A Smart Method for Picking Slot Wave Dispersion Curves Based on AttentionU-Net

Abstract

The application discloses a kind of slot wave dispersion curve intelligent pickup methods based on Attention U-Net, first to coal mine working face slot wave seismic data is preprocessed and generates frequency-velocity energy graph;Attention U-Net network with attention gate module is constructed, shallow features are filtered using high-level semantic information, and the identification effect of weak energy slender dispersion ridge line is strengthened;Through forward and measured sample combination data enhancement, network training is completed;After the energy graph to be processed is input into network to obtain preliminary identification result, construct the objective function of fusion network probability, original energy, velocity smoothing and curvature constraint, and solve the continuous dispersion curve that meets the propagation law of slot wave through dynamic programming.The application can realize the automatic stable pickup of slot wave dispersion curve under the condition of low signal-to-noise ratio and multi-modal aliasing, reduce artificial dependence, improve pickup accuracy and batch processing efficiency, and is suitable for coal mine working face slot wave exploration and geological anomaly interpretation.

Description

Technical Field

[0001] This invention belongs to the field of coal mine seismic exploration, channel wave dispersion analysis and artificial intelligence seismic data processing technology, specifically involving an intelligent method for picking channel wave dispersion curves based on AttentionU-Net. Background Technology

[0002] Channel waves are seismic guided waves that propagate along the interior of coal seams. They are highly sensitive to changes in coal seam thickness, velocity anomalies, and tectonic disturbances, making them core data for the transparency detection of coal mine working faces. Dispersion curves reflect the variation of channel wave phase velocity / group velocity with frequency. Accurate acquisition of these curve waves is a prerequisite for subsequent coal seam parameter inversion and tectonic anomaly identification, directly determining the reliability of the detection results.

[0003] Existing methods for picking channel wave dispersion curves mainly rely on manual interactive picking, peak search, energy clustering, and rule tracking. When the ridges in the dispersion energy map are clear and the signal-to-noise ratio is high, the above methods can obtain usable results. However, the underground coal mine acquisition environment is complex. Affected by factors such as roadway noise, detector coupling differences, and multi-mode superposition, dispersion energy often exhibits discontinuities, spurious peaks, and mode mixing. Traditional methods are prone to false picking, missed picking, and curve breaks, resulting in low processing efficiency and results that depend on the operator's experience, leading to poor consistency in batch data processing.

[0004] In recent years, deep learning has begun to be applied to seismic dispersion curve picking, using frequency-velocity energy maps as images for segmentation and recognition. However, ordinary convolutional neural networks pay insufficient attention to the characteristics of elongated, weak-energy, and locally missing dispersion ridges, making the recognition results prone to breaks and noise points. Moreover, existing solutions are mostly designed for conventional surface waves such as near-surface Rayleigh waves, without considering the specific characteristics of coal mine channel waves, high-frequency bands, strong roadway noise, and severe multimodal aliasing, resulting in poor performance when directly applied. In addition, most deep learning solutions rely solely on the network's direct output results without constraining them by the physical laws of channel wave propagation, leading to abrupt changes in the picking results that do not conform to geological laws, resulting in insufficient usability for subsequent interpretation.

[0005] AttentionU-Net introduces an attention gating mechanism into the skip connections of U-Net, which can use high-level semantic information from the decoder to weight and filter shallow features from the encoder, thereby enhancing the feature response of the target region and suppressing background interference. It is suitable for segmenting elongated targets in complex backgrounds. However, there is currently no specific technical solution that combines AttentionU-Net with slot dispersion curve picking and provides corresponding post-processing for dispersion physical constraints. Summary of the Invention

[0006] To address the problems of existing slot wave dispersion curve picking methods that rely on human experience, are prone to mispicking, omissions, and discontinuities under low signal-to-noise ratio and multimodal coupling conditions, this invention provides an intelligent slot wave dispersion curve picking method based on AttentionU-Net, which achieves automated, high-precision, and physically consistent stable picking of slot wave dispersion curves.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent method for picking slot wave dispersion curves based on AttentionU-Net, specifically including the following steps: S1. Data Acquisition and Preprocessing: Seismic sources and detectors are set up in the roadways on both sides of the coal mine working face to acquire channel wave seismic data; the raw data is preprocessed sequentially by bad path removal, static correction, bandpass filtering, amplitude equalization, noise suppression and normalization to eliminate the influence of abnormal paths, coupling differences and energy imbalances; the preprocessed channel wave data is converted into a frequency-velocity energy map by dispersion spectrum transformation, the velocity scanning interval is set according to the velocity range of the coal seam and surrounding rock, and the energy map is normalized and background suppressed.

