Opencast coal mine shallow goaf electrical method anomaly automatic identification method based on U-Net

Abstract

The invention provides an open pit coal mine shallow goaf electrical method anomaly automatic identification method based on U-Net, and belongs to the technical field of geophysical exploration and artificial intelligence crossing. According to the multi-scale feature fusion network, ResNet34 is combined with cavity convolution (the expansion rate is 1, 2, 4) to enlarge a receptive field, and transverse and vertical features of a goaf are captured; an adaptive filtering and channel attention mechanism (CBAM) is introduced to suppress electromagnetic interference and earth surface noise; the resistivity gradient consistency loss is embedded in the loss function, it is ensured that the segmentation result conforms to the geophysical law, experiments show that the processing time of a single cross-section diagram is smaller than or equal to 0.5 s, the IoU reaches 0.83, the recall rate is 87%, the noise immunity is remarkable, the efficiency is improved by 95% compared with manual interpretation, and the false alarm rate is smaller than or equal to 5%. The method can be widely applied to open pit coal mine goaf exploration, disaster assessment and treatment engineering, and the labor cost and the misjudgment risk are remarkably reduced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geophysical exploration and artificial intelligence, and particularly relates to a U-Net-based automatic identification method for electrical anomalies of shallow goaf in open-pit coal mine. BACKGROUND

[0002] The shallow goaf (depth ≤ 50m) in open-pit coal mine is mainly formed by disordered mining of historical small mines, and its internal includes cavities, partially filled, completely filled and other filling states, which directly affect the stability of the ground. High-density electrical method can obtain resistivity cross-section by inversion, which can represent the low-resistance characteristics of goaf, but the traditional interpretation method has the following problems:

[0003] Low efficiency: manual interpretation of a single cross-section takes more than 30 minutes, which is difficult to meet the needs of large-scale exploration;

[0004] Strong subjectivity: complex noise caused by electromagnetic interference and surface unevenness easily leads to misjudgment of low-resistance anomaly boundary, and the false positive rate is as high as 15%;

[0005] Multiple solution interference: the resistivity characteristics of goaf, water-bearing fissures and loose sediments overlap, that is, they all show low resistance, which is difficult to distinguish only by manual experience. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a U-Net-based automatic identification method for electrical anomalies of shallow goaf in open-pit coal mine, which is used for quickly identifying the spatial distribution and scale of shallow goaf in open-pit coal mine, and is suitable for mine safety production, geological disaster assessment and ecological restoration engineering.

[0007] The technical scheme adopted by the present application is: a U-Net-based automatic identification method for electrical anomalies of shallow goaf in open-pit coal mine, and the technical key points are as follows:

[0008] Obtain the high-density electrical resistivity cross-section of the shallow goaf in open-pit coal mine as the initial sample image, and perform data enhancement processing on the initial sample image to expand the number of sample images, the data enhancement processing includes introducing Gaussian noise and median noise to simulate electromagnetic interference and surface unevenness, and performing random rotation, mirror flipping and random occlusion, so as to enhance the anti-interference ability and engineering applicability of the model in complex topography of open-pit coal mine;

[0009] The improved U-Net model is constructed, ResNet34 is used as the backbone network, and a multi-scale dilated convolution layer is embedded to extract and fuse local detail features, boundary features and macro geological features in parallel to obtain fused multi-scale features; meanwhile, a space-channel attention module (CBAM) is integrated, the channel attention mechanism is used to highlight the resistivity anomaly area, and the spatial attention mechanism is used to focus on the boundary area of the goaf, so as to realize all-around capture of the horizontal and vertical features of the goaf ；

[0010] The improved U-Net model is trained by combining the fused multi-scale features with the corresponding real segmentation mask, the joint loss function is used for training, the joint loss function includes Tversky Loss and resistivity gradient consistency loss, the asymmetric penalty mechanism of light penalty for missing detection and heavy penalty for false alarm is formed by setting the parameters of Tversky Loss, and the predicted segmentation boundary is constrained to comply with the geophysical law through the resistivity gradient consistency loss, so that the model intersection over union (IoU) reaches more than 0.83, the recall rate reaches 87%, and the false alarm rate is less than or equal to 5%;

[0011] The resistivity section to be identified is input into the trained model, and a probability map is output and thresholded to obtain a binary segmentation mask;

[0012] The binary segmentation mask is post-processed and multi-source verified, a confidence heat map of the probability map is output, and a low confidence area is labeled for manual review.

