为解决矿山应急救援钻孔作业过程中井涌井漏事故预警预测困难等问题，建立了基于机器学习的钻进过程井涌井漏事故预警预测模型。首先对井涌井漏事故发生初期时的钻进参数进行事故表征参数分析；其次对事故表征参数进行数据清洗处理，在此基础上，通过XGBoost事故诊断预警模型对井涌井漏事故进行早期诊断识别；随后建立PSO-LSTM事故发展预测模型，对事故发生后的孔底压力参数发展趋势进行预测，提前掌握钻进事故发展状态；最后通过实际钻进数据对预警预测模型的有效性进行验证。结果表明：XGBoost事故诊断预警模型能根据总池体积、立管压力、出入口流量差和动力头负荷这4种钻进参数的异常变化，快速准确诊断钻进过程中的井涌井漏事故；PSO-LSTM事故发展状态预测模型能充分学习孔底压力参数发展规律，综合$ {E_{{mathrm{MAP}}}} $、$ {E_{{mathrm{MA}}}} $、$ {E_{{mathrm{RMS}}}} $、$ {R^2} $这4项误差评价指标，PSO-LSTM模型相较于BP、RNN、SVM模型的预测性能最佳，能准确预测事故发生之后的孔底压力发展趋势，提前掌握井涌井漏事故的严重程度和发展态势。研究结果丰富了钻进过程井涌井漏事故预警预测方法，提高了矿山事故地面救援的可靠性，对矿山应急救援钻孔作业过程中事故控制有着借鉴与指导作用。

In order to solve the problems such as the difficulty in early warning and prediction of kicks and lost circulation accidents during emergency rescue drilling of mine, a machine learning-based early for warning and prediction model of drilling process was established. Firstly, the accident characterization parameters of the drilling parameters in the early stage of kicks and lost circulation accidents were analyzed. Secondly, the accident characterization parameters were cleaned and processed. On this basis, XGBoost and early warning model was used to carry out the early diagnosis and identification of kicks and lost circulation accidents. Then, the PSO-LSTM accident development prediction model was established to predict the development trend of downhole pressure parameters after the accident, so as to understand the development status of drilling accidents in advance. Finally, the effectiveness of the early warning prediction model was verified by the actual drilling data. The results show that XGBoost and early warning model could quickly and accurately diagnose the kick and loss circulation accidents in the drilling process according to the abnormal changes of four drilling parameters, including the total tank volume, riser pressure, inflow-outflow differential, and power head load. The PSO-LSTM accident development status prediction model could fully learn the development law of downhole pressure parameters. With full consideration to the four error evaluation indicators, such as EMAP, EMA, ERMS and R2, the prediction performance of the PSO-LSTM models is the best compared with BP, RNN and SVM, capable of accurately predicting the development trend of the downhole pressure after the accident, thereby knowing about the severity and development situation of kick and lost circulation accidents. Generally, the research results enrich the early warning and prediction methods of kicks and lost circulation accidents in the drilling process, improve the reliability of surface rescue in mine accident, and have a reference and guiding effect on accident control during the emergency rescue drilling of mine.