This paper proposes a deep learning-based time-domain denoising method for micro-earthquake recordings by combining a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM). Based on micro-earthquake observation data from Zigong and Neijiang areas of Sichuan, the structural model and focal mechanism of the region are used to generate a synthetic noise-free dataset by numerical modeling, which is then combined with observed micro-earthquake noise to create a synthetic noisy dataset. A high-performance and stable denoising model is obtained through training of the deep learning network, demonstrating excellent generalization performance on the validation set. Compared with traditional methods, the proposed method demonstrates excellent denoising performance and better preserves the detailed characteristics of both the waveform and the spectrum of the noise-free signal. Application to micro-earthquake observation data of Zigong and Neijiang areas demonstrates the model’s strong denoising performance and generalization ability on real-world data.

提出一种基于卷积神经网络(CNN)和双向长短期记忆网络(BiLSTM)的深度学习模型, 用于时间域波形去噪。选取四川省自贡和内江地区的微震观测数据, 基于该地区的构造模型和震源机制进行数值模拟, 生成无噪声数据集, 并叠加观测微震噪声, 构建模拟含噪声数据集。通过深度学习网络的训练, 获得性能稳定且泛化能力强的去噪模型, 该模型在验证集上也表现优异。与传统去噪方法相比, 所提方法的去噪效果显著提升, 能够更好地保留信号的细节特征和频谱特征。将该模型应用于自贡和内江地区的实际微震观测数据, 结果表明能有效地去除实测数据中的噪声。