[0013] In the above scheme, the data enhancement processing includes: logarithmic transformation and standardization of resistivity values to the [0, 1] interval to eliminate the magnitude difference of different survey line data; introducing Gaussian noise and median noise; random rotation and mirror flipping; implementing random occlusion, occlusion ratio ≤20%, simulating the interference of surface vegetation coverage or terrain undulation.

[0014] In the above scheme, the multi-scale dilated convolution layer is set to three different expansion rates:

[0015] Expansion rate 1, receptive field of 3x3 pixels, used to capture the development of microstructure of strata and the difference between strata resistivity and surrounding rock resistivity caused by three-zone fracture of goaf;

[0016] Expansion rate 2, receptive field of 5x5 pixels, used to capture one-way boundary features and thin-layer goaf features of goaf;

[0017] Expansion rate 4, receptive field of 15x15 pixels, used to capture the features of horizontal distribution of goaf and vertical extension of steep goaf or goaf collapse area;

[0018] In the above scheme, when training the improved U-Net model, the Tversky Loss parameters in the joint loss function are set to false positive weight coefficient α = 0.7 and false negative weight coefficient β = 0.3 to alleviate the imbalance between foreground and background pixels; the weight coefficient of the resistivity gradient consistency loss is set to λ = 0.3, and its calculation formula is as follows:

[0019]

[0020] Where N is the number of samples. Let be the predicted resistivity gradient of the i-th sample. Let be the true resistivity gradient of the i-th sample.

[0021] In the above scheme, when training the improved U-Net model, a preset batch size is used for multiple rounds of iterative training. The learning rate is dynamically adjusted through a cosine annealing scheduling strategy, and an L2 regularization term is added to prevent overfitting. The training process terminates when the performance index of the validation set reaches a stable state.

[0022] In the above scheme, the post-processing includes closing operations to fill small holes in the segmented mask to improve the continuity inside the goaf, and corrosion operations to smooth the boundary to eliminate isolated noise points; multi-source verification uses point cloud data obtained from borehole core records and three-dimensional laser scanning to calculate the intersection-union ratio, recall rate and false alarm rate.

[0023] In the above scheme, the spatial-channel attention module works in concert with the channel attention submodule and the spatial attention submodule. The channel attention submodule generates channel weights through global average pooling and global max pooling to highlight areas with abnormal resistivity; the spatial attention submodule generates a spatial weight map through convolutional layers to focus on the boundary area of ​​the goaf.

[0024] The above scheme also includes a two-stage processing flow after the encoder output, which sequentially passes through channel attention filtering and adaptive spatial filtering to improve anti-interference capability. The two-stage noise processing flow is as follows:

[0025] The channel attention filtering module performs feature purification on the multi-scale feature map output by the encoder, generates channel weights through global average pooling and global max pooling, and normalizes them using the Sigmoid function.

[0026] The adaptive spatial filtering module dynamically selects Gaussian filtering parameters based on the local signal-to-noise ratio (SNR): when the SNR ≥ 10, a Gaussian filter with a standard deviation σ of 0.2-0.8 is used; when the SNR < 10, a Gaussian filter with a standard deviation σ of 1.2-1.7 is used, thereby effectively suppressing electromagnetic interference and surface noise.

