近年来我国东部矿区发生了多起立井井筒倾斜变形及破损灾害，严重影响了矿井安全与生产。针对厚含水松散层深立井倾斜破损灾害，以鲁南某矿深立井井筒(800 m)为研究对象，开展了井筒倾斜变形监测，研究了井筒倾斜时空变化特征，分析了井筒倾斜主要影响因素；在此基础上，基于深度学习理论，综合采用循环神经网络(RNN)、长短期记忆网络(LSTM)、门控循环单元(GRU)、一维卷积神经网络(1DCNN)4种经典深度学习方法，构建了井筒倾斜变形预测模型，并将预测结果与实测值进行对比，分析了井筒变形预测模型精度，研究了井筒整体和关键区域预测效果，验证了模型可靠性，并开展了工程应用。研究表明：(1)井筒倾斜主要发生在松散层，倾斜值由浅到深线性减小、并偏向采空区一侧，最大为352 mm，基岩层变形较小，最大为88 mm；开采引起厚松散层变形传播范围增大、底部含水层沿井壁渗流疏水及地下水渗流场的变化是导致井筒倾斜变形的主要原因。(2)模型与实测值Spearman相关系数最大为0.978，最小为0.867，4种模型与现场实测偏移量的最大差值为0.043 m，平均绝对误差E_MA 在0.003～0.009 m内，均方根误差E_RMS 在0.004～0.011 m内，整体预测效果以1 DCNN模型最优，主要倾斜方向(偏向采空区一侧的东西方向)预测精度略低于变形量较小的方向(南北方向)，且均能够满足工程需要。(3)井筒整体预测曲线与实际倾斜方向一致，井口、松散层基岩交界面E_MA与E_RMS 平均值均为0.005 m、0.006 m，井底精度略低，其对应值为0.012、0.013 m，井筒特征区域与整体预测效果均表现良好，表明基于深度学习的井筒变形预测模型具有良好的预测能力，研究成果在井筒注浆修复治理工程中得到了有效应用，为井筒安全管理提供了技术参考和数据支撑，为类似工程提供了工程实践经验。

In recent years, a number of vertical shaft tilt deformation and breakage disasters have occurred in the eastern mining areas of China, which have seriously affected mine safety and production. In response to the tilting and damage disasters of deep vertical shafts in thick water-bearing loose layers, the tilting and deformation monitoring of shafts was carried out by taking the deep vertical shaft (800 m) of a mine in Lunan as the research object, studying the spatial and temporal change characteristics of shaft tilting, and analyzing the main influencing factors of shaft tilting; based on this, based on the deep learning theory, four types of deep learning method, namely, recurrent neural network (RNN), long and short-term memory network (LSTM), gated recurrent unit (GRU), and one-dimensional convolutional neural network (1DCNN), were used. unit (GRU), and one-dimensional convolutional neural network (1DCNN) to construct a wellbore tilt deformation prediction model, and compare the prediction results with the measured values to analyze the accuracy of the wellbore deformation prediction model, validate the reliability of the model, studied overall wellbore and critical area prediction effects, and carry out engineering applications. The study shows that:(1) The wellbore tilt mainly occurs in the loose layer, the tilt value decreases linearly from shallow to deep, and is biased towards the side of the extraction zone, with a maximum of 352 mm, and the deformation of the bedrock layer is smaller, with a maximum of 88 mm;the increase in the range of deformation propagation in the thick loose layer caused by the mining, and the change of seepage hydrophobicity of the aquifer at the bottom along the wall of the well and the seepage field of the groundwater are the main causes of the tilted deformation of the wellbore. (2) The Spearman correlation coefficient between the model and the measured value is 0.978 at the maximum and 0.867 at the minimum;the maximum difference between the four models and the field measured offsets is 0.043 m, the mean absolute error E_MA is within 0.003–0.009 m, and the root mean square error E_RMS is within 0.004–0.011 m. The overall prediction is optimized by the 1DCNN model, and the main tilting direction(The prediction accuracy of the main inclined direction (east-west direction, which is inclined to the side of the mining area)is slightly lower than that of the direction with smaller deformation amount (north-south direction), and all of them can meet the engineering needs. (3) The overall prediction curve of the wellbore is consistent with the actual tilt direction, and the average values of E_MA and E_RMS of the wellhead and loose bedrock interface are 0.005 m and 0.006 m. The accuracy of the wellbore bottoming is a little bit lower, with the corresponding values of 0.012 m and 0.013 m. The wellbore characteristic area and overall prediction effect are good, indicating that the wellbore deformation prediction model based on deep learning has good prediction ability. The research results have been effectively applied in the wellbore grouting repair and management project, which provides technical reference and data support for the safe management of wellbore, and provides engineering practical experience for similar projects.