Tunnel advanced detection intelligent quantitative geological interpretation method and device based on seismic exploration while digging

Abstract

The invention provides a roadway advanced detection intelligent quantitative geological interpretation method and device based on seismic exploration while excavation, and belongs to the field of mine intelligent excavation safety geological guarantee. Under the same geological abnormal condition, different observation systems detect to obtain a reference sample and an associated sample, a two-tuple composed of the reference sample and the associated sample is input into the tunnel advanced detection intelligent quantitative geological interpretation model, the model is trained based on a reverse recognition method, multi-path recognition loss is calculated, and when the loss value is minimum, the tunnel advanced detection intelligent quantitative geological interpretation model is obtained. A trained tunnel advanced detection intelligent quantitative geological interpretation model is obtained; inputting a to-be-explained single-round roadway advanced detection imaging result image into the trained roadway advanced detection intelligent quantitative geological interpretation model for quantitative geological interpretation; the network is urged to focus on the geological abnormal imaging reflection zone part to extract features, and the advanced detection geological interpretation precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of geological safety assurance for intelligent mine excavation, and in particular to an intelligent quantitative geological interpretation method and device for tunnel advance detection based on seismic exploration while excavating. Background Art

[0002] The International Energy Agency released a report predicting that by 2040, global coal demand will increase by 15%, and the intelligent development of coal mines is imperative. As the coal mining industry moves towards automation and intelligence, the equipment level of underground mining working faces has been improved. The "imbalance between mining and excavation" caused by low excavation efficiency has become the main problem facing large underground coal mines today. Therefore, the development of fast and intelligent excavation technology has become an inevitable trend.

[0003] Geological support technology is the prerequisite for key technologies in intelligent tunneling. my country's coal resource conditions vary significantly, the geological structure is complex, and mining is difficult. According to statistics, 75% of major coal mine accidents occur during tunneling, and 80% of disaster accidents are related to hidden geological factors. The State Administration of Mine Safety Supervision requires all coal mining companies to continue to strengthen the census of hidden disaster factors, achieve fine and accurate exploration, promote geological transparency, and provide reliable geological support for the construction of smart mines and transparent mines. Among them, geophysical advance detection plays an increasingly important role in the geological support of tunneling. In order to provide more efficient and high-quality geological support and adapt to the development trend of rapid and intelligent tunneling, advance detection needs to comprehensively identify abnormal structures such as faults and collapse columns to eliminate structural influences. At present, the main technology system that can meet the needs of rapid and intelligent geological support for tunneling is the simultaneous excavation detection technology system, but there are still many key research breakthroughs in this technology system.

[0004] Accurate geological imaging and geological interpretation are the key to the success of geophysical exploration. At present, researchers have conducted extensive and in-depth research on the data processing technology of on-dig seismic exploration and have achieved a lot of research results, but there is still a lack of applicable geological interpretation methods, which makes it difficult for on-dig seismic exploration technology to meet the needs of fast and intelligent tunneling. Therefore, it is very necessary to propose an intelligent quantitative geological interpretation method for tunnel advance detection based on on-dig seismic exploration, which is of great significance to improve the accuracy of geological interpretation of advance detection and ensure safe and efficient production of coal mine tunnel excavation. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to improve the accuracy of geological interpretation in advance detection of seismic exploration while digging.

[0006] The present invention solves the above technical problems through the following technical solutions: an intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging, the method comprising:

[0007] The images of tunnel advance seismic detection imaging results under different geological anomalies are used as training samples to build an intelligent quantitative geological interpretation model for tunnel advance detection.

[0008] Under the same geological anomaly, benchmark samples and associated samples are detected by different observation systems. The binary group consisting of the benchmark samples and associated samples is input into the intelligent quantitative geological interpretation model for tunnel advance detection. The model is trained based on the reverse recognition method and the multi-path recognition loss is calculated. When the loss value is the minimum, the trained intelligent quantitative geological interpretation model for tunnel advance detection is obtained.

[0009] The imaging results of the single-round tunnel excavation seismic advance detection to be interpreted are input into the trained tunnel advance detection intelligent quantitative geological interpretation model for quantitative geological interpretation.

[0010] The present invention trains the model based on the reverse recognition method and calculates the multi-path recognition loss. By weakening the feature expression of the blank background part, the influence of the background on the recognition result of the model is reduced, the feature expression of the geological anomaly imaging reflection belt part is strengthened, and the network is urged to automatically focus on extracting features from the geological anomaly imaging reflection belt part, thereby improving the accuracy of geological interpretation in advanced detection.

