A multimodal intelligent fine classification method and system for tunnel surrounding rock

Abstract

The present invention belongs to the field of tunnel engineering and specifically discloses a multimodal intelligent fine grading method and system for tunnel surrounding rock, comprising the following steps: 1. constructing a sample library comprising drilling parameters, digital images, and corresponding surrounding rock grades of the tunnel face; 2. constructing a fine grading feature system for drilling parameters; 3. constructing a surrounding rock intelligent grading model based on multimodal data fusion; 4. dividing the rock mass in front of the tunnel face into a number of three-dimensional spatial units based on the tunnel profile and blasthole distribution characteristics; and 5. finely and intelligently identifying the surrounding rock grade of the tunnel face based on the surrounding rock intelligent grading model. The present invention constructs a surrounding rock intelligent grading model based on multimodal data fusion theory and realizes fine intelligent identification of the surrounding rock grade of the tunnel face, thereby improving the accuracy and robustness of the model's fine intelligent identification of the surrounding rock grade, and improving the surrounding rock grading level during the construction phase, providing a more comprehensive reference for subsequent construction plans, and guiding the intelligent construction of tunnels.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent tunnel construction, and in particular to a multi-modal intelligent fine grading method and system for tunnel surrounding rock. Background Art

[0002] In tunnel construction, accurate identification of surrounding rock mass is crucial to construction safety and progress. Traditional rock mass classification relies primarily on manual experience and on-site surveys, which are subject to subjectivity and low precision, making it difficult to adapt to the demands of refined construction under complex geological conditions. With the advancement of tunnel excavation technology, a large number of face drilling parameters and digital images can be acquired in real time. Intelligent identification of surrounding rock mass using artificial intelligence technologies such as deep learning and machine learning avoids human error and enables more efficient processing and analysis of complex construction environment data.

[0003] At the same time, the surrounding rock structure is complex and changeable, with significant local differences. The identification of surrounding rock grades should meet refined standards in order to formulate more accurate construction measures and reduce construction risks. Therefore, a single modality and overall classification are often insufficient to fully reflect the true situation of the rock mass in front of the face. The multimodal intelligent and refined classification of surrounding rock combines drilling parameters and face image information, comprehensively considers the characteristics of each modal data, and achieves a more comprehensive and accurate identification of surrounding rock grades. At the same time, it effectively avoids the roughness of overall identification and can identify changes in local surrounding rock grades. Therefore, the multimodal intelligent and refined classification of tunnel surrounding rock can effectively improve the intelligence level and safety of tunnel construction, assist in scientific decision-making during the construction process, and reduce engineering risks. Summary of the Invention

[0004] In order to solve the problems existing in the prior art, the present invention provides a method and system for multimodal intelligent fine grading of tunnel surrounding rock, which can realize multimodal intelligent fine grading of tunnel surrounding rock and solve the problems mentioned in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-modal intelligent fine classification method and system for tunnel surrounding rock, comprising the following steps:

[0006] S1. Build a sample library, which includes original drilling parameters, digital images, and corresponding surrounding rock grades of the tunnel face;

[0007] S2. Constructing a refined classification feature system for drilling parameters;

[0008] S3. Constructing a surrounding rock intelligent classification model based on multimodal data fusion;

[0009] S4. Divide the rock mass in front of the tunnel face into several three-dimensional spatial units based on the tunnel contour and blasthole distribution characteristics;

[0010] S5. Based on the intelligent surrounding rock classification model, the surrounding rock level of the tunnel face is intelligently identified.

[0011] Preferably, the tunnel face drilling parameters of the sample library in step S1 include feed speed, impact pressure, thrust pressure and rotary pressure automatically collected by the intelligent drilling rig, digital images of the tunnel face are collected by the rig's onboard industrial camera, and the tunnel face surrounding rock grade is determined by a standard method, including grade II, grade III, grade IV and grade V.

