煤岩单轴压缩破坏声发射轻量化三维卷积预测模型研究

Lightweight 3D convolution model for failure prediction of coal under uniaxial compression based on acoustic emission

声发射前兆信息能揭示煤岩内在破坏规律,因此基于声发射前兆信息建立判识模型是煤岩破坏监测预警的核心。基于DenseNet骨架,结合分组卷积(GC)与注意力机制中的“压缩-激励”模块(SE),构建能融合声发射时空特征的轻量化三维卷积预测模型。以红庆河矿3^(-1)煤层强冲击倾向性煤样为研究对象,在煤样单轴压缩过程中,采集不同速率加载下的声发射损伤破坏参数,并将参数预处理为能融合声发射时空特征的图像序列,作为模型输入样本,预测煤样破坏的危险等级,并采用迁移学习预测煤样破坏的剩余时间。结果显示:在预测煤样破坏危险等级的验证样本中,4种网络结构(DenseNet,DenseNet+GC,DenseNet+SE和DenseNet+GC+SE)的预测准确率均高于99.08%,高风险样本召回率均高于99.50%,表明三维卷积能有效捕获声发射时空信息。DenseNet+GC+SE网络的预测概率趋于单点分布,表明该模型能区分不同危险等级。DenseNet+GC+SE网络嵌入分组卷积与SE模块能在保证模型精度的同时,大幅下降模型复杂度与时间复杂度,提升模型效率。煤样破坏剩余时间的预测值与真实值的R2高达99.85%,表明DenseNet+GC+SE迁移模型能依据声发射信号有效预测煤岩破坏。分组卷积还能表征声发射特征多样性,SE模块能评估声发射特征重要性。

The precursor of acoustic emission(AE)can reveal the internal destructive rules of coal.Therefore,the discrimination model based on AE precursor has been a hot topic on the monitoring and prediction of coal failure.Based on DenseNet architecture combining with Group Convolution(GC)and Squeeze-and-Excitation module(SE)in attention mechanism,a lightweight three-dimensional(3D)convolution prediction model was proposed to integrate spatiotemporal information of AE signals.The coal samples from the No.3^(-1) coal seam of Hongqinghe mine,with high outburst-proneness,were used in the tests.AE signals were collected during the uniaxial compression of samples in different loading rates.Moreover,AE signals were preprocessed to generate spatial-temporal image sequences which were later used as model input to predict the destruction level of coal samples.Then,transfer learning was employed to predict the remaining failure time of coal samples.The results show that:in the validation samples for predicting the destruction level,all the networks(DenseNet,DenseNet+GC,DenseNet+SE and DenseNet+GC+SE)can obtain more than 99.08%prediction accuracy.The recall rate of high-risk samples prediction is higher than 99.50%,which indicates that 3D convolution can effectively capture spatial-temporal information of AE.The prediction probability of the DenseNet+GC+SE network is exhibited as unitary distribution,indicating the model can distinguish different destruction levels.Group Convolution and SE module can retain model accuracy while reduce the structure and time complexity,which improve DenseNet+GC+SE network efficiency greatly.The R2 between the predicted value and the true value of the remaining failure time is 99.85%,which further proves that DenseNet+GC+SE transfer model can effectively predict coal failure based on AE signals.The diversity of AE features is represented by GC,while the importance of AE features is evaluated by SE module.