Underground hidden water filling cavity detection method based on artificial intelligence

Abstract

The invention belongs to the technical field of mine engineering detection, and provides an artificial intelligence-based underground hidden water filling cavity detection method, which comprises the following steps of: making a training set of a water filling hidden cavity detection model; establishing an underground hidden water filling cavity identification model; training an underground hidden water filling cavity identification model; collecting resistivity data fed back by the on-site ore body by using a high-density resistivity method; importing the collected resistivity data into the trained underground hidden water-filling cavity identification model, identifying the spatial position information of the water-filling cavity, delineating the abnormality of each measuring line cavity, and determining the space distribution range of the cavity plane; according to the method, the technical level of the metallurgical mine industry in the aspects of hidden cavity detection, control and high-efficiency and high-quality treatment is greatly improved; and omission-free detection and safety management and control processing are carried out on the underground hidden water filling cavity, so that safety threats of the hidden cavity to mining production personnel and equipment can be eliminated, safe and smooth completion of production tasks is guaranteed, and social benefits are remarkable.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine engineering exploration, and relates to the precise detection of water-filled cavities in mining areas. Specifically, it is a method for detecting underground hidden water-filled cavities based on artificial intelligence. Background Art

[0002] In the past more than 10 years of high-intensity mining production, the production elevation has also dropped from the highest 1710 m to the lowest 1440 m. As the production elevation drops, the groundwater level gradually approaches the surface. Since 2019, water-filled hidden cavities have successively appeared in the mining area. At present, most of the research on hidden cavities focuses on unfilled hidden cavities. However, in the case of hidden water-filled cavities, under the action of the contained groundwater, not only the geophysical properties of the hidden cavities are changed, affecting the precise identification of hidden cavity anomalies, but also the rock mechanical properties of the surrounding rock change. Especially for some special soft rocks such as schist in the area, they are hard and intact in the natural state, but expand, disintegrate and soften after encountering water, and the mechanical properties of the rock will rapidly decrease significantly, making it more difficult to predict the deformation and instability of the roof of the hidden cavity, and it is extremely easy to induce sudden instability and collapse of the roof of the hidden cavity. Compared with unfilled hidden cavities, the detection and treatment of hidden water-filled cavities are more difficult and pose a greater threat to the safety of mining production, and the detection and treatment technologies for unfilled hidden cavities are not applicable to water-filled hidden cavities. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for detecting underground hidden water-filled cavities based on artificial intelligence, which combines the semantic segmentation artificial intelligence technology of convolutional neural network and the geophysical exploration technology of high-density resistivity method to accurately and efficiently complete the detection of underground hidden water-filled cavities in the mining area.

[0004] The technical solution adopted by the present invention to achieve the above purpose is as follows: A method for detecting underground hidden water-filled cavities based on artificial intelligence, comprising: S1. Making a training set for the detection model of water-filled hidden cavities S11. Through the geological research of the ore deposit, determining the main rock types of the surrounding rock interbedded with the ore layer; S12. Measuring the resistivity characteristics of the surrounding rock and the ore layer by using the small four-pole outcrop method; S13. Collecting the spatial distribution ranges of various geological bodies in the mining area, the occurrence of the ore layer and the main surrounding rock, and determining the spatial distribution characteristic parameters of various rocks of the ore layer and the surrounding rock; S14. Combining the resistivity values, distribution ranges and occurrence characteristics of each lithology in the ore deposit, constructing an open-pit mine resistivity model, and adding water-filled cavities to the model to construct an open-pit mine water-filled hidden cavity resistivity model; S15. Use the least squares method to perform forward simulation calculations on the resistivity model of the concealed water-filled cavity in the open-pit mine, and obtain the training set for the learning and training of the intelligent recognition model.

[0005] S2. Establish an identification model for underground concealed water-filled cavities For the said underground concealed water-filled cavity identification model, use the semantic segmentation method of the convolutional neural network, adopt the DeeplabV3+ model, use the residual network as the underlying network, obtain a clear object boundary by restoring spatial information, optimize the boundary for segmentation, and realize the identification of underground concealed water-filled cavities in the mining area; Furthermore, an encoder-decoder for semantic segmentation is introduced into the Deeplabv3+ model. The encoder part of the DeepLabv3+ model is jointly composed of the deep convolutional neural network Xception containing the input stream, the intermediate stream, and the output stream and the atrous spatial pyramid pooling (ASPP) module with multiple dilation rates of atrous convolutions. Use the Xception network to extract the high-level semantic features of the input image, and then transfer them to the ASPP module of the model for multi-scale sampling to generate feature maps for fusion learning to avoid information loss. The decoder module fuses the underlying features output by the Xception input stream module and the high-level features output by the encoder, and then performs bilinear interpolation upsampling to further fuse the underlying features and the high-level features, and outputs the segmentation result, ultimately improving the accuracy of the segmentation boundary.