[0012] Preferably, the step S2 constructs a refined classification feature system for drilling parameters, specifically including the following:

[0013] S21. Calculate five rock drillability indices, including equivalent propulsion force, equivalent impact force, difficulty-to-drill index, E index, and equivalent feed rate, based on four original drilling parameters.

[0014] S22. Taking the tunnel face excavation cycle as a unit, six statistical characteristics of four original drilling parameters and five rock drillability indices are calculated, including the mean, first quartile, second quartile, third quartile, standard deviation, and coefficient of variation;

[0015] S23. Taking the tunnel face excavation cycle as a unit, generate a 9×2×3 three-dimensional feature vector of the while-drilling parameters according to the original while-drilling parameters, rock drillability index, and statistical feature type.

[0016] Preferably, the step S21 specifically includes: 5 rock drillability indicators and propulsion force F f , impact force F h and the rotational torque M r The calculation formula is:

[0017] Γ ff =F f / (M r / D)

[0018] Γ fh =F h / (M r / D)

[0019] Γ v =V p / (V r D)

[0020] Γ hard =Γ ff Γ fh / Γ v

[0021]

[0022] Where: Γff is the equivalent propulsion force; F f is the thrust, in N; D is the drill diameter, in mm; M r is the rotational torque, in N·m; Γ fh is the equivalent impact force; F h is the impact force, in N; Γ v is the equivalent feed rate; V p is the feed rate, in m / min; V r is the drilling speed of the drill, in r / min; Γ hard is the difficulty to drill indicator; E is the E indicator; D f is the diameter of the rear end of the piston, in mm; P f is the propulsion pressure, unit is Pa; D hB is the diameter of the rear end of the piston, in mm; m h is the mass of the piston, in kg; P h is the impact pressure, unit is Pa; S h is the impact stroke, in mm; t h is the time of impacting the fiber tail, in seconds; P r is the rotation pressure, unit Pa; q r is the motor displacement, unit is ml / r; i r is the reduction ratio.

[0023] Preferably, step S22 specifically includes: quartiles are obtained by sorting the data from small to large and then dividing the sorted data into four equal parts, wherein the first quartile (Q1) corresponds to the value at the 25th percentile position, the second quartile (Q2) is the median, corresponding to the value at the 50th percentile position, and the third quartile (Q3) corresponds to the value at the 75th percentile position. The mean, standard deviation, and coefficient of variation are calculated as follows:

[0024]

[0025]

[0026] Where: u c is the mean of the data points; c i is a single data point; n c is the total number of data points; σ c is the standard deviation of the data points; c v,c is the coefficient of variation of the data points.

[0027] Preferably, step S23 specifically includes: combining the original while-drilling parameters and rock drillability indexes, totaling 9 features, with the first quartile and the second quartile respectively to form a first two-dimensional feature vector, then combining them with the third quartile and the standard deviation to form a second two-dimensional feature vector, and finally combining them with the mean and the coefficient of variation to form a third two-dimensional feature vector, and then stacking these three two-dimensional feature vectors vertically on a two-dimensional plane to construct a 9×2×3 three-dimensional feature vector of the while-drilling parameters.

[0028] Preferably, the step S3 of constructing a multimodal data fusion surrounding rock intelligent classification model specifically includes the following steps:

[0029] S31. Perform feature extraction through a dual convolutional neural network and obtain the classification probability of each modality;

[0030] S32. Establish a weight adaptive adjustment mechanism for the digital image modality of the face and the drilling parameter modality to achieve automatic training of the multimodal model.

[0031] Preferably, step S31 specifically includes: using a dual convolutional neural network algorithm, taking the digital image of the face and the three-dimensional feature vector of the drilling parameters as input, the network of each mode is provided with a convolution layer and a pooling layer, performing linear transformation in the fully connected layer, and using the Softmax function to calculate the classification probability of the four surrounding rock levels of each mode.

