Seismic phase picking of coal mine microseismic data based on lightweight CNN

In coal mining, the high-noise underground environment poses significant challenges to the accuracy and stability of microseismic (MS) signal first-arrival picking. This study proposes a high-robustness automatic phase picking method based on a lightweight Convolutional Neural Network (CNN). The method uses a multi-layer convolution and adaptive pooling structure to efficiently extract waveform features across temporal scales, and introduces a regression prediction strategy to directly output the first-arrival time, eliminating the quantization error induced by classification-based methods. The model was validated on 1,791 manually labelled single-component records of 597 microseismic events from the 6305 working face of Xinjulong Coal Mine, Shandong Province. On the test set, the proportion of first-arrival errors not exceeding 100 ms reached 80.27%. Compared with the traditional STA/LTA method, the overall relative error was significantly reduced, with the average relative error decreasing by approximately 50 ms. The model maintained high picking accuracy and stability under high-noise and complex geological conditions, and under high-SNR conditions it effectively reduced the subjective bias of manual picking. The findings provide reliable technical support for high-precision localization of microseismic events and the early warning of dynamic hazards in coal mines.

In recent years, with the continuous escalation in the depth and intensity of coal mining, dynamic disasters—primarily coal and gas outbursts and rockbursts—have occurred frequently, posing a fundamental challenge to the safety and sustainable development of the industry (Jiang et al., 2014). These disasters are characterized by high spontaneity and extreme destructiveness, often resulting in severe casualties and substantial economic losses within a short duration (Qi et al., 2019; Pan et al., 2023). As an advanced paradigm for monitoring coal mine dynamic disasters, microseismic (MS) monitoring technology captures MS waves generated by rock mass fracturing, enabling the real-time perception of internal stress variations and structural damage within the underground rock mass (Cong et al., 2016; Yanchun and Lei, 2023). Grounded in elastic wave propagation theory, this technology utilizes underground sensor arrays to capture weak seismic signals induced by mining disturbances. Subsequent analysis and processing allow for the inversion of the location, magnitude, and evolutionary trends of these fractures, thereby providing critical data support for the early warning and mitigation of dynamic disasters in coal mines (Nan et al., 2017; Zhao et al., 2012; Jiang et al., 2006).

Within the framework of coal mine MS monitoring, MS signal picking serves as the pivotal stage for achieving effective monitoring (Zhu et al., 2018). However, this process is frequently compromised by the complex operational environment underground. Intense background noise, generated by the vibration of mining machinery and the operation of ventilation systems, often severely obscures (or “submerges”) the weak effective signals produced by rock mass fracturing. Consequently, the interpretation of rock mass fracture information based on high-quality MS signals constitutes the essential data foundation for early disaster warning. It provides a vital basis for identifying stress concentration zones and assessing hazardous areas, thereby facilitating the scientific formulation and implementation of prevention and control strategies, such as pressure-relief mining and reinforced support (Nan et al., 2017).

Extensive research has been conducted globally concerning microseismic (MS) signal first-arrival picking, spanning both algorithmic refinement and practical engineering implementation. In the early stages, efforts predominantly focused on the adaptation and enhancement of established seismological techniques, such as the Short-Term Average to Long-Term Average ratio (STA/LTA) (Allen, 1982; Baer and Kradolfer, 1987) and the Akaike Information Criterion (AIC) (Leonard, 2000; Maeda, 1985). These classical algorithms facilitate arrival detection by constructing sliding window energy ratios or phase waveform matching criteria. While demonstrating stability in low-noise environments, their picking accuracy significantly deteriorates in high-interference coal mine scenarios due to their inherent reliance on manual threshold configurations. Since the beginning of the 21st century, with the advancement of time-frequency analysis, non-stationary signal processing methods—including the Hilbert-Huang Transform (HHT) and Wavelet Transform—have been introduced. These approaches have effectively enhanced feature extraction capabilities for complex waveforms; however, challenges regarding computational efficiency and end-point effects remain unresolved. Recently, deep learning technologies have achieved remarkable progress in phase picking. A representative paradigm is PhaseNet (Zhu and Beroza, 2019), which utilizes an end-to-end convolutional neural network (CNN) architecture to achieve automated detection of P-wave and S-wave arrivals within seismic waveforms.