[0006] S3. Train the identification model for underground concealed water-filled cavities; First, pre-train the DeepLabv3+ model on the pascalvoc dataset and learn and train the inversion data of the water-filled cavity body. Assume that the features extracted from the backbone part of the pre-trained model are general, and do not train the backbone part first, that is, first freeze most of the training parameters of the deeplabv3+ model, and then unfreeze all the parameters after the pre-training to fine-tune the parameters of the backbone part. Through repeated adjustment of the parameters in the training process, obtain the deeplabv3+ model parameters that can be used to identify underground concealed water-filled cavities.

[0007] S4. Use the high-density resistivity method to collect the resistivity feedback from the on-site ore body by using the two-electrode device and the Wenner device.

[0008] Furthermore, for the detection blind area in the ore body profile detected by the high-density resistivity method, use the cross-wiring method to avoid the influence of the blind area.

[0009] Further, according to the width of the on-site ore body, in the areas where data cannot be collected in the current inverted trapezoidal areas on both sides of the ore body, a second wiring is carried out perpendicular to the direction of the first survey line at a distance of more than 3 meters from the short side of the inverted trapezoid. The survey line spacing is 5 meters, the electrode spacing is 3 meters, and at least 5 survey lines are arranged; the complete current field area formed by the second survey line arrangement completely coincides with the blind area formed by the first survey line arrangement.

[0010] S5. Import the collected resistivity data into the trained underground hidden water-filled cavity identification model to identify the spatial position information of the water-filled cavities, delineate the cavity anomalies of each survey line, and clarify the plane spatial distribution range of the cavities.

[0011] S51. Preprocess the collected high-density resistivity data through non-value rejection, denoising, and filtering. S52. Perform inversion calculation using the least squares method to obtain the geoelectric model of the survey line profile. S53. Combine the geological characteristics and physical property characteristics of the survey area to establish a geological model. S54. Import the geoelectric model and the geological model into the trained underground hidden water-filled cavity identification model for identification calculation to identify the spatial position and range information of the water-filled cavities, and delineate the cavity anomalies of each survey line. S55. According to the abnormal identification results of each survey line, compare and analyze the abnormal characteristics of each survey line. Combine the spatial positions of the abnormal characteristics of each survey line, and connect the anomalies with similar abnormal characteristics and close spatial positions to form a geophysical exploration abnormal area of the cavity, and clarify the plane spatial distribution range of the cavity.

[0012] The beneficial effects of the present invention are: It greatly improves the technical level of the metallurgical mine industry in the detection, control, and efficient and high-quality treatment of hidden water-filled cavities. Applying this method can detect and safely control and treat underground hidden water-filled cavities without omission, eliminate the safety threats of hidden water-filled cavities to mining production personnel and equipment, ensure the safe and smooth completion of production tasks, and have significant social benefits. Description of the Drawings

[0013] Figure 1 is the flow chart of the method of the present invention; Figure 2 is the resistivity model diagram of the forward hidden water-filled cavity in the present invention; Figure 3 is the schematic diagram of collecting on-site data by using the cross-overlapping wiring method in the present invention; Figure 4 is the schematic diagram of the geoelectric model of the inverted hidden water-filled cavity in the present invention; In the figure, 1 is a hidden water-filled cavity; 2 is an iron ore layer; 3 is surrounding rock. Detailed Embodiments

[0014] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: A method for detecting hidden water-filled cavities underground based on artificial intelligence, as Figure 1 shown, includes: S1. Produce a training set for the hidden water-filled cavity detection model S11. Through geological research on the ore deposit, determine the main rock types of the surrounding rock interbedded with the ore layer; S12. Use the small four-pole outcrop method to measure the resistivity characteristics of the surrounding rock and the ore layer; S13. Collect the spatial distribution ranges of various geological bodies in the mining area, the occurrence of the ore layer and the main surrounding rock, and determine the spatial distribution characteristic parameters of various rocks in the ore layer and the surrounding rock; S14. Combine the resistivity values, distribution ranges and occurrence characteristics of each lithology in the ore deposit to construct a resistivity model of the mining area, and add water-filled cavities to the model to construct a resistivity model of hidden water-filled cavities in the mining area; S15. Use the least squares method to perform forward simulation calculations on the constructed resistivity model of hidden water-filled cavities to obtain a training set for intelligent recognition model learning and training.