[0032] Preferably, the step S32 specifically includes: establishing a weight adaptive adjustment mechanism for the tunnel face digital image modality and the while-drilling parameter modality through a cross entropy loss function, that is, a mapping relationship between the graded probabilities of different modalities and the cross entropy function:

[0033]

[0034] Where: L is the loss rate; y i is the actual label; p i is the model prediction probability; K is the total number of label categories; i is the index variable, from 1 to K; weight c is the weight of the c-mode; L c is the loss rate of the c-mode.

[0035] Based on the weight of each mode, the weighted classification probabilities of the four surrounding rock grades are calculated respectively. The weighted classification probabilities of the same surrounding rock grade under two modes are added together to obtain the classification probabilities of the four surrounding rock grades of the tunnel surrounding rock multimodal intelligent classification model. The total loss rate of the tunnel surrounding rock multimodal intelligent classification model is calculated using the cross-entropy loss function, and the loss rate is distributed according to the weight of each mode. Backpropagation is performed to update the model parameters.

[0036] Preferably, in step S4, dividing the rock mass in front of the tunnel face into a plurality of three-dimensional space units according to the tunnel profile and blasthole distribution characteristics specifically includes the following steps:

[0037] S41. Divide the tunnel face into 18 horizontal blocks along the width and height of the tunnel in the two-dimensional transverse direction according to the tunnel contour and blasthole distribution.

[0038] S42. The tunnel face area is further segmented along the tunneling direction to eventually form several three-dimensional space units.

[0039] Preferably, the step S41 specifically includes: taking the center line of the tunnel and the bottom of the pit as the reference, first drawing the horizontal line and the vertical line, and then offsetting them by 2 to 4 meters in the vertical and left and right directions respectively, and setting the maximum size of a single block to 2 to 4 meters until the outline boundary of the tunnel is covered; finally, dividing the entire tunnel outline into 18 blocks, numbered from K-1 to K-18, with an area of ​​3m 2 Up to 8m 2 between.

[0040] Preferably, step S42 specifically includes: setting the spacing of the longitudinal segments to 0.4 m, dividing the tunnel face into blocks in the width and height directions in the transverse direction, combining the longitudinal partitions, and finally extending these two-dimensional partitions along the depth in front of the tunnel face to form a plurality of three-dimensional space units.

[0041] Preferably, the refined intelligent identification of the tunnel face surrounding rock level based on the surrounding rock intelligent grading model in step S5 includes constructing a three-dimensional feature vector of the drilling parameters of each unit body, splitting the tunnel face digital image according to the same horizontal block method, and inputting the three-dimensional feature vector of the drilling parameters of each unit body and the corresponding horizontal block tunnel face digital image into the tunnel surrounding rock multimodal intelligent grading model to obtain the surrounding rock level of any unit body, thereby realizing refined intelligent identification of the tunnel face surrounding rock level.

[0042] In addition, to achieve the above-mentioned purpose, the present invention also provides the following technical solution: a multi-modal intelligent fine grading system for tunnel surrounding rock, the system comprising the following modules:

[0043] Sample library construction module: Build a sample library, which includes the drilling parameters, digital images and corresponding surrounding rock grades of the tunnel face;

[0044] Feature construction module: Constructs a refined hierarchical feature system for drilling parameters;

[0045] Surrounding rock intelligent classification model construction module: Constructing a surrounding rock intelligent classification model based on multimodal data fusion;

[0046] The rock mass division module in front of the tunnel face: divides the rock mass in front of the tunnel face into several three-dimensional space units according to the tunnel contour and blasthole distribution characteristics;

[0047] Intelligent and refined identification module for the surrounding rock grade of the tunnel face: Based on the intelligent surrounding rock classification model, it can intelligently and refinedly identify the surrounding rock grade of the tunnel face.