[0011] Preferably, the tunnel advance detection imaging result images under different geological anomaly conditions are obtained by theoretical simulation calculations using typical geological anomaly structures, and the tunnel advance detection imaging result images under different geological anomaly positions, anomaly types, and anomaly degrees are obtained under different observation systems. Typical geological anomaly structures include but are not limited to faults, collapse columns, and fracture development areas.

[0012] Preferably, the process of training the model based on the reverse recognition method and calculating the multi-path recognition loss includes:

[0013] The original images in the reference sample and the associated sample are respectively calculated to obtain inverted images, and the inverted images in the associated sample are input into the first reverse process, the original images in the associated sample are input into the first discrimination process, the inverted images in the reference sample are input into the second reverse process, and the original images in the reference sample are input into the second discrimination process for feature extraction respectively;

[0014] The output features of the first reverse process and the output features of the second discriminant process are combined into a first image pair, the output features of the first discriminant process and the output features of the second discriminant process are combined into a second image pair, and the first image pair and the second image pair are respectively input into the fully connected layer to calculate the reverse supervision loss;

[0015] Concatenate the first reverse process output feature with the first discriminant process output feature, and concatenate the second reverse process output feature with the second discriminant process output feature, respectively, to obtain two sets of feature values ​​and calculate the binary loss;

[0016] The reverse supervision loss and the bigram loss are added together to get the multi-way recognition loss.

[0017] Preferably, the first reverse process, the first discrimination process, the second reverse process, and the second discrimination process respectively include five convolution modules, the first convolution module and the second convolution module respectively include two convolution layers and a maximum pooling layer, the third convolution module, the fourth convolution module, and the fifth convolution module respectively include three convolution layers and a maximum pooling layer, and multiple attention mechanism modules are added between the first reverse process and the first discrimination process and between the second reverse process and the second discrimination process, respectively. The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is that the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process.

[0018] Preferably, the forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is a process in which the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process, including:

[0019] The output features F of each convolution module in the reverse process inverse Bilinear interpolation is performed to the same dimension as the output feature of the first convolutional layer in the discriminant process, and global average pooling and global maximum pooling are performed on each channel along the spatial axis. Each two-dimensional feature map is compressed into a real number to obtain a C×1×1 feature vector and Sum to get the eigenvector

[0020] The output feature F of the first convolutional layer of each convolutional module in the discriminant process discrimnative With the eigenvector Multiply, pass the attention information learned from the reverse process to the discriminant process, and add the multiplication result to the output feature F of the discriminant process discrimnative Get the channel refined feature F channel ;

[0021] The feature F after channel refinement channel Perform global average pooling and global maximum pooling along the channel axis to obtain a feature vector of H×W and Sum to get the eigenvector The feature F after channel refinement channel With the eigenvector Multiply and add the channel refined feature F channel Get the channel and space refined features F channel-space .

[0022] Preferably, the method for acquiring the imaging result image of the single-round tunnel excavation seismic advance detection to be interpreted is: utilizing the pseudo-random coal-rock dynamic wave signal generated during the tunnel excavation construction operation, performing real-time data collection underground, transmitting the data to the ground data processing center through the mine optical fiber ring network for data processing and imaging, and obtaining the imaging result image of the single-round tunnel excavation seismic advance detection.

[0023] Preferably, the content of the quantitative geological interpretation includes: the development location of the geological anomaly; what kind of geological anomaly it is; whether the anomaly is a high-speed anomaly or a low-speed anomaly; and the range of the anomaly compared with the physical properties of the background coal strata.

[0024] The present invention also provides an intelligent quantitative geological interpretation device for tunnel advance detection based on seismic exploration while digging, the device comprising:

[0025] A model building unit is used to build an intelligent quantitative geological interpretation model for tunnel advance detection using images of tunnel seismic advance detection imaging results under different geological abnormalities as training samples;

[0026] The model training unit obtains reference samples and associated samples by different observation systems under the same geological anomaly, inputs the binary group consisting of the reference samples and the associated samples into the intelligent quantitative geological interpretation model for tunnel advance detection, trains the model based on the reverse recognition method and calculates the multi-path recognition loss, and obtains the trained intelligent quantitative geological interpretation model for tunnel advance detection when the loss value is the smallest;

[0027] The model recognition unit is used to input the imaging result images of the single-round tunnel excavation seismic advance detection to be interpreted into the trained tunnel advance detection intelligent quantitative geological interpretation model for quantitative geological interpretation.