[0008] S2. Constructing the AttentionU-Net slot dispersion recognition network: The network adopts a U-shaped architecture of encoder-bottleneck layer-decoder. The encoder extracts texture and spatial structure features at different scales through multi-level downsampling. The bottleneck layer fuses global semantic information. The decoder restores spatial resolution step by step through multi-level upsampling. An attention gating module is connected in the skip connection path between each level of encoder and decoder. The output of the corresponding level of the decoder is used as the gating signal to spatially weight the features of the same level of the encoder, so that the network focuses on the slender dispersion ridge region and suppresses background energy, isolated noise and boundary artifacts.

[0009] S3. Network Training: Construct a training sample set containing theoretical forward modeling data, field-measured slot wave data, and corresponding dispersion curve labels; during training, incorporate data augmentation methods such as noise perturbation, bandwidth loss, energy attenuation, mode mixing, velocity perturbation, and local occlusion to improve the model's adaptability to complex acquisition conditions; the dispersion curve labels are in the form of curve centerline, curve neighborhood binary mask, or multimodal category mask; use binary cross-entropy, Dice loss, or Focal loss and their combinations as loss functions to complete iterative optimization of network parameters.

[0010] S4. Preliminary Identification: Input the frequency-velocity energy map to be processed into the trained AttentionU-Net network, and output the dispersion ridge probability map as the preliminary identification result of the dispersion curve.

[0011] S5. Physical Constraint Optimization: Combining the continuity of slot wave dispersion, velocity smoothness, curvature constraints, and maximum energy path constraints, a path optimization objective function is constructed. Based on the network output probability map and the original energy map, the global optimal path is solved through dynamic programming, shortest path search, or curve tracking algorithms to obtain a continuous, smooth, and uniform single-order dispersion curve that conforms to the propagation law of slot waves.

[0012] S6. Multimodal Differentiation and Quality Control: When it is necessary to pick multi-order slot wave dispersion curves, an iterative ridge suppression method is adopted: first, the first-order dispersion curve with the strongest energy is picked, its neighborhood energy is suppressed, and then steps S4 to S5 are repeated to pick the next-order curve; during the picking process, modal spacing constraints, cross-penalty terms, and continuity discrimination conditions are added to distinguish dispersion curves of different orders; outlier removal, local interpolation, smoothing, and manual verification are performed on the picking results, and the final output dispersion curve is as follows: The data is presented in a formal format, along with average probability, average energy, effective frequency band, and confidence level indicators. The final dispersion curve picking results are used for coal seam thickness inversion, velocity structure analysis, fault identification, collapse column identification, coal seam bifurcation and merging interpretation, or working face transparency detection.

[0013] Furthermore, in step S1, the dispersion spectrum transformation adopts any one or more combinations of phase shift method, fk transform, short-time Fourier transform, and generalized S transform; in the generated frequency-velocity energy map, the horizontal axis is the phase velocity or group velocity, the vertical axis is the frequency, and the energy intensity at the position corresponding to the pixel gray value or color value.

[0014] Furthermore, the path optimization objective function expression in step S5 is:

[0015] In the formula, The probability value of the dispersion ridge line output by the network. The energy values ​​of the original frequency-velocity energy map. The velocity change at adjacent frequency points The second-order difference component (curvature) of velocity at adjacent frequency points; For the weights of probability terms, For the weight of the energy term, For the weight of the smoothing term, The curvature term weights are all preset constants. By minimizing the objective function J, the globally optimal dispersion curve path is obtained.