[0048] The beneficial effects of the present invention are as follows: the present invention performs multimodal fusion of the face drilling parameters collected during tunnel excavation with digital images and divides the rock mass in front of the face, thereby realizing multimodal intelligent and fine grading of the tunnel surrounding rock, improving the comprehensiveness and accuracy of surrounding rock grade identification, avoiding the roughness of overall identification and the subjectivity of human identification, facilitating the formulation of more accurate construction plans, reducing construction risks, and facilitating intelligent tunnel construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic flow chart of a multi-modal intelligent fine classification method for tunnel surrounding rock in Example 1;

[0050] Figure 2 This is a schematic diagram of a digital image of a tunnel face captured by an airborne industrial camera in Example 1;

[0051] Figure 3 Schematic diagram of constructing the three-dimensional characteristic vector of drilling parameters in Example 1;

[0052] Figure 4 Schematic diagram of building a surrounding rock intelligent classification model based on multimodal data fusion in Example 1;

[0053] Figure 5 This is a schematic diagram of the horizontal division and numbering of the tunnel face in Example 1;

[0054] Figure 6 This is an example diagram of the surrounding rock intelligent fine classification results in Example 1;

[0055] Figure 7 This is a schematic diagram of the multi-modal intelligent fine classification system module for tunnel surrounding rock;

[0056] In the figure, 110 is a sample library construction module; 120 is a feature construction module; 130 is a surrounding rock intelligent classification model construction module; 140 is a rock mass classification module in front of the tunnel face; 150 is a tunnel face surrounding rock grade intelligent and refined identification module. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0058] Example 1

[0059] In tunnel construction, accurate identification of surrounding rock grade is crucial to construction safety and progress. Traditional surrounding rock grading relies primarily on manual experience and on-site surveys, which are subject to high subjectivity and low precision, making it difficult to adapt to the demands of refined construction under complex geological conditions.

[0060] To this end, through long-term research and practice, the inventors have proposed a method and system for multimodal intelligent fine grading of tunnel surrounding rock. This system utilizes multimodal fusion of face-on-drilling parameters and digital images collected during tunnel excavation, and the division of the rock mass ahead of the face. This method aims to provide a fast and accurate method for multimodal intelligent fine grading of surrounding rock, improving the accuracy and refinement of surrounding rock grading, facilitating scientific decision-making during construction, and reducing engineering risks.

[0061] See also Figure 1 This embodiment provides a technical solution: a multi-modal intelligent fine classification method for tunnel surrounding rock, which includes the following steps:

[0062] Step S1: Construct a sample library, which includes the original drilling parameters, digital images and corresponding surrounding rock grades of the face. The original drilling parameters of the sample library include four items, such as feed speed, impact pressure, thrust pressure and rotary pressure, which are automatically collected by the intelligent drilling rig. Every time 0.02m is drilled, the intelligent drilling rig automatically collects a set of drilling parameters. The digital image of the face is collected by the rig's airborne industrial camera. An example of the data of a certain blasthole of a face recorded by the intelligent drilling rig is shown in Table 1. An example of the digital image of the face collected by the airborne industrial camera is shown in Table 1. Figure 2 shown.

[0063] Table 1 Example of data of a certain blasthole on a tunnel face recorded by the intelligent drilling rig

[0064]

[0065] The surrounding rock grades of the tunnel faces in the sample library are obtained by standard classification according to the "Code for Design of Railway Tunnel TB 10003-2016". The surrounding rock grades can be divided into grades I to VI. The common surrounding rock grades in actual projects are grades II to V.

[0066] Step S2: Construct a refined classification feature system for while-drilling parameters.

[0067] 1) Based on the four original drilling parameters, five rock drillability indices including equivalent propulsion force, equivalent impact force, difficulty in drilling index, E index, and equivalent feed rate are calculated. f , impact force F h and the rotational torque M r The calculation formula is:

[0068] Γ ff =F f / (M r / D) Formula (1)

[0069] Γ fh =F h / (M r / D) Formula (2)

[0070] Γ v =V p / (V r D) Formula (3)

[0071] Γ hard =Γ ff Γ fh / Γ v Formula (4)

[0072]

