岩爆是一种由岩石破裂和弹射导致的动力灾害，会给矿山开采和隧道开挖带来巨大威胁。随着人工智能技术的不断发展，面向岩爆预测与人工智能交叉的方法已成为深部矿山开采和隧道工程领域的研究热点。由于岩爆突发性、随机性较强，相关数据难以及时收集，导致岩爆历史数据具有先天的缺陷性，其中小样本和类别不平衡问题最为显著。因此，本文以岩爆工程实例数据集为基础，结合深度学习与集成学习技术，针对岩爆数据集中的小样本和类别不平衡问题，提出了4种岩爆预测模型，分别解决了上述两种数据缺陷。以上述4种单一模型为基础提出了基于多目标麻雀搜索优化算法（Multi-objective Sparrow Search Optimization Algorithm, MSSA）的异质集成深度学习岩爆预测模型。在岩爆实例数据上验证了异质集成模型的有效性，同时对比分析了其精度。结果表明：使用异质集成深度学习模型相较于单一预测模型，具有更高的预测精度、鲁棒性和泛化性，能够有效解决岩爆数据集的小样本和不平衡问题。本文具体研究内容如下：

岩爆预测数据集的建立及数据分析。通过查阅岩爆智能预测研究文献和搜集历史岩爆数据并结合岩爆发生机理分析，考虑岩爆发生的内因和外因，选取了：岩石脆性指数、围岩最大切应力、岩石单轴抗压强度、岩石单轴抗拉强度、岩石应力系数、岩石弹性能量指数等6个指标作为岩爆预测的评价指标。根据实际需要选取岩爆四级分类法：I级(无岩爆)、II级(轻微岩爆)、III级(中级岩爆)、IV级(强烈岩爆)。建立了1个包括279组岩爆工程案例的数据集。对岩爆数据进行相关性分析以确定选取指标的合理性和有效性，并对数据进行了归一化处理。采用熵权TOPSIS方法对岩爆评价指标根据影响程度确定了权重，为后续智能模型建立奠定了数据基础。

为解决岩爆数据集中存在小样本且数据维度单一、数据高度非线性、数据自身特征少等3个问题，分别提出了基于多因子优化的麻雀搜索算法支持向量机（Progress Sparrow Search Algorithm for Optimizing Support Vectors, PSSA-SVM）岩爆预测模型、基于注意力机制优化的贝叶斯网络优化门控循环单元（Bayesian Network Optimization of Gated Recurrent Units based on Attention Mechanism, ABO-GRU）岩爆预测模型和基于函数拟合优化的因子分解机（Deep Learning Factor Decomposer Based on Function Fitting, FDeepFM）岩爆预测模型。针对岩爆数据小样本、维度单一的问题，提出了一种基于PSSA-SVM的岩爆预测模型。引入麻雀搜索算法（Sparrow Search Algorithm，SSA）优化支持向量机（Support Vector Machine，SVM）的参数配置，以提高预测模型的准确性和稳定性，融合多因子参数，对SSA算法和SVM的兼容性进行了优化。针对岩爆数据小样本、高度非线性的问题，提出了一种基于ABO-GRU的岩爆预测模型，通过加入注意力机制（Attention)对贝叶斯优化门控制循环单元（Bayesian Optimization Gated Recurrent Unit，BO-GRU）模型进行优化，以提高模型的预测性能。针对岩爆数据小样本、自身特征少的问题，提出了一种基于FDeepFM的岩爆预测模型，利用因子分解机（Factorization Machines，FM）和深度神经网络（Deep Neural Networks，DNN）来挖掘岩爆评价指标之间的深层隐性关系，利用函数拟合法快速激活，提高预测精度。通过提出的3种岩爆预测模型可有效的解决岩爆数据集中的小样本问题。

（3）为解决岩爆数据集中类别不平衡的问题，基于深度神经网络（Deep Neural Networks，DNN）技术，结合生成对抗网络（Generative Adversarial Networks，GAN）同时添加了自注意力机制（Self-Attention，SA）提出了一种基于自注意力机制深度生成对抗神经网络的混合深度学习岩爆预测模型（SA-GAN-DNN）。通过自注意力机制，提升数据特征维度，对输入数据中的关键信息进行权重处理，可有效的解决岩爆数据集中的类别不平衡问题。

（4）为同时解决岩爆数据集中存在的小样本和类别不平衡等两个问题，提出了基于PSSA-SVM、ABO-GRU、FDeepFM、SA-GAN-DNN的异质集成深度学习岩爆预测模型。异质集成深度学习岩爆预测模型能有效的集成多个不同类型的单一岩爆预测模型的优点，同时克服其不足。采用Blending集成方法，提出了一种基于MSSA的权值优化器。采用MSSA优化器对Blending方法中的集成权重值进行优化，可同时有效的解决岩爆数据集中小样本和类别不平衡的问题。

（5）以新汶孙村煤矿岩爆（冲击地压）和程潮铁矿两个典型岩爆案例数据作为工程验证。通过对各改进单一岩爆预测模型开展工程实例数据实验，验证了所提出的单一模型的有效性；通过对异质集成深度学习岩爆预测模型开展工程实例数据实验，验证了其有效性并具有最高的预测精度，为岩爆智能预测模型研究提供了一种新的方法。