[0028] Preferably, the tunnel advance detection imaging result images under different geological anomaly conditions are obtained by theoretical simulation calculations using typical geological anomaly structures, and the tunnel advance detection imaging result images under different geological anomaly positions, anomaly types, and anomaly degrees are obtained under different observation system layouts. Typical geological anomaly structures include but are not limited to faults, collapse columns, and fracture development areas.

[0029] Preferably, the process of training the model based on the reverse recognition method and calculating the multi-path recognition loss includes:

[0030] The original images in the reference sample and the associated sample are respectively calculated to obtain inverted images, and the inverted images in the associated sample are input into the first reverse process, the original images in the associated sample are input into the first discrimination process, the inverted images in the reference sample are input into the second reverse process, and the original images in the reference sample are input into the second discrimination process for feature extraction respectively;

[0031] The output features of the first reverse process and the output features of the second discriminant process are combined into a first image pair, the output features of the first discriminant process and the output features of the second discriminant process are combined into a second image pair, and the first image pair and the second image pair are respectively input into the fully connected layer to calculate the reverse supervision loss;

[0032] Concatenate the first reverse process output feature with the first discriminant process output feature, and concatenate the second reverse process output feature with the second discriminant process output feature, respectively, to obtain two sets of feature values ​​and calculate the binary loss;

[0033] The reverse supervision loss and the bigram loss are added together to get the multi-way recognition loss.

[0034] Preferably, the first reverse process, the first discrimination process, the second reverse process, and the second discrimination process respectively include five convolution modules, the first convolution module and the second convolution module respectively include two convolution layers and a maximum pooling layer, the third convolution module, the fourth convolution module, and the fifth convolution module respectively include three convolution layers and a maximum pooling layer, and multiple attention mechanism modules are added between the first reverse process and the first discrimination process and between the second reverse process and the second discrimination process, respectively. The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is that the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process.

[0035] Preferably, the forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is a process in which the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process, including:

[0036] The output features F of each convolution module in the reverse process inverse Bilinear interpolation is performed to the same dimension as the output feature of the first convolutional layer in the discriminant process, and global average pooling and global maximum pooling are performed on each channel along the spatial axis. Each two-dimensional feature map is compressed into a real number to obtain a C×1×1 feature vector and Sum to get the eigenvector

[0037] The output feature F of the first convolutional layer of each convolutional module in the discriminant process discrimnative With the eigenvector Multiply, pass the attention information learned from the reverse process to the discriminant process, and add the multiplication result to the output feature F of the discriminant process discrimnative Get the channel refined feature F channel ;

[0038] The feature F after channel refinement channelPerform global average pooling and global maximum pooling along the channel axis to obtain a feature vector of H×W and Sum to get the eigenvector The feature F after channel refinement channel With the eigenvector Multiply and add the channel refined feature F channel Get the channel and space refined features F channel-space .

[0039] Preferably, the method for acquiring the imaging result image of the single-round tunnel excavation seismic advance detection to be interpreted is: utilizing the pseudo-random coal-rock dynamic wave signal generated during the tunnel excavation construction operation, performing real-time data collection underground, transmitting the data to the ground data processing center through the mine optical fiber ring network for data processing and imaging, and obtaining the imaging result image of the single-round tunnel excavation seismic advance detection.

[0040] Preferably, the content of the quantitative geological interpretation includes: the development location of the geological anomaly; what kind of geological anomaly it is; whether the anomaly is a high-speed anomaly or a low-speed anomaly; and the range of the anomaly compared with the physical properties of the background coal strata.

[0041] The advantages provided by the present invention are:

[0042] (1) The present invention trains the model based on the reverse recognition method and calculates the multi-path recognition loss. By weakening the feature expression of the blank background part, the influence of the background on the recognition result of the model is reduced, the feature expression of the geological anomaly imaging reflection zone is strengthened, and the network is urged to automatically focus on extracting features from the geological anomaly imaging reflection zone, thereby improving the accuracy of geological interpretation in advance detection, and realizing intelligent quantitative geological interpretation of geological anomalies in tunnel excavation, which is of great significance to ensuring the safety and efficiency of coal mine tunnel excavation production.

