Acoustic emission (AE) and electromagnetic radiation (EMR) monitoring technology have been widely used in coal and rock dynamic disaster monitoring. However, the prediction of coal and gas outburst using original signal directly may cause the false and missing alarms due to the failure to identify valuable potential information, which may lead to the decrease of prediction accuracy. This paper presents a new method to identify the precursor features of coal and gas outburst using two-dimensional Convolutional Neural Network (CNN). Empirical mode decomposition (EMD) was performed for AE or EMR signals, and the Intrinsic Mode Functions (IMFs) with Variance Contribution Rate (VCR) greater than 90% were merged into modified signal. Since the frequency of AE or EMR signal can change with different coal and rock fracture stages, one-dimensional modified signal was transformed into a two-dimensional time-frequency graph using the short-time Fourier transform (STFT). The time-frequency graph was input into CNN that recognizes the time-frequency characteristics of AE or EMR using SoftMax regression at the end of the network. Compared with the original signal, the precursor characteristics of modified AE or EMR signals can be better learned by CNN. The proposed method was applied to the Jinjia Coal Mine database which was identified as having outburst risk, and it was found that the recognition accuracy was better than other approaches. The proposed method achieves maximum recognition accuracy of 98.00% (AE), and 97.20% (EMR) in Jinjia Coal Mine database, respectively.

声发射（AE）和电磁辐射（EMR）监控技术已广泛用于煤炭和岩石动态灾害监测。然而，由于未能识别出有价值的潜在信息，直接使用原始信号来预测煤与瓦斯突出可能会导致误报和丢失警报，从而可能导致预测准确性的下降。本文提出了一种使用二维卷积神经网络（CNN）识别煤与瓦斯突出特征的新方法。对AE或EMR信号进行经验模式分解（EMD），方差贡献率（VCR）大于90％的本征模式函数（IMF）合并为修改后的信号。由于AE或EMR信号的频率会随煤和岩石破裂阶段的不同而变化，使用短时傅立叶变换（STFT）将一维修改后的信号转换为二维时频图。时频图被输入到CNN中，该CNN在网络末端使用SoftMax回归来识别AE或EMR的时频特性。与原始信号相比，CNN可以更好地学习修改后的AE或EMR信号的前驱特性。将该方法应用于确定有突出危险的金家煤矿数据库，发现识别精度优于其他方法。该方法在金家煤矿数据库中的最大识别精度分别为98.00％（AE）和97.20％（EMR）。