Beyond PhaseNet, deep-learning pickers have developed rapidly. Ross et al. proposed the Generalized Phase Detection (GPD) framework, a CNN-based classifier for P/S/noise discrimination on short waveform windows (Ross et al., 2018). Mousavi et al. introduced EQTransformer, which couples convolutional, recurrent, and hierarchical self-attention layers for simultaneous event detection and phase picking (Mousavi et al., 2020). Münchmeyer et al. systematically benchmarked mainstream deep-learning pickers across multiple datasets and showed that picking performance is strongly dataset-dependent, with PhaseNet and EQTransformer leading on most benchmarks (Münchmeyer et al., 2022).

For mine and microseismic monitoring, where waveforms are typically single- or few-component, of higher frequency, and of lower signal-to-noise ratio (SNR) than regional earthquake records, dedicated networks have been developed, including Conv-LSTM-UNet for coal mine event identification and first-arrival picking (Chen H. et al., 2023), convolutional-recurrent hybrids for real-time picking of rock micro-fracture signals (Chen B-R. et al., 2023), improved U-Net variants for mine microseismic waveforms (Wang et al., 2025), physics-constrained deep learning for coal mine P-wave arrival picking (Qin et al., 2025), multi-scale CNN-Transformer hybrid networks for engineering microseismic monitoring (Song et al., 2025), denoising-assisted pickers that combine adaptive characteristic functions with AIC (Zhang et al., 2025), and parameter-efficient transfer learning that adapts earthquake-trained pickers to microseismic data (Abdullin et al., 2025). Nevertheless, most existing deep pickers are designed for three-component, 100-Hz earthquake recordings and involve millions of parameters, and lightweight architectures tailored to single-component, high-sampling-rate coal mine microseismic records that can run on resource-constrained edge devices remain scarce. This gap motivates the lightweight CNN developed in this study.

Consequently, to address the limitations of existing MS signal picking methods—specifically their insufficient anti-interference capability in intense noise environments and poor adaptability to complex geological conditions—this study employs a CNN-based deep learning approach. Focusing on MS signal picking under high-background-noise scenarios in coal mines, first-arrival picking is performed on 1,791 manually labelled single-component MS records (597 events) collected from a mine in Shandong Province. This research aims to achieve rapid and precise identification of MS arrivals, thereby establishing a critical foundation for subsequent high-precision source localization.

Engineering background

The Xinjulong Coal Mine is situated in the deep part of the Luxi Coalfield. The primary mining target, the No. 3 coal seam, is buried at depths of 800–1200 m with an average thickness of 6.2 m. The seam has a dip angle of 15°–30°, classifying it as a typical large-dip-angle “three-soft” coal seam (characterized by a soft roof, soft floor, and soft coal body). Engineering exploration revealed the F311 fault, which has a throw of up to 16 m and an extension length of 92 m; this is the largest fault encountered in the mine’s history (as shown in Figure 2). Influenced by the superposition of tectonic stress and mining disturbances, the stress concentration coefficient in the working face reaches as high as 2.8–3.5, making it highly susceptible to rockburst events. The mining process utilizes fully mechanized top-coal caving. During the period when the working face passes through the fault, “step-wise adjustment + manual opening” techniques are required due to the development of the fault fracture zone. This leads to frequent rock mass fracturing, with MS event energy being 3–5 times higher than in normal areas.

Research data were obtained from the underground MS monitoring system of the 6305 working face at the Xinjulong Coal Mine. This system comprises 12 underground triaxial accelerometers (sensitivity ≥100 mV/g, frequency response 1–5000Hz) forming a planar monitoring network to achieve full coverage of the working face and fault zones (Figure 3). The monitoring period spanned from June 11, 2023, to October 25, 2023 (137 days in total), during which 3,963 MS events were recorded, with magnitudes ranging from - 0.5 to 1.6. Of these, 597 events with reliable manual P-wave labels (1,791 single-component waveform records in total) were compiled into the dataset used for model development in this study.

The objective of seismic first-arrival picking is to accurately identify the temporal point at which the initial valid wave arrives within a seismic signal. Traditional approaches typically rely on time-series-based feature extraction and regression models. However, with the rapid advancement of deep learning technologies, Convolutional Neural Networks (CNNs) have emerged as robust tools in the field of seismic signal processing, owing to their exceptional performance in handling both temporal and image data. This section introduces an innovative CNN-based methodology that transforms 1-dimensional (1D) seismic signals into 2-dimensional (2D) images and subsequently utilizes CNN architectures to perform regression-based prediction for the first-arrival position.