[0043] (2) The present invention introduces inverted images and uses reverse supervision loss to penalize image pairs with inconsistent judgments between the original detection imaging result image and the inverted image, forcing the model to make the same recognition decision for the original detection imaging result image and the inverted image. In this way, different backgrounds will not affect the recognition results of the model, thereby weakening the feature expression of the blank background part.

[0044] (3) The present invention adds a "reverse recognition method" to the network framework and adds a multi-way attention mechanism module between the discrimination process and the reverse process to accurately locate the identifiable area, enhance the identifiable features of the imaging results, increase the learning intensity of the identifiable features, and enable the network to make the same recognition decision for the imaging results with the foreground and background reversed through training, so as to focus on the geological anomaly imaging reflection zone to extract effective features and reduce the interference of the large background area on feature extraction.

[0045] (4) The present invention performs weighting in the channel and spatial dimensions, making the features more recognizable, and can effectively alleviate the problem of "the same type of anomalies with different degrees of anomaly have small feature differences".

[0046] (5) It can effectively alleviate the problem of insufficient effective information in the tunnel advance detection imaging results, and also reduce the impact of disturbance noise in the background area. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of an intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration during excavation provided by an embodiment of the present invention;

[0048] Figure 2 An image of an imaging result of a tunnel advance detection of an intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration during excavation provided in an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of a reverse identification method in an intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging provided by an embodiment of the present invention;

[0050] Figure 4 A diagram of the architecture of an intelligent quantitative geological interpretation model for tunnel advance detection in an intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while excavating provided in an embodiment of the present invention;

[0051] Figure 5 A schematic diagram of image acquisition of imaging results of a single round of tunnel advance detection based on tunnel seismic exploration provided by an embodiment of the present invention;

[0052] Figure 6 A schematic diagram of the contents of quantitative geological interpretation in the intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while excavating provided by an embodiment of the present invention;

[0053] Figure 7 A schematic diagram of an intelligent quantitative geological interpretation device for tunnel advance detection based on seismic exploration while excavating provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the technical solution of the present invention is clearly and completely described below in combination with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] like Figure 1As shown, this embodiment provides an intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging, including the following steps:

[0056] Step 1: Using the images of tunnel advance seismic detection imaging results under different geological anomalies as training samples, an intelligent quantitative geological interpretation model for tunnel advance detection is constructed.

[0057] See also Figure 2 The tunnel advance detection imaging result images under different geological anomaly conditions are obtained by theoretical simulation calculation based on typical geological anomaly structures. The tunnel advance detection imaging result images under different geological anomaly positions, anomaly types, and anomaly degrees are obtained under different observation systems. Typical geological anomaly structures include but are not limited to faults, collapse columns, and fracture development areas.

[0058] Step 2. Input the binary group consisting of the benchmark sample and the associated sample into the intelligent quantitative geological interpretation model for tunnel advance detection, train the model based on the reverse recognition method and calculate the multi-path recognition loss. When the loss value is the smallest, the trained intelligent quantitative geological interpretation model for tunnel advance detection is obtained; the benchmark sample and the associated sample are detected by different observation systems under the same geological anomaly condition.

[0059] like Figure 4 As shown in the figure, the model includes 4 weight sharing processes, namely 2 discriminant processes and 2 reverse processes. The 2 discriminant processes use the original images of the reference sample and the associated sample image pair as input, respectively, and extract features through the convolution module. The 2 reverse processes use the inverted images of the reference sample and the associated sample image pair as input. The network structure of the reverse process is the same as that of the discriminant process, and both include five convolution modules. Each network is composed of the first convolution module, the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module from its input end to its output end.

[0060] The process of training the model based on the reverse recognition method and calculating the multi-way recognition loss includes:

[0061] The inverted images are obtained by calculating the original images in the reference sample and the associated sample respectively. Figure 4 The original image refers to the original detection imaging result image of the green background and the blue / red abnormal reflection band, and the inverted image refers to the image of the black background and the blue / red abnormal reflection band obtained by calculation from the original detection imaging result image. The inverted image in the associated sample is input into the first reverse process, the original image in the associated sample is input into the first discrimination process, the inverted image in the reference sample is input into the second reverse process, and the original image in the reference sample is input into the second discrimination process for feature extraction respectively;

[0062] See also Figure 3, the features of different processes are merged into two feature maps, the output features of the first reverse process and the output features of the second discriminant process are merged into the first image pair, the first image pair corresponds to the inverted image in the associated sample and the original image in the reference sample, the output features of the first discriminant process and the output features of the second discriminant process are merged into the second image pair, the second image pair corresponds to the original image in the associated sample and the original image in the reference sample, through a global average pooling layer, the merged first image pair and the second image pair are respectively input into the fully connected layer to calculate the reverse supervision loss L reverse ;

[0063] Concatenate the first reverse process output feature with the first discriminant process output feature, and the second reverse process output feature with the second discriminant process output feature, respectively, to obtain two sets of feature values ​​and calculate the binary loss L triplet ; Figure 4 The C in the middle represents the classification and decision-making process.

