The accuracy of microseismic source localization depends largely on the quality of the velocity model. Due to the anisotropy of the rock mass, the current uniform velocity model is no longer sufficient for high-precision localization. Additionally, the time-varying property of the velocity model will influence the accuracy of the estimated source location. Focusing on these challenges, we propose an iterative source location estimation and simplified anisotropic velocity inversion method based on the neural network algorithm and dynamic reduction of solution interval. We first introduce a simplified anisotropic velocity model and establish an objective function for source localization. The t-distribution is embedded in the neural network algorithm to increase the probability of jumping out of the local optimum. In each iteration, the solution interval is narrowed down and then the source location is estimated by the neural network algorithm. The initial solution interval is determined from the inversion results of the uniform velocity model. The performance of the proposed method is evaluated by the numerical and blasting experiments. The location accuracy of the proposed method is at least 40% higher than that of the conventional method. Test results indicate that our method is effective to locate the sources in the areas with heterogeneous and complex media.

Passive source seismic interferometry is a cost-effective geophysical method that converts noise signals into valuable information. The fidelity of the resultant common-shot gather is pivotal for effective imaging. The quality of reconstructed records via seismic interferometry directly correlates with the duration of background noise observation. However, practical applications often encounter difficulties in obtaining stable and usable long-duration observations of noise-based passive seismic records. Short-duration observations may introduce spurious physical events, thereby compromising the reliability of seismic wavefield imaging and geological interpretation. In this study, we introduce MDUNETR, an advanced passive data reconstruction network amalgamating Transformer and Multi-scale Dense Blocks (MDB) to enhance accuracy. By integrating Transformer and MDB, the network effectively captures both global and local information. Utilizing the MDUNETR network, we can reconstruct accurate passive source interferometric seismic records from short-duration noise interference signals. This overcomes the time limitations imposed by seismic interferometry on the original noise records. Theoretical data applications demonstrate the stability and fidelity of the seismic records reconstructed by this network, ensuring reliable results.

The pivotal role of seismic velocity inversion in oil and gas exploration and geological research has been widely acknowledged. However, conventional methods face challenges such as strong reliance on initial models and high computational costs. Based on the mode of seismic event generation, seismic data can be classified into active seismic data and passive seismic data, which collectively constitute the multisource data discussed in this article. Velocity inversion based on deep learning primarily relies on active seismic data, training neural networks to learn the mapping between seismic records and subsurface velocities. In contrast, signals in passive seismic data typically originate from noise at certain depths within the Earth, encompassing valuable information about deep subsurface structures that is crucial for velocity inversion, thus presenting a potential complement to active seismic data. This study proposes a seismic velocity inversion method that combines active and passive seismic data, utilizing deep learning techniques to adaptively integrate data from both sources, enabling joint inversion. The proposed neural network architecture combines transformer and convolutional neural network (CNN), enhancing the accuracy and robustness of velocity inversion.

The passive seismic interferometry (SI), harnessing ambient noise or unconventional seismic sources, has garnered widespread attention in the fields of Earth science and resource exploration. Conventional SI requires several assumptions to be satisfied, including uniform distribution of subsurface sources, an adequate number of sources, and long recording periods. However, these assumptions often fall short in real-world scenarios, leading to suboptimal reconstruction quality and subsequently impacting imaging results. Therefore, we propose a passive SI method with deep transfer learning. This method can extract real-time empirical Green’s functions directly from noisy datasets without prior preprocessing. Importantly, this technique goes beyond simple data retrieval; it demonstrates the ability to accurately reconstruct the entire wavefield. We establish a joint transformer-CNN network and conduct supervised training on intricate velocity models. Subsequently, we employ transfer learning to fine-tune the model, adapting it to new data that differ from the training dataset. Notably, our method requires only a small amount of data and can be applied to other velocity models without additional training for new neural networks. The validity of our method is demonstrated through a series of numerical experiments. Compared to conventional methods, real-time passive SI offers greater efficiency and accuracy in reconstructing subsurface structural response.

Recognition of microseismic events is the primary task of microseismic monitoring. Aiming at the low signal-to-noise ratio (SNR) of weak microseismic events and the high cost of labeling them, an unsupervised learning method for recognizing microseismic events is proposed. The method first recognizes microseismic events from monitoring data segments by simultaneous deep clustering and then performs a second clustering to further pick the first arrival times of the detected microseismic events by multistage deep clustering. The networks in this two-step clustering framework are built on a newly designed residual shrinkage dense block (RSDB). To better suppress the noise in microseismic data, RSDB adds a densely connected hybrid dilated convolution and an improved threshold module to the deep residual shrinkage network. The autoencoder built by the RSDB and U-Net architecture is combined with simultaneous deep clustering and multistage deep clustering to recognize microseismic events and their first arrival times, respectively. Finally, tests on the synthetic data and field microseismic data demonstrate the feasibility and superiority of the proposed method.

Locating microseismic sources in time is a challenging problem in microseismic monitoring. In order to improve the accuracy and efficiency of locating sources, this paper presents a method for locating microseismic sources using deep reinforcement learning. We first construct and train a convolutional autoencoder to preprocess the seismic records in the microseismic waveform database. Then, the problem of locating the source is described as a Markov decision process for the application of deep reinforcement learning. We decompose the task of locating the source into three subtasks and design the critical elements of deep reinforcement learning. Three agents independently learn optimal policies for their respective subtasks in the framework of a deep Q-network (DQN) and jointly determine the precise location of the microseismic source. Finally, we evaluate the proposed method using synthetic data generated from the Marmousi model and the 3D velocity model. The experiment results indicate that the proposed method can locate microseismic sources efficiently and accurately.

The performance of the microseismic monitoring technique depends greatly on the accuracy of microseismic event localization. Recently, machine learning (ML) methods have been extensively implemented for the localization of microseismic events. These neural networks are typically trained using numerous microseismic events labeled with known source locations. Obtaining enough microseismic events with good source locations can be difficult and costly. To overcome this shortcoming, we present a microseismic events localization method using semi-supervised generative adversarial networks (GANs). We utilize limited labeled seismograms and large amounts of unlabeled seismograms to train the semi-supervised GANs, thus improving the prediction ability of the networks. Finally, we evaluate the performance of the proposed method using synthetic microseismic data and field data. Comparison with the supervised learning methods on the same microseismic data shows that the proposed method can significantly improve the accuracy of locating microseismic sources in the lack of sufficient source labels.