[0027] The beneficial effects of this invention are as follows: This U-Net-based automatic identification method for electrical anomalies in shallow goaf areas of open-pit coal mines significantly enhances the model's ability to capture goaf boundaries and internal structures through the collaborative work of multi-scale void convolution and spatial-channel attention mechanisms; the two-stage noise processing effectively suppresses electromagnetic interference and surface inhomogeneity noise, maintaining a segmentation accuracy decrease of no more than 5% even in complex environments with a signal-to-noise ratio ≤10dB, demonstrating strong anti-interference capabilities; the introduction of resistivity gradient consistency loss constraints ensures that the segmentation results conform to geophysical laws, improving the consistency rate with borehole verification, supporting the interpretation of resistivity profile data from various devices under complex terrains in open-pit coal mines, adapting to various goaf filling types, and suitable for various terrains (hills, plains, etc.), making it widely applicable in engineering. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the automatic identification method for electrical anomalies in shallow goaf areas of open-pit coal mines based on U-Net, as described in this embodiment of the invention.

[0030] Figure 2 This is an improved U-Net model architecture diagram (including dilated convolution and CBAM modules) in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram illustrating data preprocessing and enhancement in an embodiment of the present invention;

[0032] Figure 4 This is a cross-sectional view of the original detection data in an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram illustrating the results of the conventional method in the embodiments of the present invention;

[0034] Figure 6 This is a schematic diagram of the U-Net automatic segmentation and recognition results in an embodiment of the present invention. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with the accompanying drawings. Figures 1-6 The present invention will be further described in detail below with reference to specific embodiments.

[0036] This embodiment employs an automatic electrical anomaly identification method based on U-Net for shallow goaf areas in open-pit coal mines, comprising the following steps:

[0037] S1. Steps for data preparation and enhancement.

[0038] Input data: Two-dimensional resistivity profiles obtained by inversion using a high-density electrical resistivity Wenner α and β device, targeting typical scenarios such as voids, partial filling, and complete filling in goaf areas.

[0039] Tag data: Based on borehole verification and 3D laser scanning results inside the borehole, the goaf mask was manually labeled as tag data;

[0040] The above data undergoes preprocessing, including normalization, noise addition, geometric transformation, and random occlusion. The specific process is as follows:

[0041] Normalization: The resistivity values ​​are logarithmically transformed to resolve dimensional differences, and then standardized to the [0, 1] interval to eliminate magnitude differences between different measurement lines. The aim is to provide stable and standardized input for subsequent model training by standardizing and enhancing the original resistivity data.

[0042] Noise addition: Gaussian noise with a mean of 0 and a variance of 0.01 is introduced to simulate electromagnetic interference, and median noise with a probability of 5% is introduced to simulate surface inhomogeneity interference.

[0043] Geometric transformations: random rotation (±15°) and mirror flipping enhance the model's adaptability to different directional distributions of goaf areas.

[0044] By introducing noise and geometric changes, the generalization ability and robustness of the model are improved in complex noise environments and under different goaf distribution directions, which supports the high accuracy and strong anti-interference of this embodiment.

[0045] Random occlusion: Simulates surface disturbances (such as vegetation cover, terrain undulation), with an occlusion rate of ≤20%, to avoid overfitting the model to local features. The purpose is to ensure that when applied to real-world scenarios, even with incomplete or partially disturbed data, relatively robust judgments can be made.

[0046] S2. Improved U-Net model construction:

[0047] S2.1. Steps for achieving multi-scale feature fusion through dilated convolution.

[0048] The encoder in this embodiment uses a ResNet34 pre-trained model. The pre-trained model is trained using a forward model established by high-density electrical resistivity data verified from shallow goaf areas of open-pit coal mines in the same region, so that it has the ability to identify the resistivity characteristics of shallow goaf areas of open-pit coal mines in the detection area.

[0049] This embodiment introduces dilated convolution based on the ResNet34 pre-trained model. Dilated convolution expands the receptive field by inserting holes in the convolution kernel without increasing the number of parameters or computational complexity. Setting the dilation rate of the dilated convolution to 1, 2, and 4 respectively expands the receptive field to 3×3, 5×5, and 15×15 pixels.