[0064] The reverse supervision loss L reverse , bigram loss L triplet Add together to get the multi-path recognition loss L duolu :

[0065] L duolu =L triplet +L reverse

[0066] The present invention uses reverse supervision loss to punish image pairs with inconsistent judgments on the original detection imaging result image and the inverted image, forcing the model to make the same recognition decision on the original detection imaging result image and the inverted image. Then, different backgrounds will not affect the recognition results of the model, that is, the feature expression of the blank background part is weakened, and the feature expression of the geological anomaly imaging reflection zone part is strengthened, urging the network to automatically focus on extracting features from the geological anomaly imaging reflection zone part.

[0067] Continue reading Figure 4 The first reverse process, the first discrimination process, the second reverse process, and the second discrimination process each include five convolution modules, and the first convolution module and the second convolution module each include two convolution layers ( Figure 4 Medium blue rectangle), a maximum pooling layer ( Figure 4 The third convolution module, the fourth convolution module, and the fifth convolution module include three convolution layers and one maximum pooling layer respectively. Multi-way attention mechanism modules are added between the first reverse process and the first discrimination process, and between the second reverse process and the second discrimination process respectively. The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and its reverse process is that the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process.

[0068] The present invention utilizes a multi-way attention mechanism module based on channel domain and spatial domain attention, which acts between the discrimination process and the reverse process to link the discrimination process and the reverse process.

[0069] In the present invention, multi-way attention mechanism modules are added between the first reverse process and the first discrimination process, and between the second reverse process and the second discrimination process respectively. The action mode of the attention mechanism module between the first reverse process and the first discrimination process is the same as that between the second reverse process and the second discrimination process. The action mode of the attention mechanism module between the first reverse process and the first discrimination process is taken as an example for explanation below. For the convenience of description, the first reverse process is referred to as the reverse process, and the first discrimination process is referred to as the discrimination process. Those skilled in the art should know that when the action mode of the attention mechanism module between the second reverse process and the second discrimination process is taken as an example for explanation, the reverse process refers to the second reverse process, and the discrimination process refers to the second discrimination process.

[0070] The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is the process of transferring attention information from the deep convolution layer in the reverse process to the shallow convolution layer in the discrimination process, including:

[0071] The output features F of each convolution module in the reverse process inverse Bilinear interpolation is performed to the same dimension as the output feature of the first convolutional layer in the discriminant process, and global average pooling and global maximum pooling are performed on each channel along the spatial axis. Each two-dimensional feature map is compressed into a real number to obtain a C×1×1 feature vector and Sum to get the eigenvector The eigenvalue in represents the importance of each channel. The larger the eigenvalue, the more important the corresponding channel is.

[0072] The output feature F of the first convolutional layer of each convolutional module in the discriminant process discrimnative With the eigenvector Multiply, pass the attention information learned from the reverse process to the discriminant process, and add the multiplication result to the output feature F of the discriminant process discrimnative Get the channel refined feature F channel ;

[0073] The feature F after channel refinement channel Perform global average pooling and global maximum pooling along the channel axis to obtain a feature vector of H×W and Sum to get the eigenvector Each eigenvalue represents the importance of each position, and the channel refined feature Fchannel With the eigenvector Multiply and add the channel refined feature F channel Get the channel and space refined features F channel-space .

[0074] The present invention uses the reverse recognition method to add a multi-way attention mechanism module between the discrimination process and the reverse process to accurately locate the identification area, enhance the identification features of the imaging results, and increase the learning intensity of the identification features. Through training, the network makes the same identification decision for the imaging results with the foreground and background reversed, prompting the network to focus on the geological anomaly imaging reflection zone to extract effective features and reduce the interference of the large background area on feature extraction.

[0075] Step 3: Input the image of the single-round tunnel advance detection imaging result to be interpreted into the trained tunnel advance detection intelligent quantitative geological interpretation model for quantitative geological interpretation.