The core philosophy of the proposed method involves the transformation of raw 1D seismic signals into 2D image representations. This transformation process not only preserves the intrinsic temporal characteristics of the signal but also effectively maps the frequency information into the spatial structure of the image. Consequently, it provides a richer multi-dimensional feature space for subsequent feature extraction and first-arrival picking.

Signal modeling and transformation

In the context of coal mine microseismic (MS) monitoring, the discrete time-series signal x [ n ] received by sensors can be regarded as the result of the source signal propagating through the underground medium and being superimposed with background noise. The mathematical model for this process is defined in Equation 1: where n represents the discrete sampling index, s n denotes the source wavelet generated by rock mass fracturing g n is the Green’s function along the propagation path from the source to the sensor, * represents the convolution operator, and η n signifies the additive background noise. Under the complex operational conditions of underground coal mines, due to vibrations from mining machinery and electromagnetic interference, η n is typically manifested as intense noise with a non-Gaussian distribution. This results in an extremely low Signal-to-Noise Ratio (SNR), whereby the effective signal s n is severely obscured.

To overcome the limitations of low feature discernibility associated with 1D time-domain signals under intense noise backgrounds, this study performs preprocessing—including zero-mean normalization and standardization—on the raw signals to eliminate scale discrepancies. More importantly, a dimensionality elevation strategy is introduced. We map the 1D time-series into 2D time-frequency feature maps. This process can be formalized as a nonlinear transformation operator τ, as shown in Equation 2: where x ∼ is the standardized signal sequence and X is the generated 2D feature matrix (image). The core theoretical rationale for this transformation is that the first arrival of an MS signal is characterized not only by amplitude jumps in the time domain but also by instantaneous shifts in energy distribution within the frequency domain. By mapping the temporal dimension to the horizontal axis and frequency components to the vertical axis, Convolutional Neural Networks (CNNs) can leverage their superiority in image texture recognition to simultaneously capture the local singularity (temporal mutations) and spectral structure characteristics (frequency energy clusters) of the signal. This achieves a significant “denoising gain” during the feature extraction stage. Specifically, the transformation τ is implemented as the short-time Fourier transform (STFT) with a 64-sample (128 ms) window and an 8-sample (16 ms) hop; the logarithmic magnitude of the STFT is normalized to zero mean and unit variance, so that each 1,600-sample trace yields a 33 × 193 time-frequency image serving as the network input.

Network architecture design

To address the specific requirements of real-time performance and hardware deployment in coal mine MS monitoring, traditional deep networks (such as ResNet) often suffer from parameter redundancy and are highly susceptible to overfitting when MS samples are limited. Consequently, this study constructs a specialized lightweight CNN architecture. The design rationale for this architecture is primarily based on the waveform characteristics of MS signals. Superiority of Shallow Features over Deep Semantics: P-wave arrivals belong to low-level waveform mutation features rather than complex semantic features. Excessively deep networks tend to lose these critical transient details.

Balance between Noise Robustness and Sensitivity: The model adopts a multi-stage hierarchical convolutional structure.

The first layer (Block 1): Utilizes a relatively large 5 × 5 convolutional kernel. Due to the complex noise environment underground, a larger receptive field acts as a smoothing filter, suppressing high-frequency random noise while extracting large-scale contours of the signal.

Subsequent layers (Block 2-n): Transition to smaller 3 × 3 convolutional kernels. This design is intended to focus on local fine-grained details and capture the waveform onset points (i.e., the mutation features at the arrival time) that are preserved after the initial denoising stage.

The network comprises three convolutional blocks with 16, 32, and 64 channels and kernel sizes of 5 × 5, 3 × 3, and 3 × 3, respectively; each block applies a convolutional layer followed by batch normalization, a ReLU activation, and 2 × 2 max pooling (Figure 1). An adaptive average pooling layer and three fully connected layers (with ReLU activations and a dropout rate of 0.5 after the first two) then regress the first-arrival time. In total, the network contains only 58,529 trainable parameters (≈0.059 M, ≈0.22 MB in 32-bit floating point) and requires approximately 4.7 M multiply–accumulate operations (≈9.4 M FLOPs) per trace, which confirms its lightweight nature and suitability for real-time deployment on resource-constrained mine monitoring hardware.