[0073] Where: Γ ff is the equivalent propulsion force; F f is the thrust, in N; D is the drill diameter, in mm; M r is the rotational torque, in N·m; Γ fh is the equivalent impact force; F h is the impact force, in N; Γ v is the equivalent feed rate; V p is the feed rate, in m / min; V r is the drilling speed of the drill, in r / min; Γ hard is the difficulty to drill indicator; E is the E indicator; D f is the diameter of the rear end of the piston, in mm; P f is the propulsion pressure, unit is Pa; D hB is the diameter of the rear end of the piston, in mm; m h is the mass of the piston, in kg; P h is the impact pressure, unit is Pa; S h is the impact stroke, in mm; t h is the time of impacting the fiber tail, in seconds; P ris the rotation pressure, unit Pa; q r is the motor displacement, unit is ml / r; i r is the reduction ratio.

[0074] 2) Taking the tunnel face excavation cycle as the unit, six statistical characteristics of four original drilling parameters and five rock drillability indices were collected, including the mean, first quartile, second quartile, third quartile, standard deviation, and coefficient of variation.

[0075] Quartiles are obtained by sorting the data from smallest to largest and then dividing the sorted data into four equal parts. The first quartile (Q1) corresponds to the value at the 25th percentile, the second quartile (Q2) is the median, corresponding to the value at the 50th percentile, and the third quartile (Q3) corresponds to the value at the 75th percentile. The formulas for calculating the mean, standard deviation, and coefficient of variation are:

[0076]

[0077] Where: u c is the mean of the data points; c i is a single data point; n c is the total number of data points; σ c is the standard deviation of the data points; c v,c is the coefficient of variation of the data points.

[0078] 3) Taking the tunnel face excavation cycle as a unit, a 9×2×3 three-dimensional feature vector of the drilling parameters is generated based on the original drilling parameters, rock drillability index, and statistical feature types.

[0079] like Figure 3 As shown in Figure 1, f1 to f9 are the original while-drilling parameters and rock drillability indicators, totaling 9 features. Q1, Q2, Q3, Std, Mean, and CV are the first quartile, second quartile, first tertile, standard deviation, mean, and coefficient of variation. The original while-drilling parameters and rock drillability indicators, totaling 9 features, are first combined with the first quartile and second quartile to form the first two-dimensional feature vector (M Q1,Q2 ), and then combined with the third quartile and standard deviation to form the second two-dimensional feature vector (M Q3,Std ), and finally combined with the mean and coefficient of variation to form the third two-dimensional feature vector (M Mean,CV ), and then these three two-dimensional feature vectors are stacked vertically on the two-dimensional plane to construct a 9×2×3 three-dimensional feature vector of the drilling parameters.

[0080] Step S3: Constructing a surrounding rock intelligent classification model based on multimodal data fusion.

[0081] 1) Feature extraction is performed through a dual convolutional neural network, and the classification probability of each modality is obtained;

[0082] like Figure 4 As shown in the figure, a dual convolutional neural network algorithm is used, with the digital image of the tunnel face and the three-dimensional feature vector of the drilling parameters as input. The network of each mode is equipped with a convolution layer and a pooling layer, a linear transformation is performed in the fully connected layer, and the Softmax function is used to calculate the classification probability of the four surrounding rock grades for each mode.

[0083] 2) Establish an adaptive weight adjustment mechanism for the digital image modality of the face and the drilling parameter modality to achieve automatic training of the multimodal model.

[0084] like Figure 4 As shown in the figure, after calculating the classification probabilities of the two modalities, the weight adaptive adjustment mechanism of the tunnel face digital image modality and the drilling parameter modality is established through the cross entropy loss function, that is, the mapping relationship between the classification probabilities of different modalities and the cross entropy function:

[0085]

[0086] Where: L is the loss rate; y i is the actual label; p i is the model prediction probability; K is the total number of label categories; i is the index variable, from 1 to K; weight c is the weight of the c-mode; L c is the loss rate of the c-mode.