[0015] S2. Establish an underground hidden water-filled cavity recognition model, use the semantic segmentation method of convolutional neural network, adopt the DeeplabV3+ model, use the residual network as the underlying network, obtain clear object boundaries by restoring spatial information, optimize the boundaries for segmentation, and realize the recognition of hidden water-filled cavities underground in the mining area; among them, the encoder-decoder for semantic segmentation is introduced in the Deeplabv3+ model. The encoder part of the DeepLabv3+ model is jointly composed of the deep convolutional neural network Xception containing the input stream, the middle stream and the output stream and the multi-scale spatial pyramid pooling module (atrous spatial pyramid pooling, ASPP) containing multiple dilation rates of dilated convolutions. Use the Xception network to extract the high-level semantic features of the input image, and then transfer them to the ASPP module of the model for multi-scale sampling to generate feature maps for fusion learning to avoid information loss. The decoder module fuses the underlying features output by the Xception input stream module and the high-level features output by the encoder, and then performs bilinear interpolation upsampling to further fuse the underlying features and the high-level features, and outputs the segmentation result, finally improving the accuracy of the segmentation boundary.

[0016] S3. Train the underground hidden water-filled cavity recognition model; First, pre-train the DeepLabv3+ model on the Pascal VOC dataset and learn and train the inversion data of the water-filled cavity body. Assume that the features extracted from the backbone part of the pre-trained model are general, and do not train the backbone part first, that is, freeze most of the training parameters of the DeepLabv3+ model. After the pre-training is completed, unfreeze all the parameters to fine-tune the parameters of the backbone part. By repeatedly adjusting the parameters of the training process, obtain the DeepLabv3+ model parameters that can be used to identify underground hidden water-filled cavities.

[0017] S4. Use the high-density resistivity method to collect the resistivity feedback from the on-site ore body using the secondary device and the Wenner device. For the detection blind area in the ore body profile detected by the high-density resistivity method, use the cross-wiring method to avoid the influence of the blind area. According to the width of the on-site ore body, in the area where data cannot be collected in the current inverted trapezoidal area on both sides of the ore body, conduct a second wiring perpendicular to the first survey line direction with a distance of more than 3 meters from the short side of the inverted trapezoid. The survey line spacing is 5 meters, the electrode spacing is 3 meters, and at least 5 survey lines are arranged. The complete area of the current field formed by the second survey line arrangement completely coincides with the blind area formed by the first survey line arrangement.

[0018] S5. Import the collected resistivity data into the trained underground hidden water-filled cavity identification model to identify the spatial position information of the water-filled cavity, delineate the cavity anomalies of each survey line, and clarify the plane spatial distribution range of the cavity. S51. Preprocess the collected high-density resistivity data by removing non-values, denoising, and filtering. S52. Use the least squares method for inversion calculation to obtain the geoelectric model of the survey line profile. S53. Combine the geological characteristics and physical property characteristics of the survey area to establish a geological model. S54. Import the geoelectric model and the geological model into the trained underground hidden water-filled cavity identification model for identification calculation to identify the spatial position and range information of the water-filled cavity, and delineate the cavity anomalies of each survey line. S55. According to the anomaly identification results of each survey line, compare and analyze the anomaly characteristics of each survey line. Combine the spatial positions of the anomaly characteristics of each survey line, and connect the anomalies with similar anomaly characteristics and close spatial positions to form a geophysical exploration anomaly area of the cavity, and clarify the plane spatial distribution range of the cavity. Embodiment

[0019] Apply the above method to detect and study the water-filled cavities in the open-pit iron ore deposit in the study area. S1. Produce a training set for the detection model of concealed water-filled cavities: Through detailed geological research on iron ore deposits, determine that the main rock types interbedded with iron ore layers are meta-diabase, sericite quartz schist, iron ore bodies, and carbonaceous schist; use the small four-electrode outcrop method to measure the resistivity characteristics of the surrounding rock and iron ore layers; collect information on the spatial distribution ranges of various geological bodies in the mining area and the occurrence of iron ore layers and main surrounding rocks, and determine the spatial distribution characteristic parameters of iron ore layers, meta-diabase, carbonaceous schist, etc.; combine the resistance values, distribution ranges, and occurrence characteristics of each lithology in the iron ore deposit to construct a resistivity model of an open-pit iron ore, and add water-filled cavities to the model to construct a resistivity model of a concealed water-filled cavity in an open-pit iron ore; use the least squares method to perform forward simulation calculations on the constructed resistivity model of the iron ore deposit containing water-filled cavities, as Figure 2 shown, to obtain a set of pictures for recognition training and learning.