Traditional methods often treat phase picking as a classification problem (i.e., determining whether a time segment contains the first arrival), which introduces quantization errors determined by the time window length. To break through this precision bottleneck, this study adopts an end-to-end continuous regression strategy. Specifically, the fully connected layer of the network directly outputs a continuous value t ^, representing the predicted first-arrival time. The model is trained using Mean Squared Error (MSE) as the loss function, defined in Equation 3: where t p r e d is the model-predicted time and t t r u e is the manually labeled ground truth. By minimizing the Euclidean distance between the predicted and true values, the model can approximate the actual first-arrival point on a continuous time axis, thereby eliminating the discrete errors caused by human-defined grids and achieving higher precision in automated picking.

MSE is adopted as the loss function for three reasons. First, under the regression formulation the picking task reduces to estimating a continuous arrival time, and MSE is the maximum-likelihood estimator when labelling errors follow an approximately Gaussian distribution, which is consistent with the near-symmetric, bell-shaped distribution of manual picking residuals observed in our dataset. Second, compared with the mean absolute error (L1) loss, the quadratic form of MSE imposes increasingly heavy penalties on large deviations, suppressing gross picking errors that are the most damaging to downstream source location. Third, MSE is smooth and continuously differentiable everywhere, providing stable gradients and faster convergence near the optimum than L1.

It should be noted that picking errors of opposite signs may, in principle, affect source location asymmetrically. In our experiments the test-set residuals are approximately zero-centred without a systematic early or late bias, so the symmetric MSE loss does not introduce a directional bias into the located hypocentres; designing asymmetric losses that explicitly couple the picking objective with the location objective is left as a direction for future work.

Model training and key evaluation metrics

The training and testing data for this study were sourced from the MS monitoring records of the 6305 working face at the Xinjulong Coal Mine in Shandong Province. A total of 3,963 MS events were monitored, of which 597 events (1,791 single-component waveform records) were manually labelled and used in this study, with magnitudes ranging from - 0.5 to 1.6 and a sampling rate of 500 Hz, covering the typical MS activity characteristics during the mining process. The dataset consists of waveform signal files and their corresponding first-arrival time labels. Due to variations in energy release intensity and monitoring distances across different MS events, the amplitude range of the raw signals fluctuates significantly. Such amplitude discrepancies might cause the model to over-focus on the features of strong signals while neglecting information from weak signals. Therefore, normalization was performed on all signals to ensure that signals of different intensities participate in model training on the same scale, thereby enhancing the fairness of feature learning.

The labelled dataset comprising 1,791 single-component waveform records was randomly divided into training, validation, and test sets at a ratio of 60%/20%/20% (1,074, 358, and 359 records, respectively; the random seed was fixed to ensure reproducibility), and the validation set was used for early stopping. Model training was implemented based on the PyTorch deep learning framework, with the core parameter configuration shown in Table 1.

The Adam algorithm was selected as the optimizer (Kingma and Ba, 2015; Ruder, 2016). Its combination of momentum and adaptive learning rates allows the model to effectively adapt to non-stationary objective functions. The initial learning rate was set to 0.001, and the batch size was set to 32. This configuration fully utilizes GPU memory resources while enhancing the stability of parameter updates through mini-batch gradient descent. The total training duration consisted of 100 epochs, during which an early stopping mechanism was implemented to save the model parameters with the minimum validation loss, thereby mitigating the risk of overfitting. Mean Squared Error (MSE) was adopted as the loss function. By minimizing the squared error between the predicted values and the true first-arrival times, the regression capability of the model for continuous target variables was optimized.

To comprehensively evaluate the precision of the first-arrival picking, Mean Absolute Error (MAE), minimum error, and graded accuracy were utilized as the core evaluation metrics. The specific results obtained from the test set are presented in Table 2. All errors are reported in sampling points; at the 500-Hz sampling rate of the monitoring system, one sampling point corresponds to 2 ms, so the ≤50-sampling-point criterion in Table 2 is equivalent to ≤100 ms, consistent with the accuracy reported in the Abstract.

In terms of overall performance, the model achieved an MAE of 45.61 sampling points on the test set, indicating that the picking deviation across samples was well-controlled. The accuracy for errors within ≤50 sampling points reached 80.27%, demonstrating that the picking precision for the majority of samples satisfies the fundamental requirements for coal mine microseismic localization.

As illustrated in Figure 4, the distribution of error differences exhibits a distinct left-skewed characteristic, with most of the data points falling within the negative region to the left of the zero line (i.e., the CNN error is lower than the STA/LTA error). Statistical analysis reveals that the mean relative error difference is −24.98 (indicated by the green dash-dotted line). Considering the system sampling rate of 500 Hz, this implies that, in a statistical sense, the picking precision of the CNN method is approximately 50 ms higher than that of the STA/LTA method. This systematic negative shift robustly proves that the data-driven CNN model can capture first-arrival features more accurately than traditional methods reliant on energy thresholds, significantly reducing picking errors in the complex and noisy environments of underground coal mines.