[0087] Based on the weight of each mode, the weighted classification probabilities of the four surrounding rock grades are calculated respectively. The weighted classification probabilities of the same surrounding rock grade under two modes are added together to obtain the classification probabilities of the four surrounding rock grades of the tunnel surrounding rock multimodal intelligent classification model. The total loss rate of the tunnel surrounding rock multimodal intelligent classification model is calculated using the cross-entropy loss function, and the loss rate is distributed according to the weight of each mode. Backpropagation is performed to update the model parameters.

[0088] Step S4: Divide the rock mass in front of the tunnel face into several three-dimensional space units according to the tunnel contour and blasthole distribution characteristics.

[0089] 1) In the two-dimensional transverse direction of the tunnel face, along the width and height of the tunnel, the tunnel face is divided into 18 transverse blocks according to the tunnel contour and blasthole distribution;

[0090] like Figure 5As shown, with the tunnel centerline and pit bottom as the reference, first draw the horizontal and vertical lines, and then offset them 2 to 4 meters in the up and down, left and right directions respectively. The maximum size of a single block is set to 2 to 4 meters until the outline of the tunnel is covered. Finally, the entire tunnel outline is divided into 18 blocks, numbered from K-1 to K-18, with an area of ​​3m 2 Up to 8m 2 between.

[0091] 2) The tunnel face area is further segmented along the tunneling direction, ultimately forming several three-dimensional spatial units. The longitudinal segment spacing is set at 0.4m. By dividing the tunnel face in the transverse width and height directions, combined with the longitudinal partitioning, these two-dimensional partitions are ultimately extended along the depth in front of the tunnel face to form several three-dimensional spatial units.

[0092] Step S5: Based on the intelligent surrounding rock classification model, fine-grained intelligent identification of the tunnel face surrounding rock level. Fine-grained intelligent identification of the tunnel face surrounding rock level based on the intelligent surrounding rock classification model includes: constructing a three-dimensional feature vector of the drilling parameters of each unit body, splitting the tunnel face digital image according to the same horizontal block method, inputting the three-dimensional feature vector of the drilling parameters of each unit body and the corresponding horizontal block tunnel face digital image into the tunnel surrounding rock multimodal intelligent classification model, obtaining the surrounding rock level of any unit body, and realizing fine-grained intelligent identification of the tunnel face surrounding rock level. An example of the surrounding rock fine-grained identification result is as follows: Figure 6 shown.

[0093] Example 2

[0094] Based on the same inventive concept as the above method embodiment, the present application embodiment also provides a tunnel surrounding rock multi-modal intelligent fine classification system, which can realize the functions provided by the above method embodiment, such as Figure 7 As shown, the system modules include the following:

[0095] Sample library construction module 110: constructs a sample library, which includes original drilling parameters, digital images and corresponding surrounding rock grades of the tunnel face;

[0096] Feature construction module 120: constructing a refined classification feature system for drilling parameters;

[0097] Surrounding rock intelligent classification model construction module 130: constructing a surrounding rock intelligent classification model based on multimodal data fusion;

[0098] The rock mass division module 140 in front of the tunnel face is used to divide the rock mass in front of the tunnel face into a number of three-dimensional space units according to the tunnel contour and the blasthole distribution characteristics;

[0099] Intelligent and refined identification module 150 for the surrounding rock grade of the tunnel face: Based on the intelligent surrounding rock classification model, it performs refined and intelligent identification of the surrounding rock grade of the tunnel face.

[0100] The present invention realizes multimodal intelligent grading of tunnel surrounding rock based on the face drilling parameters and digital images automatically collected by the intelligent drilling rig and the airborne industrial camera. After the intelligent drilling rig and the airborne industrial camera collect the face drilling parameters and digital images, the method can call the multimodal data fusion surrounding rock intelligent grading model to realize multimodal intelligent grading of surrounding rock without the participation of geological personnel. Compared with the traditional manual surrounding rock grading method, it avoids the subjectivity of manual identification of surrounding rock grade, reduces the error of manual identification, improves the accuracy of surrounding rock grade identification, and enhances the refinement level of surrounding rock grade identification. This method contributes to scientific decision-making during the construction process and facilitates the intelligent construction of tunnels.