Rock burst is a dynamic disaster caused by rock fracture and ejection, which poses a huge threat to mining and tunnel excavation. With the continuous development of artificial intelligence technology, the intersection of rock burst prediction and artificial intelligence has become a research hot spot in the field of deep-seated mining and tunnel engineering. Due to the suddenness and randomness of rock burst, it is difficult to collect related data in time, which leads to inherent defects in the historical data of rock burst, with the most prominent problems being small sample size and class imbalance. Therefore, this paper takes the engineering example data set of rock burst as the basis, combining deep learning and ensemble learning techniques, and proposes four rock burst prediction models to solve the small sample size and class imbalance problems in the data set. Based on the above four individual models, a heterogeneous integrated deep learning rock burst prediction model is proposed using the multi-objective sparrow search optimization algorithm (MSSA). The effectiveness of the heterogeneous integrated model is verified on the rock burst example data set, and its accuracy is compared and analyzed. The results show that using the heterogeneous integrated deep learning model instead of the single prediction model has higher prediction accuracy, robustness, and generalization ability, and can effectively solve the small sample data and class imbalance problems in the rock burst data set. The specific research content of this paper is as follows:

The establishment and data analysis of rockburst historical data set. By reviewing the literature on intelligent prediction of rockburst and collecting historical rockburst data and combining the analysis of the causal mechanism of rockburst, considering both internal and external factors, six evaluation indicators were selected: rock brittleness index, maximum shear stress of surrounding rock, uniaxial compressive strength of rock, uniaxial tensile strength of rock, stress coefficient of rock, and energy index of rock elasticity. According to the actual needs, the rockburst classification method of level IV was selected: level I (no rockburst), level II (mild rockburst), level III (moderate rockburst), and level IV (severe rockburst). A data set including 279 rockburst engineering cases was established. The correlation analysis was conducted on the rockburst data to determine the rationality and effectiveness of the selected indicators, and the data was normalized. The entropy weight TOPSIS method was used to determine the weights of the rockburst evaluation indicators based on their impact levels, laying a data foundation for the subsequent establishment of intelligent models.

To solve the three problems of small sample data, low data dimensionality, and highly nonlinear data in the rockburst data set, three rockburst prediction models were proposed, namely, the Progress Sparrow Search Algorithm for Optimizing Support Vectors (PSSA-SVM), the Bayesian Network Optimization of Gated Recurrent Units based on Attention Mechanism (ABO-GRU), and the Deep Learning Factor Decomposer Based on Function Fitting (FDeepFM). First, to solve the problems of small sample data and low data dimensionality in the rockburst data set, a rockburst prediction model based on PSSA-SVM was proposed. The PSSA algorithm was introduced to optimize the parameter configuration of the support vector machine (SVM) to improve the accuracy and stability of the prediction model, and the compatibility between the PSSA algorithm and SVM was optimized by combining multiple factor parameters. The PSSA-SVM model was constructed. Second, to solve the problems of small sample data and high nonlinearity in the rockburst data set, a rockburst prediction model based on ABO-GRU was constructed by adding an attention mechanism to the Bayesian optimization gated recurrent unit (BO-GRU) model to improve the prediction performance of the model. To address the problems of small sample size and limited intrinsic features in rockburst data, a rockburst prediction model based on FDeepFM was constructed, which utilizes Factorization Machines (FM) and Deep Neural Networks (DNN) to mine the deep latent relationships between rockburst evaluation indicators, and utilizes function approximation to quickly activate and improve prediction accuracy. By constructing three types of rockburst prediction models, the small sample problem in the rockburst data set can be effectively solved.

To solve the problem of class imbalance in the rockburst data set, a hybrid deep learning rockburst prediction model based on Self-Attention (SA) and Generative Adversarial Networks (GAN) was proposed, combining Deep Neural Networks (DNN) technology. The model adds a self-attention mechanism to enhance the feature dimension of the data, and weights the key information in the input data to effectively solve the problem of class imbalance in the rockburst data set.

To solve both the small sample and class imbalance problems in the rockburst data set simultaneously, a heterogeneous integrated deep learning rockburst prediction model based on PSSA-SVM, ABO-GRU, FDeepFM, and SA-GAN-DNN was proposed. The heterogeneous integrated deep learning rockburst prediction model can effectively integrate the advantages of multiple different types of individual rockburst prediction models and overcome their limitations. Using the blending integration method, a weight optimization algorithm based on MSSA was proposed. By optimizing the integration weight values in the blending method using the MSSA optimizer, the small sample and class imbalance problems in the rockburst data set can be effectively solved simultaneously.

(5) The two typical rockburst (impact pressure) case data of Xinwen Suncun coal mine and Chengchao Iron Mine were used as engineering verification. By conducting engineering example data experiments on the improved single rockburst prediction model, the effectiveness of the proposed single model was verified. By conducting engineering example data experiments on the heterogeneous integrated deep learning rockburst prediction model, its effectiveness and highest prediction accuracy were verified, providing a new method for the research of rockburst intelligent prediction model.