~在矿山安全中，井下巷道变形的监测一直是学者们研究的热点。其中光纤光栅传感器以其体积小、表征变形量多、布设方便等优点可用于巷道变形监测。由于井下环境复杂，光纤光栅传感器监测的数据往往伴随着噪声，导致监测数据曲线不够平滑，不利于预警系统的分析和决策。且岩层运动具有随机性和不确定性，使用传统的预测手段难以达到准确预测巷道变形的目的，因此本文将去噪算法和人工智能领域的算法结合，引入光纤光栅监测巷道变形系统中，可有效解决光纤光栅数据噪声和巷道变形预测难的问题。
本文的数据来源于杨伙盘煤矿主平硐的光纤光栅变形监测系统。通过对历史数据曲线的观察，发现数据曲线中存在大量毛刺和矩形锯齿状冗余的问题，因此判断光纤光栅监测系统采集的数据中混入了噪声数据。针对数据中存在的噪声问题，选取了滑动均值、卡尔曼滤波以及完全自适应噪声集合经验模态分解（Complete EEMD with Adaptive Noise，CEEMDAN）法三种方法对数据进行去噪，采用去噪结果和原始数据间的误差评估去噪效果。为得到最优的预测神经网络结构，将去噪的数据经归一化后，引入了在遗传算法优化下的多层前馈（Back Propagation，BP）神经网络、长短时记忆（Long Short-Term Memory，LSTM）神经网络和LSTM-BP神经网络三种网络结构的预测研究。为对比数据处理前后的预测效果，采用原始监测数据，通过LSTM-BP神经网络结构进行预测。本文对神经网络预测的表现，均采用均方误差（Mean Squared Error，MSE）、平均绝对误差（Mean Absolute Error，MAE）、均方根误差（Root Mean Squard Error，RMSE）及决定系数（R-Square，R2）四种评估方式进行评估。
实验结果表明，采用CEEMDAN法处理数据噪声，数据平滑性好，与原始数据的MSE、MAE最小，提取有效信息最多，处理数据噪声的效果最优。经遗传算法（Genetic Algorithm，GA）对模型结构寻优后，采用LSTM-BP神经网络的预测效果最优，其预测值和真实值之间的误差最小，拟合优度最大，均能达到0.93以上，且平均拟合优度为0.96。原始监测数据的预测效果在误差和拟合优度上的表现均不如去噪后的数据预测效果。通过本文的数据去噪处理和变形预测研究，可为巷道安全预警提供理论支持和现场实用依据。

~In mine safety，the deformation monitoring of underground roadway has always been a hot topic for scholars. The fiber grating sensor can be used for roadway deformation monitoring because of its small size，large deformation and convenient layout. Due to the complex underground environment，the data monitored by fiber grating sensors are often accompanied by noise，resulting in the monitoring data curve is not smooth，which is not conducive to the analysis and decision-making of early warning systems. And the rock movement has randomness and uncertainty，it is difficult to achieve the purpose of accurate prediction of roadway deformation by traditional prediction methods. Therefore，this paper combines the denoising algorithm and the algorithm in the field of artificial intelligence，and introduces the fiber grating monitoring roadway deformation system，which can effectively solve the problem of fiber grating data noise and roadway deformation prediction.
The data in this paper are from the fiber grating deformation monitoring system of the main adit in Yanghuopan Coal Mine. Through the observation of the historical data curve，it is found that there are a large number of burrs and rectangular sawtooth redundancy problems in the data curve，so it is judged that the noise data is mixed in the data collected by the fiber grating monitoring system. Aiming at the noise problem in the data，three methods of sliding mean，Kalman filter and Complete EEMD with Adaptive Noise（CEEMDAN）are selected to denoise the data. The denoising effect is evaluated by the error between the denoising results and the original data. In order to obtain the optimal prediction neural network structure，the denoised data are normalized，and the prediction research of three network structures，namely，multilayer feedforward（BP）neural network，long short-term memory（LSTM）neural network and LSTM-BP neural network optimized by genetic algorithm，is introduced. In order to compare the prediction effect before and after data processing，the original monitoring data is used to predict by LSTM-BP neural network structure.In this paper，the performance of neural network prediction is evaluated by four evaluation methods : mean squared error（MSE），mean absolute error（MAE），root mean squared error（RMSE）and determination coefficient（R-Square，R2）.
The experimental results show that the CEEMDAN method for processing data noise has good data smoothness，the minimum MSE and MAE with the original data，the most effective information extracted，and the best effect for processing data noise. After the model structure is optimized by genetic algorithm (GA)，the prediction effect of LSTM-BP neural network is the best. The error between the predicted value and the real value is the smallest，and the goodness of fit is the largest，which can reach more than 0.93，and the average goodness of fit is 0.96. The prediction effect of original monitoring data is not as good as that of denoised data in error and goodness of fit. The research on data denoising and deformation prediction in this paper can provide theoretical support and practical basis for roadway safety early warning.