<p>矿山围岩变形的监测与预测是保证矿井安全及实现智能化开采的重要前提和基础。现有光纤感知围岩变形的研究工作通过相似材料物理模型试验和现场应用实践，建立了采动覆岩变形的光纤表征和光纤与巷道围岩间的应变传递机制。然而，缺乏对光纤监测数据蕴含的围岩变形时空特征规律的深入解析和围岩变形预测方法的研究。采动覆岩变形是由煤层开采引起上覆岩层应力重新分布和动态调整，导致岩层破断、垮落和移动而形成；采动覆岩变形结合煤层厚度、煤层倾角以及推进速度等地质条件和采矿工艺参数因素的影响与工作面周期来压具有时空相关性。而巷道围岩变形是由于应力分布不平衡引起。可见，采动覆岩变形和巷道围岩变形的成因不同，构建一个统一的预测模型颇具挑战。为此，本文利用深度学习强大的非线性表征能力，分别研究光纤监测两带高度、工作面来压步距和巷道围岩变形的预测方法，对煤矿智能化开采具有重要意义。</p> <p>主要研究内容如下：</p> <p>(1) 光纤监测数据的异常数据识别方法。监测数据数量和数据表现特征的不同，使异常数据具有不同的特点。针对数据量小且长期缓慢变化的监测数据，提出了改进聚类密度的快速识别方法。首先，采用K-means聚类算法快速定位单点异常和区间异常范围，然后采用基于最近邻密度算法识别异常数据。针对数据量大且正常与异常数据不平衡的监测数据，提出了结合不同采样方法的随机森林异常数据识别方法。正常多数类数据进行去重复的欠采样，异常少数类数据进行合成少数类的过采样，对新数据集采用随机森林方法进行异常数据识别。最后以工程应用中的光纤监测数据验证了两种异常数据识别方法的有效性。</p> <p>(2) 光纤监测数据的缺失数据填补方法。不同采样方式下光纤传感器监测数据中缺失数据的特点不同，针对分布式光纤多采样点少量随机缺失值的填补问题，提出了最小二乘支持向量机(LSSVM)的多采样点单属性缺失数据快速填补方法。针对单采样点光纤传感器时间序列数据连续缺失数据的填补问题，利用同类光纤传感器监测数据间的相关性，提出了遗传算法(GA)优化极端梯度提升(XGBoost)的缺失数据填补方法。以物理模型试验和工程应用中的采动覆岩变形和巷道围岩中平硐变形的不同监测数据，验证了缺失数据填补方法的有效性。</p> <p>(3) 分布式光纤监测采动覆岩变形两带高度动态预测方法。在光纤频移值对采动覆岩变形时空特征表征框架的基础上，考虑岩层的抗压强度、岩层弹性模量、煤层厚度以及开采速度等因素对采动应力的影响，提出了集成经验模态分解(EEMD)和集成深度学习的两带高度动态预测方法。使用EEMD分解方法消除监测数据噪声和非平稳性；对光纤全采样点频移值分量建立集成学习CNN-GRU-BP的预测方法，预测分量叠加得到全采样点的预测频移值曲线。根据频移值曲线的阶跃高度预测出垮落带和裂隙带发育高度。在三维和平面物理模型试验中验证了预测方法的有效性和鲁棒性。</p> <p>(4) 分布式光纤监测采动覆岩变形工作面来压步距动态预测方法。在光纤频移变化度对工作面周期来压时空特征表征框架的基础上，定义了光纤频移值的覆岩应变信息熵，定量表征不同岩层变形对工作面来压的影响，并考虑了煤层厚度、开采高度、工作面推进距离以及煤层倾角等影响因素，提出了双分支卷积循环神经网络工作面来压步距预测方法。设计了融合注意力机制的APMCNN-GRU网络分支，提取不同空间覆岩应变信息熵蕴含的周期来压的时空特征。设计了ID-CNN-GRU网络分支，深入挖掘光纤频移变化度数据序列和周期来压影响因素中蕴含的来压步距特征。通过BP神经网络将提取的特征进行融合，并输出预测的周期来压步距。在三维和平面物理模型试验中验证了预测方法的有效性和鲁棒性。</p> <p>(5) 光纤监测巷道围岩中平硐变形多步预测方法。光纤监测巷道围岩变形的工程中监测数据含有大量噪声，影响多步预测的精度并产生较大的误差累积。在巷道围岩中平硐等缓慢变形场景下，提出了结合EEMD方法的TCN-LSTM平硐变形多步预测方法。采用EEMD方法消除监测数据噪声并使用模糊熵提取有效分量，构建了多路TCN-LSTM深度神经网络提取不同分量的长时间域特征和非线性特征，并采用多输出策略，有效降低了累积误差。叠加预测分量输出未来一段时间的平硐变形预测结果。利用光纤监测平硐变形的现场应用数据，验证了多步预测方法的有效性和鲁棒性。</p> <p>本文基于深度学习建立了光纤监测围岩变形预测方法，为围岩控制提供了理论和技术支撑，对推动矿山围岩变形智能预测预报技术的研发与应用具有积极意义。</p>

<p>The monitoring and predicting the surrounding rock deformation in mines is a crucial prerequisite and foundation for ensuring mine safety and achieving intelligent mining. Existing work on optical fiber sensing of surrounding rock deformation has established the optical fiber characterization of mining overburden deformation and the strain transfer mechanism between optical fiber sensors and tunnel surrounding rock through similar material physical model tests and field applications. However, there is a lack of in-depth exploration on the spatiotemporal characteristics of surrounding rock deformation implied by fiber optic monitoring data and on methods for predicting surrounding rock deformation. Mining-induced overburden deformation is caused by the redistribution and dynamic adjustment of stress in overlying strata due to coal extraction, leading to rock fracture, collapse, and movement. The deformation of strata combined with by geological conditions such as coal seam thickness, coal seam dip angle, and advancement speed, as well as mining process parameters, and &nbsp;exhibits spatiotemporal correlation with the periodic pressure of the working face. In contrast, deformation of tunnel surrounding rock is caused by imbalanced stress distribution. It is evident that the causes of mining-induced overburden deformation and roadway surrounding rock deformation differ, making it quite challenging to establish a unified prediction model. Therefore, this thesis leverages the powerful nonlinear representation capabilities of deep learning to separately study prediction methods for fiber optic monitoring of the two-zone height, the periodic pressure distance of the working face, and roadway surrounding rock deformation. This research holds significant importance for the intelligent mining of coal.</p> <p>&nbsp;The main research contents are as follows:</p> <p>(1)&nbsp; Anomaly detection methods for fiber optic monitoring data. The differences in the quantity and characteristics of monitoring data result in varying features of abnormal data. For monitoring data with small quantities and long-term slow changes, a fast identification method based on improved clustering density is proposed. Firstly, the K-means clustering algorithm is used to quickly locate the ranges of single-point anomalies and interval anomalies, and then the nearest neighbor density-based algorithm is employed to identify abnormal data. For monitoring data with large quantities and imbalances between normal and abnormal data, a random forest anomaly detection method combining different sampling techniques is proposed. The method involves undersampling the normal majority class data by removing duplicates and oversampling the abnormal minority class data by synthetic minority oversampling. The random forest method is then applied to the new dataset for anomaly detection. Finally, the effectiveness of both anomaly detection methods is validated using fiber optic monitoring data from engineering applications.