与LSTM算法相结合, 应用在矿井突水水源快速识别中。 实验样本采自淮南矿区, 以砂岩水和老空水为原始样本, 并将砂岩水和老空水按照不同比例混合配置成5种混合水样, 共7种待测水样进行实验。 首先采用最大最小值归一化(MinMaxScaler)、 平滑滤波(SG)以及标准正态变量变换(SNV)三种预处理方法对原始光谱数据进行预处理, 减少原始光谱数据存在的噪声和干扰信息。 之后为防止数据量过大, 维度过高, 将包括原始光谱数据在内的四组数据再进行LDA降维至3维。 最后分别搭建LSTM识别模型, 从测试集预测准确率、 训练集准确率变化趋势以及训练集损失函数变化趋势三个方面进行比较, 选择最优模型。 其中SG+LDA+LSTM和Original+LDA+LSTM在测试集预测准确率上都能达到100%, MinMaxScaler+LDA+LSTM测试集预测准确率在98.57%, SNV+LDA+LSTM准确率最低, 只有87.14%; 在训练集准确率变化趋势表现上, SG+LDA+LSTM能够保持良好的学习, 很快达到100%, Original+LDA+LSTM和MinMaxScaler+LDA+LSTM也能达到100%的准确率, 但在前几次训练过程中会有准确率下降的情况出现, SNV+LDA+LSTM训练集准确率在训练次数内并未达到100%; SG+LDA+LSTM损失函数变化趋势也具有很好的收敛性和稳定性, Original+LDA+LSTM, MinMaxScaler+LDA+LSTM以及SNV+LDA+LSTM在损失函数变化趋势上表现并不出色。 结果表明, 4组模型中, SG+LDA+LSTM模型是最适合应用于矿井突水识别, 该方法补充了矿井突水水源识别工作的内容, 为矿井突水识别提供了新的思路。

Mine water disasteris a great threat to the safety production of a coal mine, so the rapid identification of mine water inrush source is of great significance to the early warning and post-disaster rescue work. Laser-induced fluorescence (LIF) technology has high speed, high efficiency and high sensitivity, which overcomes the shortcomings of long recognition time in traditional hydrochemical methods. Circulating neural network (RNN) has obvious shortcomings in solving the problems of gradient disappearance and gradient explosion in long sequence training, while the special variant RNN, long and short term memory (LSTM) neural network, makes up for the shortcomings of RNN.In this paper, the combination of LIF technology and LSTM algorithm is applied to rapidly identify mine water inrush source.The experimental samples were collected from Huainan Mining Area. The sandstone water and goaf water were taken as the original samples, and the sandstone water and goaf water were mixed into 5 kinds of mixed water samples. According to different proportions, 7 kinds of water samples experimented. Firstly, MinMaxSxalerr, SG and SNV were used to preprocess the original spectral data to reduce the noise and interference. After that, to prevent the data from being too large and too high a dimension, the dimension of four groups of data, including the original spectral data, was reduced to 3 dimensions by LDA.Finally, the LSTM recognition models are built respectively, and the optimal model is selected by comparing the prediction accuracy of the test set, the changing trend of the accuracy and the loss function of the training set.Thereinto, SG+LDA+LSTM and Original+LDA+LSTM can reach 100% in the test set prediction accuracy, MinMaxScaler+LDA+LSTM test set prediction accuracy is 98.57%, SNV+LDA+LSTM accuracy is the lowest, only 87.14%;In terms of the trend of training set accuracy, SG+LDA+LSTM can keep good learning and reach 100% soon. Original+LDA+LSTM and MinMaxScaler+LDA+LSTM can also reach 100% accuracy. However, at the beginning of the training process, the accuracy will decline, and the SNV+LDA+LSTM training set does not reach 100% within the training times; The trend of SG+LDA+LSTM loss function also has good convergence and stability. Original+LDA+LSTM, MinMaxScalerr+LDA+LSTM and SNV+LDA+LSTM do not perform well in the trend of loss function.The results show that the SG+LDA+LSTM model is the most suitable for mine water inrush identification among the four models. This method supplements the work of mine water inrush source identification and provides a new idea for mine water inrush identification.