[0050] The dilation rate 1 is a conventional convolution with a receptive field of view of 3×3 pixels. It is used to extract local details, boundary features, and lateral and vertical distribution characteristics of the goaf, such as the microstructural development of the strata, and the difference in resistivity between the strata caused by the fracturing of the three zones in the goaf and the surrounding rock. Among these, the microstructural development of the strata refers to the collapse of the roof or the bulging of the floor in the goaf, and the three zones in the goaf refer to the caving zone, the fracture zone, and the tortuous zone. It identifies small, discontinuous low-resistivity or high-resistivity anomaly stripes or patches. On the resistivity map, these features reflect that the surrounding rock fractures caused by mining have been filled with water or formed voids.

[0051] The expansion ratio 2 expands the field of view to 5×5 pixels to capture geological features ranging from centimeters to meters, such as the unidirectional boundary features of goaf areas and thin-layer goaf features. By acquiring pixels or edges where resistivity values ​​change drastically as abrupt change points, the transition area between the goaf area and the intact surrounding rock can be determined. Its purpose is to determine whether there is an anomaly in a small area of ​​the image, as well as the basic morphology and electrical characteristics of the anomaly.

[0052] Expansion Rate 4: The field of view is expanded to 15×15 pixels, enabling the capture of geological features of goaf areas at the 10-meter to 100-meter level. This includes features used to capture the lateral distribution of goaf areas and the vertical extension of steeply dipping goaf areas or goaf collapse areas. Specifically, by observing the overall outline of low-resistivity anomaly areas, it can determine whether their lateral extension exceeds 20 meters, whether they are isolated or connected, and thus determine the horizontal extension range, overall shape, and spatial arrangement of goaf areas. By observing the depth at which low-resistivity anomaly areas begin to appear and end, it can determine whether it is a single-layer or multi-layer goaf, and thus determine the vertical distribution layers of goaf areas, whether they are within the shallow layer ≤50m, and the vertical extension height.

[0053] Feature maps are extracted in parallel using three dilated convolutional layers with different dilation rates. These multi-scale feature maps are then fused, and the fused multi-scale features are passed to the decoder via skip connections for fusion with upsampled features. This approach restores spatial resolution while maintaining the integrity of multi-scale features, ensuring accurate localization of goaf boundaries and accurate identification of internal structures. The collaborative work of multi-scale dilated convolutions effectively expands the model's receptive field, achieving comprehensive capture of the lateral and vertical extension features of shallow goaf areas in open-pit coal mines, significantly improving the accuracy and robustness of goaf identification.

[0054] S2.2 Steps for optimizing feature weights using the over-attention mechanism.

[0055] The decoder introduces a Spatial-Channel Attention Module (CBAM), which comprises two sub-modules: a channel attention mechanism and a spatial attention mechanism. The channel attention mechanism captures the overall feature distribution through global average pooling and extracts significant resistivity features for each channel using global max pooling to form a feature map. Based on the results of global average pooling and global max pooling, weights for each feature are calculated through fully connected layers or convolutional layers. During this process, a ReLU activation function can be used to introduce a non-linear transformation to enhance representational power, generating weights for the S2.1 feature map. These weights are used to adjust the importance of each feature map, thereby making the model more focused on resistivity anomaly regions related to goaf areas.

[0056] The spatial attention mechanism in this embodiment is as follows: based on channel attention, the spatial distribution of the feature map is further refined. The spatial attention mechanism calculates a weight for each position in the S2.1 feature map, that is, it calculates spatial weights, highlights important spatial regions, and smooths the boundaries, thereby making the boundaries more accurate.

[0057] By focusing on resistivity anomalies using channel attention weights, spatial attention mechanisms can refine boundaries.

[0058] Using the attention-weighted features described above, a probability map is generated in the model's output layer using the Sigmoid activation function. This probability map reflects the probability that each pixel is a mined-out area. After thresholding (the default threshold is 0.5), a binary segmentation mask is output to determine whether the pixel belongs to a mined-out area or not.

[0059] S2.3. Steps for suppressing noise interference through two-stage filtering.