[0016] Compared with the prior art, the present invention has the following advantages: 1. This invention transforms the acquisition of slot wave dispersion curves into a fully automated processing flow of "deep learning probability recognition + dispersion physical constraint optimization", which greatly reduces the workload of manual interaction, improves the efficiency and consistency of batch data processing, and reduces the dependence on the experience of operators.

[0017] 2. This invention adopts the AttentionU-Net architecture with attention gating. Targeting the characteristics of slot wave dispersion ridges being thin, weak in energy, and prone to discontinuity, it significantly enhances the response intensity of dispersion ridges under low signal-to-noise ratio by guiding shallow feature weighting through high-level semantics. This effectively suppresses roadway noise, spurious peaks, and background interference, and the recognition accuracy is superior to that of ordinary convolutional networks.

[0018] 3. This invention constructs a multi-objective optimization function that integrates network semantic probability, original energy, velocity smoothing and curvature constraints. It solves the global optimal path through dynamic programming, so that the picking result not only utilizes the semantic recognition capability of deep learning, but also strictly conforms to the physical law of slot wave propagation, thus solving the problem that pure network output is prone to jumps and breaks.

[0019] 4. This invention supports the picking of basic and higher-order multimodal dispersion curves. By reducing mode aliasing through iterative suppression and modal constraints, it can provide more complete dispersion information for the interpretation of geological anomalies such as coal seam thickness variation, faults, collapse columns, and coal seam bifurcation and merging, and has a wider range of applicable scenarios.

[0020] 5. The training samples of this invention are compatible with both theoretical forward modeling and field measured data. Combined with multi-dimensional data enhancement strategies, they can adapt to engineering scenarios with different coal seam conditions and different acquisition qualities, and have strong engineering application value and promotion potential. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0022] Figure 2 This is an example schematic diagram of the slot wave frequency-velocity energy diagram in this invention.

[0023] Figure 3 This is a schematic diagram of the AttentionU-Net slot dispersion recognition network in this invention.

[0024] Figure 4 This is a schematic diagram comparing the results of automatic picking with those of manual picking in this invention. Detailed Implementation

[0025] The present invention will be further described below.

[0026] like Figure 1 As shown, this embodiment performs intelligent picking of the slot wave dispersion curve according to the following steps: S1. Data Acquisition and Preprocessing: Seismic source points and receiver points are set up on both sides of the coal mine working face to acquire single-component or three-component channel wave seismic records. The raw data is processed in the following order: bad channel removal, removing seismic channels with abnormal energy or no signal; static correction, eliminating time shifts caused by differences in surface excitation and reception conditions; bandpass filtering, suppressing random noise outside the effective frequency band of the channel wave; amplitude equalization, compensating for energy differences between different channels; noise suppression, further suppressing coherent noise in the roadway; and finally, amplitude normalization of the data.

[0027] The preprocessed slot wave data is subjected to a dispersion spectrum transform using the generalized S-transform to obtain a frequency-velocity energy map, as shown below. Figure 2 As shown in the figure, the vertical axis represents frequency, ranging from 0 to 500 Hz; the horizontal axis represents phase velocity, ranging from 600 to 2200 m / s; and the pixel color depth represents the energy intensity of the corresponding frequency-velocity point. In the processing, the velocity scanning interval is set according to the velocity range of the target coal seam, and the energy map is subjected to background suppression and 0-1 normalization to unify the numerical range of the input network.

[0028] S2. AttentionU-Net Network Construction: The constructed network structure is as follows... Figure 3 As shown, the overall architecture is a U-shaped structure with 4 levels of downsampling and 4 levels of upsampling, and its specific components are as follows: The encoder contains four levels of downsampling units. Each level consists of two 3×3 convolutional layers, a batch normalization layer, a ReLU activation layer, and a 2×2 max pooling layer. It downsamples the input feature map size to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original size, respectively, to extract multi-scale features from fine-grained texture to global structure.

[0029] Bottleneck layer: Located between the encoder and decoder, it consists of two convolutional layers and is used to fuse global semantic information and model the overall shape and distribution of the dispersion curve.

[0030] Decoder: Contains 4 levels of upsampling units. Each unit is first upsampled by 2x through bilinear interpolation or transposed convolution, and then fused with the encoder features of the corresponding level through skip connections. After that, it goes through two 3×3 convolutional layers for feature extraction, and the spatial resolution is restored step by step.

