摘要
为深入了解煤岩的变形和破裂过程,建立基于声发射前兆信息的判别模型进行煤岩破坏的监测和预警。通过构建整合声发射时空特征的轻量级三维卷积煤岩破坏识别模型,研究煤岩不同破坏阶段识别模型的预测效果,并验证模型的泛化能力。在识别煤岩损伤危险阶段的验证样本中,煤岩破坏识别模型预测准确率为99.39%,高风险样本的召回率超99.2%,表明三维卷积模型能有效捕捉煤岩破坏声发射波形的时空耦合信息。ConDenseNet-SE模型通过知识蒸馏优化进一步降低模型过拟合程度并获得性能和准确率兼备的煤岩破坏识别模型,验证了优化后的ConDenseNet-SE模型在识别煤岩破坏及破坏预警方面的优越性。
In order to deeply understand the deformation and rupture process of coal samples,the early warning discrim⁃ination model of coal rock damage monitoring based on acoustic emission precursor information was established to pro⁃vide an important basis for mine safety production.By constructing a lightweight three⁃dimensional convolutional coal rock damage identification model integrating acoustic emission temporal and spatial features,the prediction effect of the identification model for different stages of coal rock damage was studied,and the model′s generalization ability was ver⁃ified.The prediction accuracy of the coal rock damage recognition model was 99.39%in the validation samples of iden⁃tifying the damage hazard stages of coal samples,and the recall rate of the high⁃risk samples was also more than 99.20%,which indicated that 3D convolution could effectively captured the coupled spatio⁃temporal information of the acoustic emission waveforms of coal sample damage.Moreover,the ConDenseNet with SE model could be optimized by knowledge distillation to further reduce the degree of model overfitting and obtain a coal damage recognition model with both performance and accuracy,which verified the superiority of the optimized ConDenseNet with SE model in identif⁃ying coal damage and damage warning.
作者
高贤成
GAO Xiancheng(Shaqu No.2 Coal Mine,Huajin Coking Coal Co.,Ltd.,Lüliang 033000,Shanxi,China)
出处
《隧道与地下工程灾害防治》
2024年第4期90-98,共9页
Hazard Control in Tunnelling and Underground Engineering
基金
国家自然科学基金面上资助项目(51874014)
关键词
深度学习
煤岩
变形破坏
知识蒸馏
三维卷积神经网络
deep learning
coal
deformation and failure
knowledge distillation
3D convolutional neural network