</p> <p>(2) Methods for filling missing data in fiber optic monitoring data. The characteristics of missing data in fiber optic sensor monitoring data vary with different sampling methods. To address the problem of filling a small number of randomly missing values at multiple sampling points in distributed fiber optics, a fast filling method for single-attribute missing data at multiple sampling points using the Least Squares Support Vector Machine (LSSVM) is proposed. For addressing the problem of continuously missing data in time series from single sampling point fiber optic sensors, a method utilizing the correlation among data from similar fiber optic sensors is proposed. This method employs a Genetic Algorithm (GA) to optimize Extreme Gradient Boosting (XGBoost) for filling in the missing data. The effectiveness of this method is validated using different monitoring data from physical model tests and engineering applications, including mining-induced overburden deformation and tunnel deformation in roadway surrounding rock.</p> <p>(3) Dynamic prediction method of two-zone height for distributed fiber optic monitoring of mining-induced overburden deformation. Building upon the framework of fiber optic frequency shift values for characterizing the spatiotemporal features of mining-induced overburden deformation, this factors such as rock compressive strength, rock elastic modulus, coal seam thickness, and mining speed affecting mining stress affecting mining stress is considered. It proposes an integrated approach using Empirical Mode Decomposition (EEMD) and integrated deep learning for dynamic prediction of two-zone height variations. The EEMD decomposition method is utilized to eliminate noise and non-stationarity in monitoring data. A prediction method is developed using integrated learning CNN-GRU-BP for the frequency shift components at all fiber optic sampling points, and these components are combined to obtain the predicted frequency shift curve across all sampling points. Based on the step heights of the frequency shift curve, the development heights of collapse zones and fracture zones are predicted. The effectiveness and robustness of the prediction method are validated through physical modeling experiments in both three-dimensional and planar setups.</p> <p>(4) Dynamic prediction method for the step pressure on the working face of mining-induced overburden deformation using distributed fiber optic monitoring. Building upon the framework where the fiber frequency shift variability characterizes the spatiotemporal features of pressure on the working face using fiber optic frequency shift variations, the overburden strain information entropy of the fiber frequency shift value is defined. This quantitatively characterizes the impact of different rock strata deformations on the working face pressure.Considering factors such as coal seam thickness, mining height, working face advancement distance, and coal seam dip angle, a dual-branch convolutional recurrent neural network method for predicting the step pressure on the working face is proposed. Additionally, an ID-CNN-GRU network branch was developed to delve into the features of the step pressure implied by the fiber optic frequency shift variations and factors affecting periodic stress. These features are fused using a BP neural network to predict the periodic advancing step distance. The effectiveness and robustness of the prediction method are validated through experiments in both three-dimensional and planar physical models.</p> <p>(5)&nbsp; Multi-step prediction method for deformation of adit deformation in tunnel surrounding rock monitored by fiber optics. In the engineering practice of fiber optic monitoring of surrounding rock deformation in tunnel roadways, the monitoring data contains a significant amount of noise, which affects the accuracy of multi-step predictions and leads to considerable error accumulation. In scenarios of slow surrounding rock deformation, such as adit deformation, a multi-step prediction method for adit deformation combining EEMD and TCN-LSTM is proposed. The EEMD method is used to eliminate noise from the monitoring data, and fuzzy entropy is utilized to extract effective components. A multi-channel TCN-LSTM deep neural network is constructed to extract long-term domain features and nonlinear characteristics from different components. By employing a multi-output strategy, the method effectively reduces cumulative error. The predicted components are superimposed to output the future deformation prediction results for the tunnel. The effectiveness and robustness of the multi-step prediction method are validated using field application data from fiber optic monitoring of tunnel deformation.</p> <p>This thesis establishes a prediction method for surrounding rock deformation based on fiber optic monitoring using deep learning. This method provides theoretical and technical support for surrounding rock control and has significant positive implications for advancing the research and application of intelligent prediction and forecasting technology for mine surrounding rock deformation.</p>