[0031] Attention gating modules: There are four modules in total, each connected to a different hop path. Each module receives upsampled features from the decoder as a gating signal. Features at the same level as the encoder The features are jointly processed through convolution and sigmoid activation to generate a spatial attention weight map, which is then multiplied with the original encoder features to obtain a weighted feature. This mechanism can automatically focus on the dispersive ridge region using high-level semantic information, suppress background noise and non-target energy responses, and especially improve the recognition effect of weak energy and discontinuous ridges.

[0032] Output layer: Uses 1×1 convolution with Sigmoid activation to output a dispersion ridge probability map with the same size as the input. The pixel value ranges from 0 to 1, and the higher the value, the greater the probability that the point belongs to the dispersion ridge.

[0033] S3. Network Training: The training sample set consists of three parts: ① Simulated channel wave data and corresponding theoretical dispersion curve labels obtained by finite difference forward modeling based on coal seam geological model; ② Dispersion curve labels of field measured channel wave data annotated by professionals; The label format is a binary mask generated by expanding the center line of the curve by 3~5 pixels.

[0034] During training, data augmentation was employed to improve model generalization, specifically including: adding random Gaussian noise, randomly occluding local frequency bands, randomly attenuating local energy, adding multimodal aliasing signals, and applying small velocity offsets. The loss function used was a combination of Dice loss and binary cross-entropy loss, each with a weight of 0.5, to address the sample imbalance problem caused by the low proportion of pixels along the dispersion ridge. The optimizer was Adam, with an initial learning rate of 0.001 and 100 training epochs. Training stopped when the validation set loss did not decrease for 10 consecutive epochs.

[0035] S4. Preliminary identification of dispersion curves: After adjusting the frequency-velocity energy map to be processed to the network input size, input it into the trained AttentionU-Net model, and obtain the dispersion ridge probability map through forward inference, which serves as the preliminary identification result.

[0036] S5. Path optimization under physical constraints: In order to eliminate local breaks, isolated noise points and pseudo-peaks in the preliminary identification results, the objective function is further constructed and the optimal path is solved by dynamic programming.

[0037] The objective function in this embodiment is:

[0038] Among them, frequency range ~ Set according to the effective frequency band; weighting coefficient is taken as follows: , , , .

[0039] A dynamic programming algorithm is used to find the minimum cost path: the cumulative cost of each velocity point is calculated row by row along the frequency axis from low frequency to high frequency, and the optimal preceding position of each point is recorded; after traversing to the highest frequency, backtracking is performed from the minimum cost point to obtain the globally optimal continuous dispersion curve path. This path simultaneously satisfies multiple constraints such as high probability, high energy, smooth velocity change, and small curvature, which conforms to the physical laws of slot wave propagation.

[0040] S6. Multimodal Picking and Quality Control: For picking multi-order dispersion curves, an iterative suppression strategy is adopted: first, the most energetic basic order dispersion curve is picked according to the above method, and then the energy of the curve's neighborhood is suppressed in the original energy map and probability map. The identification and path optimization steps are repeated to obtain higher-order curves. In the process, modal spacing constraints are added to limit the minimum velocity difference between adjacent order curves, and a cross-penalty term is added to avoid curve crossing.

[0041] Quality control is performed on the final acquired dispersion curves: outliers deviating from the overall trend are removed, linear interpolation is performed on locally missing frequency bands, and smoothing is done using moving average; after manual review, the output is... The dispersion curve is displayed in the form of a frequency response curve, and quality indicators such as average confidence level and effective bandwidth are also output. The picking results are as follows: Figure 4 As shown, the results are highly consistent with those obtained manually, and good continuity is maintained even in low signal-to-noise ratio and local power loss frequency bands.

[0042] The resulting dispersion curves can be directly used for coal seam thickness inversion, coal seam velocity structure analysis, fault and collapse column identification, and coal seam bifurcation and merging interpretation, supporting the transparent detection and geological support work of coal mine working faces.