[0060] To improve the model's stability in noisy environments, this embodiment introduces an independent channel attention filter module at the encoder backend, which works in conjunction with adaptive filtering to form a two-stage noise processing flow. This flow cleanses the multi-scale feature map output in step S2.1, aiming to suppress noise-related interference patterns in the feature map while enhancing useful features related to goaf areas.

[0061] This embodiment introduces an independent channel attention filter module at the encoder's back end, which works in conjunction with the adaptive filter to form a two-stage noise processing flow. The processing chain of the noise robustness enhancement design is as follows: the encoder output is sent to the channel attention filter module, the channel attention filter module then outputs to the adaptive spatial filter module, and finally the adaptive spatial filter module outputs to the decoder.

[0062] The channel attention filtering module is located after the encoder. The multi-scale feature maps extracted by the encoder are first sent to the channel attention filtering module for feature purification, and then passed to the decoder through skip connections for upsampling and fusion.

[0063] First, channel attention filtering is performed: for each channel of the input feature map, global average pooling and global max pooling are performed to obtain two global descriptor vectors. These two pooling vectors are then input into a shared small weight generation network. This network consists of two fully connected layers sequentially: the first fully connected layer is followed by a ReLU activation function for non-linear transformation, while the second fully connected layer does not use an activation function. This network shares parameters across all channels, and its output is the initial weight for each channel. The specific workflow of this shared small weight generation network is as follows: global average pooling and global max pooling are performed on each feature map to obtain two different global descriptor vectors; these two vectors are then input into the same small network consisting of two fully connected layers, where the first fully connected layer uses the ReLU activation function; the weight vectors output by the two networks are element-wise summed; finally, the sum is normalized to the [0,1] interval using the Sigmoid function to obtain the final weight for each channel. The normalized channel weights are then multiplied channel-by-channel by the corresponding channel of the original feature map.

[0064] Subsequently, adaptive spatial filtering is performed: the standard deviation or variance of pixel values ​​is calculated within a local window on the feature map after channel attention filtering to estimate the noise intensity of the region.

[0065] Based on the estimated local noise intensity, the filtering method is dynamically selected: when the signal-to-noise ratio (SNR) is ≥10, a Gaussian filter with a standard deviation σ satisfying 0.2≤σ≤0.8 is used; when the SNR is <10, a Gaussian filter with a standard deviation σ satisfying 1.2≤σ≤1.7 is used, or a median filter is used when the noise interference probability is ≥5%.

[0066] The feature map, after spatial filtering, is passed to the decoder via skip connections. This completes the creation of the improved U-Net model.

[0067] S3. After completing the construction of the improved U-Net model, the model training phase begins. This embodiment uses training data to train the model, and the specific steps are as follows:

[0068] S3.1 Loss function for training optimization.

[0069] This embodiment uses a joint loss function to train the model. The joint loss function L... total Tversky Loss and resistivity gradient uniformity loss L grad The weighted composition is calculated using the following formula:

[0070] L total =L Tversky +λ·L grad

[0071] Where λ is the weighting coefficient for the resistivity gradient consistency loss, set to λ = 0.3

[0072] This embodiment utilizes Tversky Loss to address the severe imbalance in the number of foreground and background pixels during goaf area identification. The formula is as follows:

[0073] The formula for calculating Tversky Loss is:

[0074]

[0075] Here, ε is a very small positive number used to ensure numerical stability and avoid zero in the denominator. When both the numerator and denominator are very small, ε can prevent instability or overflow issues in gradient calculation. α is the penalty weight for missed detections; as α increases, the model focuses more on reducing missed detections. β is the penalty weight for false positives; as β increases, the model focuses more on reducing false positives. Because missed detections in open-pit coal mine goaf areas can lead to serious safety hazards, the parameters are set to α = 0.7 and β = 0.3, increasing the weight of the foreground and emphasizing the importance of foreground pixels, making the model more focused on reducing missed detections in goaf areas.