[0101] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-modal intelligent fine classification method for tunnel surrounding rock, characterized by: The following steps are involved: S1. Build a sample library, which includes original drilling parameters, digital images, and corresponding surrounding rock grades of the tunnel face; S2. Constructing a refined classification feature system for drilling parameters; specifically including the following: S21. Calculate five rock drillability indices: equivalent propulsion force, equivalent impact force, difficulty-to-drill index, E index, and equivalent feed rate based on four original drilling parameters. S22. Taking the tunnel face excavation cycle as a unit, six statistical characteristics of four original drilling parameters and five rock drillability indices are collected, including the mean, first quartile, second quartile, third quartile, standard deviation, and coefficient of variation; S23, taking the tunnel face excavation cycle as a unit, generating a 9×2×3 three-dimensional feature vector of the while-drilling parameters according to the original while-drilling parameters, rock drillability index, and statistical characteristics; S3. Constructing a surrounding rock intelligent classification model based on multimodal data fusion; specifically including the following: S31. Perform feature extraction through a dual convolutional neural network and obtain the classification probability of each modality; S32. Establishing a weight adaptive adjustment mechanism for the digital image modality of the face and the drilling parameter modality to achieve automatic training of the multimodal model; S4. Divide the rock mass in front of the tunnel face into several three-dimensional spatial units based on the tunnel contour and blasthole distribution characteristics. The details include the following: S41. Divide the tunnel face into 18 horizontal blocks along the width and height of the tunnel in the two-dimensional transverse direction according to the tunnel contour and blasthole distribution. S42, further segmenting the tunnel face area along the tunneling direction to ultimately form a number of three-dimensional space units; S5. Based on the intelligent surrounding rock classification model, the surrounding rock grade of the tunnel face is refined and intelligently identified. Specifically, this includes: constructing a three-dimensional feature vector of the drilling parameters of each unit body, splitting the digital image of the tunnel face according to the same horizontal block method, and inputting the three-dimensional feature vector of the drilling parameters of each unit body and the corresponding horizontally divided digital image of the tunnel face into the tunnel surrounding rock multimodal intelligent classification model to obtain the surrounding rock grade of any unit body, thereby realizing refined intelligent identification of the tunnel face surrounding rock grade.

2. The multimodal intelligent fine classification method for tunnel surrounding rock according to claim 1 is characterized by: The original drilling parameters of the tunnel face in the sample library in step S1 include feed speed, impact pressure, thrust pressure, and rotary pressure automatically collected by the intelligent rock drilling rig; digital images of the tunnel face are collected by the rig's onboard industrial camera; and the tunnel face surrounding rock grade is determined by a standard method, including Grade II, Grade III, Grade IV, and Grade V.

3. The multimodal intelligent fine classification method for tunnel surrounding rock according to claim 1 is characterized by: In step S21, the five rock drillability indicators and the propulsion force F f , impact force F h and rotational torque M r The calculation formula is: Γ ff =F f / (M r / D) Γ fh =F h / (M r / D) Γ v =V p / (V r D) C hard =C ff C fh / C v Where: Γ ff is the equivalent propulsion force; F f is the propulsion force; D is the drill diameter; M r is the rotational torque; Γ fh is the equivalent impact force; F h is the impact force; Γ v is the equivalent feed rate; V p is the feed rate; V r is the drilling speed of the drilling tool; Γ hard is the difficulty to drill indicator; E is the E indicator; D f is the diameter of the rear end of the piston; P f is the propulsion pressure; D hB is the diameter of the rear end of the piston; m h is the piston mass; P h is the impact pressure; S h is the impact stroke; t h is the time of impact on the fiber tail; P r is the rotation pressure; q r is the motor displacement; i r is the reduction ratio; In step S22, the quartiles are obtained by sorting the data from small to large and then dividing the sorted data into four equal parts, where the first quartile Q1 corresponds to the value at the 25th percentile position, the second quartile Q2 is the median, corresponding to the value at the 50th percentile position, and the third quartile Q3 corresponds to the value at the 75th percentile position. The mean, standard deviation, and coefficient of variation are calculated as follows: Where: u c is the mean of the data points; c i is a single data point; n c is the total number of data points; σ c is the standard deviation of the data points; c v,c is the coefficient of variation of the data points; In step S23, the original drilling parameters and rock drillability index, a total of 9 features, are first combined with the first quartile and the second quartile to form the first two-dimensional feature vector, then combined with the third quartile and the standard deviation to form the second two-dimensional feature vector, and finally combined with the mean and the coefficient of variation to form the third two-dimensional feature vector. These three two-dimensional feature vectors are then stacked vertically on a two-dimensional plane to construct a 9×2×3 three-dimensional feature vector of the drilling parameters.