[0043] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligently picking slot wave dispersion curves based on AttentionU-Net, characterized in that, Includes the following steps: S1. Collect channel wave seismic data from the coal mine working face. After preprocessing the channel wave seismic data, generate a frequency-velocity energy map through dispersion spectrum transformation. S2. Construct the AttentionU-Net slot dispersion recognition network, which includes an encoder, a bottleneck layer, a decoder, skip connections, and an attention gating module; The encoder is used to extract multi-scale dispersion features, the bottleneck layer is used to fuse global semantic information, the decoder is used to restore spatial resolution step by step, and the attention gating module is connected to the skip connection path. Based on the high-level semantic information output by the decoder, the shallow features output by the encoder are weighted and filtered to enhance the dispersion ridge features and suppress background noise. S3. Construct a training sample set to train the AttentionU-Net slot wave dispersion recognition network to obtain the trained recognition network model; S4. Input the frequency-velocity energy map to be processed obtained in step S1 into the trained recognition network model, and output the preliminary recognition result of the slot wave dispersion curve. S5. Combine the physical constraints of slot wave dispersion to construct the path optimization objective function. Using the preliminary identification results and the original frequency-velocity energy map as input, the optimal path is solved by dynamic programming algorithm to obtain a continuous and smooth single-order slot wave dispersion curve. S6. After performing quality control on the picked-up slot wave dispersion curve, output the final dispersion curve picking result.

2. The method according to claim 1, characterized in that, In step S1, preprocessing includes bad sector removal, static correction, bandpass filtering, amplitude equalization, noise suppression, and data normalization.

3. The method according to claim 1, characterized in that, In step S1, the dispersion spectrum transformation adopts any one or more combinations of phase shift method, fk transform, short-time Fourier transform, and generalized S transform; in the generated frequency-velocity energy map, the horizontal axis is the phase velocity or group velocity, the vertical axis is the frequency, and the energy intensity at the position corresponding to the pixel gray value or color value.

4. The method according to claim 1, characterized in that, In step S2, the encoder includes multiple levels of downsampling convolutional units, each level of which includes a convolutional layer, a batch normalization layer, and a ReLU activation layer; the decoder includes multiple levels of upsampling convolutional units corresponding to the encoder levels; the attention gating module is correspondingly set in each skip connection, receives the gating signal of the corresponding level of the decoder, spatially weights the output features of the encoder at the same level, and suppresses the response of background noise and non-ridge features.

5. The method according to claim 5, characterized in that, In step S3, the training sample set includes theoretical forward modeling simulation of tank wave data, field measured tank wave data, and corresponding labeled dispersion curves; the dispersion curve labels are in the form of curve centerline, curve neighborhood binary mask, or multimodal category mask.

6. The method according to claim 1, characterized in that, During the training process in step S3, data augmentation is performed using any one or more of the following methods: noise perturbation, bandwidth loss, energy attenuation, modal mixing, velocity perturbation, and local occlusion. The loss function for network training uses any one or more combinations of binary cross-entropy loss, Dice loss, and Focal loss.

7. The method according to claim 1, characterized in that, In step S5, the expression for the path optimization objective function J is: in, To identify the dispersion ridge probability map output by the network. This is the original frequency-velocity energy diagram. The velocity change at adjacent frequency points The change in velocity curvature at adjacent frequency points. , , , These are the preset weighting coefficients.

8. The method according to claim 1, characterized in that, When it is necessary to pick up multi-order slot wave dispersion curves, an iterative ridge suppression method is adopted: first, pick up the first-order dispersion curve with the strongest energy, suppress its neighborhood energy, and then repeat steps S4 to S5 to pick up the next order curve; during the picking process, add modal spacing constraints, cross penalty terms and continuity discrimination conditions to distinguish dispersion curves of different orders.

9. The method according to claim 1, characterized in that, In step S6, quality control includes any one or more of the following: outlier removal, local interpolation, smoothing, and manual review; the final output dispersion curve is... It is represented in a formal way, and includes average probability, average energy, effective bandwidth and confidence level indicators.

10. The method according to claim 1, characterized in that, The final dispersion curve picking results are used for coal seam thickness inversion, velocity structure analysis, fault identification, collapse column identification, coal seam bifurcation and merging interpretation, or working face transparency detection.

Images