[0076] like Figure 5 As shown, the method for acquiring the imaging result image of the single-round tunnel advance detection during excavation is as follows: using the pseudo-random coal-rock dynamic wave signal generated during the tunnel excavation operation, real-time data collection is carried out underground, and the data is transmitted to the ground data processing center through the mine optical fiber ring network for data processing and imaging, so as to obtain the imaging result image of the single-round tunnel advance detection.

[0077] like Figure 6 As shown in the figure, the image of the imaging result of the single-round tunnel advance detection to be interpreted is input into the trained intelligent quantitative geological interpretation model of tunnel advance detection to determine the characteristic value of the geological anomaly condition of this round, which includes the location of the geological anomaly, the type of anomaly, and the degree of anomaly. The content of quantitative geological interpretation includes: the development location of the geological anomaly; what kind of geological anomaly it is, the types mainly include lithological interface, fault fracture zone, collapse column, fracture development zone, etc.; whether the anomaly is a high-speed anomaly or a low-speed anomaly, and how much is its anomaly range compared with the physical properties of the background coal and rock formations.

[0078] In quantitative geological interpretation, it is assumed that the image of the associated sample comes from the detection result of geological anomaly abnormal1, and the image of the benchmark sample also comes from the detection result of geological anomaly abnormal1. If the model correctly identifies the geological anomaly information in the image, then the anomaly recognition decision should be independent of the color of the imaging result image and the image background. As long as the anomaly information matches the same network decision, the model should make the same recognition decision for the original image and the inverted image. Therefore, the model will make the same recognition judgment for the inverted image in the associated sample and the original image in the benchmark sample, and the original image in the associated sample and the original image in the benchmark sample. In the associated sample, no matter the original image or the inverted image, the condition feature value of geological anomaly abnormal1 will be quantitatively identified, and then the geological anomaly abnormal1 can be quantitatively interpreted.

[0079] The present invention also provides an intelligent quantitative geological interpretation device for tunnel advance detection based on seismic exploration while digging, comprising:

[0080] The model building unit is used to build an intelligent quantitative geological interpretation model for tunnel advance detection using the tunnel seismic advance detection imaging result images under different geological anomaly conditions as training samples; the tunnel advance detection imaging result images under different geological anomaly conditions are obtained by theoretical simulation calculations using typical geological anomaly structures, and the tunnel advance detection imaging result images under different geological anomaly positions, anomaly types, and anomaly degrees are obtained under different observation systems. Typical geological anomaly structures include but are not limited to faults, collapse columns, and fracture development areas.

[0081] The model training unit is used to obtain benchmark samples and associated samples detected by different observation systems under the same geological anomaly conditions, input the binary group consisting of the benchmark samples and the associated samples into the intelligent quantitative geological interpretation model of tunnel advance detection, train the model based on the reverse recognition method and calculate the multi-path recognition loss, and when the loss value is the smallest, obtain the trained intelligent quantitative geological interpretation model of tunnel advance detection.

[0082] The process of training the model based on the reverse recognition method and calculating the multi-way recognition loss includes:

[0083] The original images in the reference sample and the associated sample are respectively calculated to obtain inverted images, and the inverted images in the associated sample are input into the first reverse process, the original images in the associated sample are input into the first discrimination process, the inverted images in the reference sample are input into the second reverse process, and the original images in the reference sample are input into the second discrimination process for feature extraction respectively;

[0084] The output features of the first reverse process and the output features of the second discriminant process are combined into a first image pair, the output features of the first discriminant process and the output features of the second discriminant process are combined into a second image pair, and the first image pair and the second image pair are respectively input into the fully connected layer to calculate the reverse supervision loss;

[0085] Concatenate the first reverse process output feature with the first discriminant process output feature, and concatenate the second reverse process output feature with the second discriminant process output feature, respectively, to obtain two sets of feature values ​​and calculate the binary loss;

[0086] The reverse supervision loss and the bigram loss are added together to get the multi-way recognition loss.

[0087] Among them, the first reverse process, the first discrimination process, the second reverse process, and the second discrimination process respectively include five convolution modules, the first convolution module and the second convolution module respectively include two convolution layers and a maximum pooling layer, the third convolution module, the fourth convolution module, and the fifth convolution module respectively include three convolution layers and a maximum pooling layer, and multiple attention mechanism modules are added between the first reverse process and the first discrimination process and between the second reverse process and the second discrimination process. The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is that the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process.