[0020] S2. Establish a recognition model for concealed water-filled cavities. Adopt the semantic segmentation technology of convolutional neural networks and use the DeeplabV3+ model. Use the residual network as the underlying network to obtain clear object boundaries by restoring spatial information, thereby optimizing the boundaries for segmentation. At the same time, the Deeplabv3+ network draws on the idea of transfer learning. An encoder-decoder commonly used in semantic segmentation is introduced into the model. The encoder module of the DeepLabv3+ model is jointly composed of a deep convolutional neural network Xception containing an input stream, an intermediate stream, and an output stream and a multi-scale spatial pyramid pooling module (atrous spatial pyramid pooling, ASPP) containing multiple dilation rates of atrous convolutions. The encoding end uses the Xception network to extract high-level semantic features of the input image, and then transfers them to the ASPP module for multi-scale sampling to generate feature maps for fusion learning to avoid information loss. The decoding end fuses the underlying features output by the Xception input stream module and the high-level features output by the encoding end and then performs bilinear interpolation upsampling, that is, it further fuses the underlying features and high-level features to output the segmentation result, ultimately improving the accuracy of the segmentation boundary.

[0021] S3. Use the training set to train the recognition model of concealed water-filled cavities. First, pre-train the Deeplabv3+ model on the pascal voc dataset. When learning and training the inversion data of water-filled cavities, assume that the features extracted from the backbone part of the pre-trained model are general, and do not train the backbone part first, that is, first freeze the first 358 layers of the Deeplabv3+ model, so that the number of training parameters is reduced from the original 41,253,330 parameters to 3,183,218, a reduction of 92%. After the first step of training is completed, unfreeze and train all parameters to fine-tune the parameters of the backbone part. Through repeated adjustment of the parameters in the training process, finally obtain the model parameters of Deeplabv3+ that can be applied to the intelligent recognition of water-filled cavities.

[0022] S4. Apply the high-density resistivity method, and use the dipole-dipole array and Wenner array to collect the resistivity feedback from the on-site ore body; aiming at the technical problem of detection blind areas in the ore body profile detected by the high-density resistivity method, study and apply the cross-wiring method to avoid the influence of blind areas; as Figure 3 shown, according to the width of the on-site ore body, in the area where data cannot be collected in the current inverted trapezoidal area on both sides of the ore body, conduct a second wiring perpendicular to the direction of the first survey line by more than 3 meters beyond the short side of the inverted trapezoid, with a survey line spacing of 5 meters, an electrode spacing of 3 meters, and at least 5 survey lines are arranged; the complete current field area formed by the second survey line arrangement completely coincides with the blind area formed by the first survey line arrangement. Through this wiring method of cross-wiring and overlapping wiring, the blind area is reduced and the effective range of geophysical exploration is increased.

[0023] S5. After preprocessing the collected high-density resistivity data by non-value rejection, denoising, filtering, etc., use the least squares method for inversion calculation, as Figure 4 shown, to obtain the geoelectric model of the survey line profile. Then, combined with the geological characteristics and physical property characteristics of the survey area, establish a geological model; import it into the trained concealed water-filled cavity identification model for identification calculation to identify the spatial information such as the spatial position and scope of the water-filled cavity body, and delineate the cavity body anomalies of each survey line; according to the anomaly identification results of each survey line, compare and analyze the anomaly characteristics of each survey line, and combine the spatial positions of the anomaly characteristics of each survey line to connect the anomalies with similar anomaly characteristics and close spatial positions to form a geophysical exploration anomaly area of the cavity body, and clarify the plane spatial distribution range of the cavity body.

[0024] The method of the present invention combines the semantic segmentation artificial intelligence technology of convolutional neural network and the geophysical exploration technology of high-density resistivity method. On the basis of analyzing various geological factors affecting the high-density resistivity imaging method, a resistivity model of water-filled cavities conforming to the actual geological conditions of the iron ore deposit in the study area is constructed, and the semantic segmentation technology of convolutional neural network is first applied to the intelligent identification of water-filled cavities in the iron ore deposit in the study area, realizing the automatic, intelligent, and high-precision identification of detection anomalies of water-filled cavities in the study area, and establishing an artificial intelligence precise identification and spatial three-dimensional restoration technology for concealed water-filled cavity anomalies. Since the promotion and application of this technology in Yuanjiacun Iron Mine, 32 concealed cavities and 11 water-filled concealed cavities have been discovered by using the high-density resistivity method.