[0076] In this embodiment, the resistivity gradient consistency loss is used to constrain the segmentation results of the model, ensuring that the predicted goaf boundary remains consistent with the actual resistivity gradient abrupt change region. This loss term is calculated using the mean squared error, and its formula is as follows:

[0077]

[0078] Where N is the number of samples. Let be the predicted resistivity gradient of the i-th sample. Let be the true resistivity gradient of the i-th sample.

[0079] S3.2 Training parameter settings.

[0080] In this embodiment, the weighting coefficient λ = 0.3 for the resistivity gradient consistency loss is used to constrain the prediction segmentation boundary to be consistent with the resistivity gradient abrupt change region.

[0081] The model was trained using a batch size of 16 and 150 epochs. Training was terminated when the Tversky index on the validation set stabilized at 0.86.

[0082] Cosine annealing scheduling with an initial learning rate of 1e-4 is used, and L2 regularization (λ = 1e-5) is added to prevent overfitting. The batch size is 16.

[0083] S4. Post-processing and verification:

[0084] Perform a closing operation on the segmentation results (kernel size 5×5): fill the small holes in the segmentation mask and improve the continuity inside the goaf.

[0085] Etching operation (core size 3×3): Fill the pores and smooth the boundary burrs, eliminating isolated noise points;

[0086] Multi-source validation: The segmentation accuracy is validated using multi-source data obtained from borehole core records and 3D laser scanning models. Interaction rate (IoU), recall, and false alarm rate (False Alarm Rate) are calculated. The accuracy metrics for multi-source validation are IoU ≥ 0.80, recall ≥ 85%, and false alarm rate ≤ 5%.

[0087] Uncertainty analysis: Output a confidence heatmap of the probability plot, and mark low confidence areas (such as areas with ambiguous boundaries) for manual review.

[0088] The model training and performance verification process in this embodiment is as follows:

[0089] High-density electrical resistivity data (electrode spacing 2m, measurement line length 200m, combined measurement with Wenner α and β devices) from an open-pit coal mine area were used to generate 60 resistivity profile maps, and the boundaries of the goaf were manually marked.

[0090] After data augmentation, the dataset was expanded to 300 images, and the training, validation, and test sets were divided in a 7:2:1 ratio.

[0091] The improved U-Net model was trained and iterated for 150 epochs, and the Tversky index on the validation set remained stable at 0.86.

[0092] The test set segmentation result IoU = 0.82, with a borehole matching rate of 89% and a false alarm rate of ≤5%.

[0093] Inputting a three-dimensional resistivity data volume of a mining area, U-Net automatically segments shallow goaf areas (depth ≤ 50m) and outputs probability maps and confidence heat maps; combined with geological modeling software, it generates a three-dimensional distribution model of the goaf area to guide the design of grouting treatment engineering schemes; compared with traditional manual interpretation, the construction period is shortened by 70% and the treatment cost is reduced by 20%.

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A U-Net-based automatic identification method for open-pit coal mine shallow goaf electrical anomalies, characterized in that, The method comprises the following steps: obtaining a high-density electrical resistivity section of a shallow goaf of an open-pit coal mine as an initial sample image, and performing data enhancement processing on the initial sample image to expand the number of sample images, wherein the data enhancement processing comprises introducing Gaussian noise and median noise to simulate electromagnetic interference and surface unevenness, and performing random rotation, mirror flipping and random occlusion, so as to enhance the anti-interference ability and engineering applicability of the model under complex topography of the open-pit coal mine; The improved U-Net model is constructed, ResNet34 is used as a backbone network, and a multi-scale hollow convolution layer is embedded to extract and fuse local detail features, boundary features and macro geological features in parallel, so that multi-scale features are obtained; meanwhile, a space-channel attention module is integrated, the channel attention mechanism is used to highlight the resistivity anomaly area, the space attention mechanism is used to focus on the boundary area of the mined-out area, so that the all-around capture of the horizontal and vertical features of the mined-out area is realized ； the improved U-Net model is trained by combining the fused multi-scale features and the corresponding true segmentation mask, a joint loss function is used for training, the joint loss function comprises Tversky Loss and resistivity gradient consistency loss, an asymmetric penalty mechanism of light penalty for missed detection and heavy penalty for false alarm is formed by setting parameters of the Tversky Loss, and the resistivity gradient consistency loss is used to constrain the predicted segmentation boundary to comply with the geophysical law, so that the model intersection over union (IoU) is greater than 0.83, the recall rate is 87%, and the false alarm rate is less than or equal to 5%; the resistivity section to be identified is input into the trained model, a probability map is output, and a binary segmentation mask is obtained by thresholding; the binary segmentation mask is post-processed and verified by multiple sources, a confidence heat map of the probability map is output, and a low-confidence area is labeled for manual review.

2. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, The data enhancement processing comprises: performing logarithmic transformation on the resistivity value and standardizing it to the [0, 1] interval to eliminate the magnitude difference of different survey line data; introducing Gaussian noise and median noise; performing random rotation and mirror flipping; and implementing random occlusion with an occlusion ratio of less than or equal to 20% to simulate surface vegetation coverage or terrain fluctuation interference.

3. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, The multi-scale hollow convolution layer is provided with three different expansion rates: an expansion rate 1 with a receptive field of 3*3 pixels, which is used to capture the development of microstructure of the stratum and the difference between the stratum resistivity in the goaf and the surrounding stratum resistivity caused by the three-zone fracture of the goaf; an expansion rate 2 with a receptive field of 5*5 pixels, which is used to capture the one-way boundary feature of the goaf and the thin-layer goaf feature; an expansion rate 4 with a receptive field of 15*15 pixels, which is used to capture the features of the transverse distribution of the goaf and the vertical extension of the steep goaf or goaf collapse area.

4. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, When the improved U-Net model is trained, the Tversky Loss parameter in the joint loss function is set to a false positive weight coefficient α=0.7 and a false negative weight coefficient β=0.3 to alleviate the imbalance between foreground and background pixels; and the weight coefficient of the resistivity gradient consistency loss is set to λ=0.3, and the calculation formula is: where N is the number of samples, is the predicted resistivity gradient for the i-th sample, is the true resistivity gradient for the i-th sample.

5. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, When the improved U-Net model is trained, a preset batch size is used for multiple rounds of iterative training, the learning rate is dynamically adjusted by using a cosine annealing scheduling strategy, and an L2 regularization term is added to prevent overfitting, and the training process is terminated when the performance indicators of the verification set reach a stable state.

6. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, The post-processing comprises a closing operation to fill small holes in the segmentation mask to improve the continuity of the goaf, and an erosion operation to smooth the boundary to eliminate isolated noise points; the multiple-source verification uses the drilling core records and point cloud data obtained by three-dimensional laser scanning to calculate the intersection over union, the recall rate and the false alarm rate.

7. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, The space-channel attention module highlights the resistivity anomaly area by generating channel weights through global average pooling and global maximum pooling through the synergistic effect of the channel attention submodule and the spatial attention submodule; the spatial attention submodule generates a spatial weight map through a convolution layer to focus on the boundary area of the mined-out area.

8. The U-Net-based automatic identification method of electrical anomalies of shallow goaf in surface coal mine according to claim 1, characterized in that, The application also comprises a double-stage processing procedure to improve the anti-interference ability by sequentially passing through the channel attention filtering and the adaptive spatial filtering after the encoder output. The channel attention filtering module performs feature purification on the multi-scale feature map output by the encoder, generates channel weights through global average pooling and global maximum pooling, and normalizes the channel weights using a Sigmoid function; The adaptive spatial filtering module dynamically selects Gaussian filtering parameters according to the local signal-to-noise ratio: when the signal-to-noise ratio SNR is greater than or equal to 10, a Gaussian filter with a standard deviation sigma of 0.2-0.8 is used; when the signal-to-noise ratio SNR is less than 10, a Gaussian filter with a standard deviation sigma of 1.2-1.7 is used, so as to effectively suppress electromagnetic interference and surface noise.

Images