[0088] Specifically, the forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is the process of transferring attention information from the deep convolution layer in the reverse process to the shallow convolution layer in the discrimination process, including:

[0089] The output features F of each convolution module in the reverse process inverse Bilinear interpolation is performed to the same dimension as the output feature of the first convolutional layer in the discriminant process, and global average pooling and global maximum pooling are performed on each channel along the spatial axis. Each two-dimensional feature map is compressed into a real number to obtain a C×1×1 feature vector and Sum to get the eigenvector

[0090] The output feature F of the first convolutional layer of each convolutional module in the discriminant process discrimnative With the eigenvector Multiply, pass the attention information learned from the reverse process to the discriminant process, and add the multiplication result to the output feature F of the discriminant process discrimnative Get the channel refined feature F channel ;

[0091] The feature F after channel refinement channelPerform global average pooling and global maximum pooling along the channel axis to obtain a feature vector of H×W and Sum to get the eigenvector The feature F after channel refinement channel With the eigenvector Multiply and add the channel refined feature F channel Get the channel and space refined features F channel-space .

[0092] The model recognition unit is used to input the image of the imaging result of the single-round tunnel excavation seismic advance detection to be interpreted into the trained intelligent quantitative geological interpretation model of tunnel advance detection for quantitative geological interpretation. The pseudo-random coal-rock dynamic wave signal generated during the tunnel excavation construction operation is used to collect real-time data underground, and the data is transmitted to the ground data processing center through the mine optical fiber ring network for data processing and imaging, so as to obtain the image of the imaging result of the single-round tunnel excavation seismic advance detection. The content of quantitative geological interpretation includes: the development location of geological anomalies; what kind of geological anomalies; whether the anomaly is a high-speed anomaly or a low-speed anomaly; and the range of the anomaly compared with the physical properties of the background coal and rock formations.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging, characterized by: The method comprises: The images of tunnel advance seismic detection imaging results under different geological anomalies are used as training samples to build an intelligent quantitative geological interpretation model for tunnel advance detection. Under the same geological anomaly, benchmark samples and associated samples are detected by different observation systems. The binary group consisting of the benchmark samples and associated samples is input into the intelligent quantitative geological interpretation model for tunnel advance detection. The model is trained based on the reverse recognition method and the multi-path recognition loss is calculated. When the loss value is the minimum, the trained intelligent quantitative geological interpretation model for tunnel advance detection is obtained. The imaging results of the single-round tunnel excavation seismic advance detection to be interpreted are input into the trained tunnel advance detection intelligent quantitative geological interpretation model for quantitative geological interpretation.

2. The intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration during excavation according to claim 1 is characterized by: The tunnel advance detection imaging result images under different geological anomaly conditions are obtained by theoretical simulation calculations using typical geological anomaly structures, and the tunnel advance detection imaging result images under different geological anomaly positions, anomaly types, and anomaly degrees are obtained under different observation systems. The typical geological anomaly structures include but are not limited to faults, collapse columns, and fracture development areas.

3. The intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration during excavation according to claim 1 is characterized by: The process of training the model based on the reverse recognition method and calculating the multi-path recognition loss includes: The original images in the reference sample and the associated sample are respectively calculated to obtain inverted images, and the inverted images in the associated sample are input into the first reverse process, the original images in the associated sample are input into the first discrimination process, the inverted images in the reference sample are input into the second reverse process, and the original images in the reference sample are input into the second discrimination process for feature extraction respectively; The output features of the first reverse process and the output features of the second discriminant process are combined into a first image pair, the output features of the first discriminant process and the output features of the second discriminant process are combined into a second image pair, and the first image pair and the second image pair are respectively input into the fully connected layer to calculate the reverse supervision loss; Concatenate the first reverse process output feature with the first discriminant process output feature, and concatenate the second reverse process output feature with the second discriminant process output feature, respectively, to obtain two sets of feature values ​​and calculate the binary loss; The reverse supervision loss and the bigram loss are added together to get the multi-way recognition loss.

4. The intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging according to claim 3 is characterized by: The first reverse process, the first discrimination process, the second reverse process, and the second discrimination process respectively include five convolution modules, the first convolution module and the second convolution module respectively include two convolution layers and a maximum pooling layer, the third convolution module, the fourth convolution module, and the fifth convolution module respectively include three convolution layers and a maximum pooling layer, and multiple attention mechanism modules are added between the first reverse process and the first discrimination process, and between the second reverse process and the second discrimination process. The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is that the deep convolution layer in the reverse process transfers attention information to the shallow convolution layer in the discrimination process.