Claims

1. A method for detecting underground hidden water-filled cavities based on artificial intelligence, characterized in that: include: S1. Create a training set for the water-filled hidden cavity detection model; S2. Establish a model for identifying underground hidden water-filled cavities; S3, training a model for identifying underground hidden water-filled cavities; S4. Use high-density resistivity method to collect resistivity data fed back from the on-site ore body; S5. Import the collected resistivity data into the trained underground hidden water-filled cavity identification model to identify the spatial location information of the water-filled cavities, circle the cavity anomalies of each survey line, and clarify the spatial distribution range of the cavity plane.

2. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 1, characterized in that: The specific steps of S1 include: S11. Determine the main rock types of the surrounding rocks interbedded with the ore layers through geological research on the ore deposit; S12. Use the small quadrupole outcrop method to measure the resistivity characteristics of surrounding rocks and ore layers; S13. Collect the spatial distribution range of geological bodies in the mining area, the occurrence of ore layers and main surrounding rocks, and determine the spatial distribution characteristic parameters of various types of rocks in the ore layers and surrounding rocks; S14. Based on the resistivity values, distribution range and occurrence characteristics of each lithology of the ore deposit, an open-pit mine resistivity model is constructed, and water-filled cavities are added to the model to construct a resistivity model of water-filled hidden cavities in the open-pit mine; S15. Use the least square method to perform forward simulation calculations on the constructed open-pit mine water-filled hidden cavity resistivity model to obtain a training set for intelligent recognition model learning and training.

3. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 1 is characterized in that: The underground hidden water-filled cavity identification model uses a convolutional neural network semantic segmentation method, adopts the DeeplabV3+ model, and uses a residual network as the underlying network. It obtains clear object boundaries by restoring spatial information, optimizes the boundaries for segmentation, and realizes the identification of underground hidden water-filled cavities in mining areas.

4. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 3 is characterized in that: The Deeplabv3+ model introduces an encoder-decoder for semantic segmentation. The encoder part of the DeepLabv3+ model is composed of a deep convolutional neural network Xception containing an input stream, an intermediate stream, and an output stream, and a multi-scale spatial pyramid pooling module containing multiple hole convolution expansion rates. The Xception network is used to extract high-level semantic features of the input image, which are then transferred to the ASPP module of the model for multi-scale sampling to generate feature maps for fusion learning. The decoder module fuses the underlying features output by the Xception input stream module and the high-level features output by the encoder, and then performs bilinear interpolation upsampling to further fuse the underlying features with the high-level features and output the segmentation result.

5. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 1 is characterized by: The specific contents of the S3 training of the underground hidden water-filled cavity identification model include: Firstly, the DeepLabv3+ model is pre-trained on the PascalVOC dataset, and the inversion data of water-filled cavities is trained. Assuming that the features extracted from the backbone of the pre-trained model are universal, the backbone is not trained first, that is, most of the training parameters of the DeepLabv3+ model are frozen first, and then all the parameters are unfrozen after the pre-training is completed to fine-tune the parameters of the backbone. By repeatedly adjusting the parameters of the training process, the DeepLabv3+ model parameters that can be used to identify hidden underground water-filled cavities are obtained.

6. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 1 is characterized by: The detection blind area in the ore body profile detected by the high-density resistivity method uses a cross wiring method to avoid the influence of the blind area.

7. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 6 is characterized by: The cross wiring method is based on the width of the on-site ore body. In the areas where data cannot be collected in the current inverted trapezoidal areas on both sides of the ore body, a second wiring is performed perpendicular to the direction of the first survey line with a length of 3 meters exceeding the short side of the inverted trapezoid. The survey line spacing is 5 meters, the electrode spacing is 3 meters, and at least 5 survey lines are arranged; the complete area of ​​the current field formed by the second survey line arrangement completely overlaps with the blind area formed by the first survey line arrangement.

8. The method for detecting underground hidden water-filled cavities based on artificial intelligence according to claim 1 is characterized by: The specific contents of S5 include: S51, preprocessing the collected high-density resistivity data by non-value elimination, denoising and filtering; S52, using the least square method to perform inversion calculation to obtain a geoelectric model of the survey line profile; S53. Establish a geological model based on the geological characteristics and physical properties of the survey area; S54, the geoelectric model and the geological model are imported into the trained underground hidden water-filled cavity identification model for identification calculation, the spatial position and range information of the water-filled cavity is identified, and the cavity anomalies of each survey line are delineated; S55. According to the anomaly identification results of each survey line, the anomaly characteristics of each survey line are compared and analyzed. Combined with the spatial position of the anomaly characteristics of each survey line, the anomalies with similar anomaly characteristics and close spatial positions are connected to form a cavity geophysical anomaly area and clarify the spatial distribution range of the cavity plane.

Images