4. The multi-modal intelligent fine classification method for tunnel surrounding rock according to claim 1 is characterized by: In step S31, a dual convolutional neural network algorithm is used, with the digital image of the tunnel face and the three-dimensional feature vector of the drilling parameters as input. The network of each mode is set with a convolution layer and a pooling layer, a linear transformation is performed in the fully connected layer, and a softmax function is used to calculate the classification probability of the four surrounding rock grades for each mode; In step S32, a weight adaptive adjustment mechanism for the tunnel face digital image modality and the drilling parameter modality is established through the cross entropy loss function, that is, the mapping relationship between the graded probability of different modalities and the cross entropy function: Where: L is the loss rate; y i is the actual label; p i is the model prediction probability; K is the total number of label categories; i is the index variable, from 1 to K; weight c is the weight of the c-mode; L c is the loss rate of mode c; Based on the weight of each mode, the weighted classification probabilities of the four surrounding rock grades are calculated respectively. The weighted classification probabilities of the same surrounding rock grade under two modes are added together to obtain the classification probabilities of the four surrounding rock grades of the tunnel surrounding rock multimodal intelligent classification model. The total loss rate of the tunnel surrounding rock multimodal intelligent classification model is calculated using the cross-entropy loss function, and the loss rate is distributed according to the weight of each mode. Backpropagation is performed to update the model parameters.

5. The multi-modal intelligent fine classification method for tunnel surrounding rock according to claim 1 is characterized by: The step S41 specifically includes: taking the center line of the tunnel and the bottom of the pit as the reference, first drawing the horizontal line and the vertical line, and then offsetting them by 2 to 4 meters in the vertical and left and right directions respectively, with the maximum size of a single block set to 2 to 4 meters, until the outline boundary of the tunnel is covered; finally, the entire tunnel outline is divided into 18 blocks, numbered from K-1 to K-18, with an area of ​​3m 2 Up to 8m 2 between.

6. The multi-modal intelligent fine classification method for tunnel surrounding rock according to claim 1 is characterized by: The step S42 specifically includes setting the spacing of the longitudinal segments to 0.4 m, dividing the tunnel face into blocks in the width and height directions, combining the longitudinal partitions, and finally extending these two-dimensional partitions along the depth in front of the tunnel face to form a plurality of three-dimensional spatial units.

7. A system according to the multi-modal intelligent fine classification method for tunnel surrounding rock according to any one of claims 1 to 6, characterized in that: Includes the following modules: Sample library construction module (110): constructing a sample library, the sample library including the drilling parameters, digital images and corresponding surrounding rock levels of the tunnel face; Feature construction module (120): constructing a refined classification feature system for drilling parameters; Surrounding rock intelligent classification model construction module (130): constructing a surrounding rock intelligent classification model based on multimodal data fusion; The rock mass division module (140) in front of the tunnel face is used to divide the rock mass in front of the tunnel face into a number of three-dimensional space units according to the tunnel contour and the blasthole distribution characteristics; Intelligent and refined identification module for tunnel face surrounding rock grade (150): Based on the surrounding rock intelligent classification model, the tunnel face surrounding rock grade is intelligently identified.

Images