Validation with real data and performance analysis

To comprehensively evaluate the generalization performance and engineering applicability of the proposed lightweight CNN model in complex mine environments, this study conducted a systematic validation based on a real microseismic (MS) monitoring dataset from the 6305 working face of the Xinjulong Coal Mine in Shandong Province. The experimental design encompasses a complete evaluation process, ranging from macro-statistical metric analysis to detailed examinations of typical waveform morphologies, and performs an in-depth comparison between the proposed method and the classic Short-Term Average to Long-Term Average (STA/LTA) ratio method, the Akaike Information Criterion (AIC) picker, and a PhaseNet-style deep-learning baseline. It should be noted that the compared methods follow different paradigms: STA/LTA and AIC are threshold- or statistics-based pickers; PhaseNet performs per-sample probabilistic classification, in which a U-Net outputs the probability of each sample being the P-wave onset (trained with Gaussian-shaped labels) and the pick is taken at the probability peak; whereas the proposed lightweight CNN performs whole-trace regression and directly outputs the onset time. Because the published PhaseNet model was trained on three-component, 100-Hz regional earthquake recordings, directly applying its pre-trained weights to single-component, 500-Hz coal mine microseismic data suffers from severe domain shift; for a fair comparison, a PhaseNet-style U-Net was therefore retrained on the same training set used by our model (identical 60%/20%/20% split, Gaussian labels with σ = 20 ms) and all methods were evaluated on the identical test set.

Evaluated on the identical test set, the proposed lightweight CNN achieved the lowest mean absolute error (MAE). The classical AIC picker and the retrained PhaseNet-style U-Net yielded markedly higher MAEs of 79.4 and 86.6 sampling points, respectively, and their accuracies within the engineering-relevant ≤100 ms tolerance (65.2% and 66.3%) remained well below that of the proposed method (78.0%). These results show that neither a refined statistics-based picker nor a generic deep-learning picker outperforms the proposed model on single-component, high-sampling-rate coal mine microseismic data, underscoring the benefit of an architecture tailored to this setting.

Statistical error analysis

Based on the prediction results of 1,189 MS samples in the test set, Figure 5 illustrates the performance discrepancies between the two methods across various evaluation dimensions. Mean Absolute Error (MAE), as a key metric for measuring the overall deviation of predictions, intuitively reflects the convergence capability of the model. Statistical results indicate that the CNN model achieved an MAE of 55.38 sampling points. Compared to the 65.71 points of the STA/LTA method, the overall error was reduced by approximately 15.7%. This demonstrates that through the end-to-end regression strategy, the CNN effectively circumvents the systematic bias introduced by sliding window discretization in traditional methods.

Furthermore, the accuracy distribution under various error thresholds (Figure 5B) reveals the advantages of the model in refined picking. Under an extremely stringent error tolerance (error <5 points), the accuracy of the CNN reached 10.6%, outperforming the 7.7% of STA/LTA. As the tolerance threshold was relaxed to 50 points (equivalent to 100 ms), which is commonly used in engineering, the accuracy of the CNN significantly improved to 78.0%, whereas that of STA/LTA was only 62.9%. The aforementioned data suggest that the proposed model not only aligns more closely with the true first arrivals in terms of overall trends but also demonstrates stronger competitiveness within the “small-error” intervals required for high-precision localization.

Comparison of typical phase picking effects

To explore the specific performance of the model under varying Signal-to-Noise Ratio (SNR) conditions, this study selected representative high-SNR and low-SNR waveforms for visual analysis. The SNR of each trace is quantified as the power ratio between a signal window after the manually labelled first arrival and a noise window before it: SNR = 10·log 10 (P signal /P noise) dB, where P signal is the mean power of the demeaned 100-sample (200 ms) window starting at the labelled onset t 0, and P noise is the mean power of the demeaned 300-sample (600 ms) window ending 10 samples (20 ms) before t 0; the 20-ms guard interval excludes emergent signal energy from the noise estimate. Computed with this definition, the SNR of the labelled dataset has a median of 17.5 dB (mean 15.8 dB), and the high- and low-SNR examples in Figures 6, 7 are annotated with their measured SNR values. Picking Precision in High-SNR Environments Figure 6 illustrates microseismic (MS) signals representative of high-SNR environments. These signals are characterized by low background noise, strong P-wave onset energy, and distinct waveform contours. Under such ideal conditions, the traditional STA/LTA method (green dashed line) can typically achieve effective picking. However, constrained by the sensitivity settings of the long- and short-term window energy ratios, it is susceptible to “phase drift” at local energy fluctuations near the first-arrival point. This often results in picking positions that slightly lag or lead the ground truth (grey solid line).