5. The intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging according to claim 4 is characterized by: The forward process of the attention mechanism module receives the first layer features of the convolution module in the discrimination process, and the reverse process is the process of transferring attention information from the deep convolution layer in the reverse process to the shallow convolution layer in the discrimination process, including: The output features F of each convolution module in the reverse process inverse Bilinear interpolation is performed to the same dimension as the output feature of the first convolutional layer in the discriminant process, and global average pooling and global maximum pooling are performed on each channel along the spatial axis. Each two-dimensional feature map is compressed into a real number to obtain a C×1×1 feature vector and Sum to get the eigenvector The output feature F of the first convolutional layer of each convolutional module in the discriminant process discrimnative With the eigenvector Multiply, pass the attention information learned from the reverse process to the discriminant process, and add the multiplication result to the output feature F of the discriminant process discrimnative Get the channel refined feature F channel ; The feature F after channel refinement channel Perform global average pooling and global maximum pooling along the channel axis to obtain a feature vector of H×W and Sum to get the eigenvector The feature F after channel refinement channel With the eigenvector Multiply and add the channel refined feature F channel Get the channel and space refined features F channel-space .

6. The intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration while digging according to claim 1 is characterized by: The method for obtaining the imaging result image of the single-round tunnel excavation seismic advance detection to be interpreted is: using the pseudo-random coal and rock dynamic wave signals generated during the tunnel excavation construction operation, real-time data collection is carried out underground, and the data is transmitted to the ground data processing center through the mine optical fiber ring network for data processing and imaging, so as to obtain the imaging result image of the single-round tunnel excavation seismic advance detection.

7. The intelligent quantitative geological interpretation method for tunnel advance detection based on seismic exploration during excavation according to claim 1 is characterized by: The content of the quantitative geological interpretation includes: the development location of the geological anomaly; what kind of geological anomaly it is; whether the anomaly is a high-speed anomaly or a low-speed anomaly; and what is the range of the anomaly compared with the physical properties of the background coal and rock formations.

8. An intelligent quantitative geological interpretation device for tunnel advance detection based on seismic exploration while digging, characterized by: The device comprises: A model building unit is used to build an intelligent quantitative geological interpretation model for tunnel advance detection using images of tunnel seismic advance detection imaging results under different geological anomalies as training samples; The model training unit is used to obtain reference samples and associated samples detected by different observation systems under the same geological anomaly, input the binary group consisting of the reference samples and the associated samples into the intelligent quantitative geological interpretation model for tunnel advance detection, train the model based on the reverse recognition method and calculate the multi-path recognition loss, and obtain the trained intelligent quantitative geological interpretation model for tunnel advance detection when the loss value is the smallest; The model recognition unit is used to input the imaging result images of the single-round tunnel excavation seismic advance detection to be interpreted into the trained tunnel advance detection intelligent quantitative geological interpretation model for quantitative geological interpretation.

9. The intelligent quantitative geological interpretation device for tunnel advance detection based on seismic exploration while digging according to claim 8 is characterized by: The tunnel advance detection imaging result images under different geological anomaly conditions are obtained by theoretical simulation calculations using typical geological anomaly structures, and the tunnel advance detection imaging result images under different geological anomaly positions, anomaly types, and anomaly degrees are obtained under different observation systems. The typical geological anomaly structures include but are not limited to faults, collapse columns, and fracture development areas.

10. The intelligent quantitative geological interpretation device for tunnel advance detection based on seismic exploration while digging according to claim 8, characterized in that: The process of training the model based on the reverse recognition method and calculating the multi-path recognition loss includes: The original images in the reference sample and the associated sample are respectively calculated to obtain inverted images, and the inverted images in the associated sample are input into the first reverse process, the original images in the associated sample are input into the first discrimination process, the inverted images in the reference sample are input into the second reverse process, and the original images in the reference sample are input into the second discrimination process for feature extraction respectively; The output features of the first reverse process and the output features of the second discriminant process are combined into a first image pair, the output features of the first discriminant process and the output features of the second discriminant process are combined into a second image pair, and the first image pair and the second image pair are respectively input into the fully connected layer to calculate the reverse supervision loss; Concatenate the first reverse process output feature with the first discriminant process output feature, and concatenate the second reverse process output feature with the second discriminant process output feature, respectively, to obtain two sets of feature values ​​and calculate the binary loss; The reverse supervision loss and the bigram loss are added together to get the multi-way recognition loss.

Images