机器学习预测方法在带压开采煤层底板破坏深度预测评价中得到广泛应用，然而在预测模型构建过程中常常存在着实测数据获取成本高、采集难度大以及随机性强等问题，基于少量样本构建的模型预测的性能往往受到预测精度和泛化能力的严重制约。通过文献调研，收集获得50组实测数据样本，引入MTD （Mega-Trend-Diffusion）类分布虚拟样本生成技术，生成虚拟样本，进一步实现煤层底板破坏深度实测样本的扩充增强；采用SaDE-ELM （Self Adaptive Differential Evolution- Extreme Learning Machine）、GA-PSO-BP （Genetic Algorithm-Particle Swarm Optimization-Back Propagation Neural Network）、BP （Back Propagation Neural Network）等机器学习算法，构建虚拟样本增强前后煤层底板破坏深度预测模型，并对增强前后的模型预测精度进行对比分析。结果表明：所提方法生成的虚拟样本与实测样本的分布基本一致；采用虚拟样本增强后的预测模型精度均显著提高，其中基于MTD类分布小样本增强的PCA-SaDE-ELM （Particle Swarm Optimization-Self Adaptive Differential Evolution-Extreme Learning Machine）预测模型的预测效果最优，增强前后误差可降低42.95%～51.27%。采用MTD类分布生成技术进行小样本增强，并构建带压煤层底板破坏深度预测模型能够更加精确地预测多因素影响下煤层底板破坏深度，通过与规范经验预测结果和滑移线场理论预测结果进行对比分析，所提方法预测得到的云驾岭矿19105工作面底板破坏深度相对较大，有利于工作面安全生产管理。相关研究成果对奥灰承压水上煤层的安全高效回采提供了有利支撑。

The prediction method of machine learning has been widely used in the prediction and evaluation of the failure depth of coal seam floor in pressure mining. However, there are often some problems in the construction of the prediction model, such as high acquisition cost, difficulty in collecting and strong randomness of the measured data. The prediction performance of the model built based on a small number of samples is often severely restricted by the prediction accuracy and generalization ability. Through literature research, 50 sets of measured data samples were collected, and MTD similar distribution virtual sample generation technology was introduced to generate virtual samples to further expand and enhance the measured samples of coal seam floor failure depth.Machine learning algorithms such as ADE-ELM, GA-PSO-BP and BP were used to build a prediction model of coal seam floor failure depth before and after virtual sample enhancement, and the prediction accuracy of the model before and after enhancement was compared and analyzed. The results show that the distribution of virtual samples generated by this method is basically consistent with that of measured samples. The accuracy of the prediction models enhanced with virtual samples is significantly improved, among which the PCA-ADE-ELM prediction model enhanced with small distributed samples of MTD class has the best prediction effect, and the error of the enhanced model can be reduced by 42.95%～51.27%. MTD similar distribution virtual sample generation technology is used to enhance small samples, and the prediction model of failure depth of coal seam floor under pressure can be built to more accurately predict the failure depth of coal seam floor under the influence of multiple factors. Through comparison and analysis with the standard empirical prediction results and the slip line field theory prediction results, the failure depth of 19105 working face of Yunjialing Mine predicted by this method is relatively large, which is conducive to the safe production management of working face. The relevant research results provide favorable support for the safe and efficient mining of confined above-water coal seam of Ordovician limestone.