In contrast, the CNN model (red dashed line) does not rely solely on changes in amplitude energy. Instead, it extracts the textural and morphological features of the waveform through convolutional layers, allowing it to lock onto the onset point more precisely during the gradual initial phase of the waveform. For instance, in the second sample, the predicted values of the CNN exhibit extremely high alignment with the manual labels. This demonstrates the potential of the model to serve as a substitute for manual picking in high-quality signal processing, effectively eliminating the subjective visual errors associated with manual operations. 2. Robustness Validation in Intense Noise Environments Figure 7 illustrates microseismic (MS) signals representative of low-SNR conditions, which constitute the most common complex operational scenarios in underground coal mines. These signals are characterized by high-amplitude background noise (e.g., shearer cutting noise and conveyor vibrations) preceding the arrival of the first wave. Consequently, the onset characteristics of the effective signals are severely obscured or even submerged, with waveforms exhibiting non-stationary random fluctuations.

In this scenario, the limitations of the STA/LTA method are fully exposed. Since the high energy of background noise frequently induces mutations in the characteristic function, the STA/LTA method is highly susceptible to false triggers before the actual first arrival, leading to significant picking bias. As shown in the third signal in Figure 7, the picking position of STA/LTA (green dashed line) deviates significantly from the true onset point.

Conversely, the lightweight CNN model proposed in this study (red dashed line) exhibits excellent anti-noise robustness. Benefiting from the joint learning capability of deep networks regarding local features and contextual information, the model can “filter” specific waveform patterns associated with P-wave arrivals out of intense noise backgrounds. Even in extreme low-SNR samples where identification is difficult for the human eye—such as the first signal sample—the CNN still yields predictions highly approximate to the ground truth, with errors maintained within a minimal range. These results robustly demonstrate that the proposed method maintains stable phase recognition capabilities under intense noise interference, effectively addressing the critical engineering challenge of “difficulty in extracting weak signals” in coal mine MS monitoring.

To address the challenges of insufficient anti-interference capability in intense noise environments and poor adaptability to complex geological conditions in coal mine microseismic (MS) monitoring, this study proposed a phase picking method based on a small-size CNN. Through theoretical modeling, architecture design, and validation with engineering data, the following conclusions were reached: The constructed CNN architecture effectively adapts to the temporal characteristics of coal mine MS signals. By utilizing multi-layer convolutional operations, the model automatically extracts multi-scale waveform features, such as amplitude mutations. Furthermore, the employment of a regression model to directly output continuous temporal predictions circumvents the quantization errors inherent in classification-based discretization, thereby providing robust technical support for high-precision first-arrival localization.

Engineering validation demonstrates that the proposed method performs excellently under intense noise and complex geological conditions. Experimental results based on 1,791 manually labelled single-component MS records (597 events) from the Xinjulong Coal Mine in Shandong Province indicate a Mean Absolute Error (MAE) of 45.61 sampling points on the test set, with an accuracy for errors ≤50sampling points reaching 80.27%. This satisfies the fundamental precision requirements for MS localization in coal mines. Real data validation further confirms that in high-noise and low-signal-to-noise ratio (SNR) scenarios, the model can precisely identify the first arrivals of effective signals “submerged” by noise; in high-SNR scenarios, it mitigates the impact of subjective errors and visual fatigue associated with manual picking, achieving more stable and accurate localization.

The proposed method enables the rapid and precise picking of MS first arrivals, effectively covering various signal-quality scenarios. It establishes a reliable data foundation for subsequent high-precision localization calculations and possesses significant practical value for enhancing early warning capabilities for coal mine dynamic disasters—such as coal and gas outbursts and rockbursts—while ensuring the safety of deep mining operations.

Future research will focus on expanding the dataset across diverse geological conditions, optimizing the lightweight design of the model, and improving adaptability and real-time performance in extremely complex operational environments to provide comprehensive technical support for the industrial application of MS monitoring technology.