CN-120913053-B - Dual-mode three-layer multi-scale multi-task fusion intelligent judgment method and system for tunnel face geological features

CN120913053BCN 120913053 BCN120913053 BCN 120913053BCN-120913053-B

Abstract

The invention belongs to the field of tunnel engineering, and particularly discloses a bimodal three-layer multi-scale multi-task fusion intelligent judgment method and system for geological features of tunnel face, wherein the method comprises the steps of collecting images of the face after tunnel field blasting is completed; the method comprises the steps of collecting and arranging construction site data, integrating and cleaning original while-drilling parameters and geological sketch information of corresponding face, combining the face images and geological feature information extracted from geological sketch to construct a multi-mode data set, constructing a surrounding rock level intelligent judgment model based on multi-mode data fusion, and carrying out intelligent judgment. According to the invention, through fusing the multi-dimensional information of the face image and the while-drilling parameters, intelligent identification of the geological features of the face is realized, the accuracy and the robustness of the identification of the geological features are remarkably improved, the stratum judgment and the blasting parameter adjustment in the construction process are optimized, and the intelligent development of tunnel construction is effectively promoted.

Inventors

TONG JIANJUN
MIAO XINGWANG
ZHANG YE
YE PEI
WANG MINGNIAN
XIANG LULU
HUANG HE

Assignees

西南交通大学

Dates

Publication Date: 20260508
Application Date: 20250609

Claims (8)

1. A bimodal three-layer multi-scale multi-task fusion intelligent judgment method for geological features of tunnel face is characterized by comprising the following steps: S1, acquiring images of a tunnel face after tunnel field blasting is completed; S2, collecting and sorting construction site data, including original while-drilling parameters and geological sketch information of a corresponding face; S3, integrating and cleaning original while-drilling parameters, and constructing a multi-mode data set by combining the tunnel face image and geological feature information extracted from geological sketch; s4, constructing a multi-modal data fusion surrounding rock level intelligent judgment model based on the multi-modal data set, and intelligently judging the surrounding rock level of the face, wherein the surrounding rock level intelligent judgment model is a multi-modal fusion judgment model with a three-layer cross-modal attention pyramid structure, and specifically comprises the following steps: 1) The bottom layer-GeoDeformNet network is used for focusing on local features of lithology texture mutation areas and parameter mutation areas while drilling in images, fusing image pixels and the parameter while drilling of corresponding blastholes through a cross-modal attention mechanism, focusing on lithology texture mutation areas and parameter sharp mutation points, extracting irregular texture features by combining deformable convolution, and realizing fine granularity alignment by utilizing pixel and blasthole coordinate mapping; 2) The middle layer-GeoPyramidGAT network is used for extracting regional characteristics of joint distribution trend and local variation trend of parameters while drilling in an image, dividing the image into a plurality of geological structure surface regions by adopting a super-pixel segmentation algorithm, extracting texture characteristics of each region and statistics of parameters while drilling covering blastholes of each region, constructing a regional level graph structure, inputting the regional level graph structure into a graph attention network GAT, and capturing fusion characteristics of middle layer scales by using the regional characteristics; 3) The top-GeoAugmented Vision Mamba network, namely the GAV-Mamba network, is used for capturing the overall stability characteristics of the image and the overall distribution characteristics of main variables in the parameters while drilling, improving the structure on the basis of a vision mamba model, mapping the parameters while drilling into a two-dimensional matrix consistent with the image size, fusing the two-dimensional matrix with the image characteristic diagram, and constructing overall fusion input; The three-layer features realize dynamic weighted fusion through a gating mechanism, the weight of each scale information is dynamically adjusted according to different geological conditions, the adaptability and the robustness of the model are enhanced, and finally geological feature results are output through the full-connection layer.
2. The method for intelligently judging the geological features of the tunnel face by means of bimodal three-layer multi-scale multi-task fusion is characterized in that in the step S1, when image acquisition is carried out on the face after the field blasting of a tunnel constructed by a drilling and blasting method is completed, after ventilation is carried out for at least 30 minutes, a light scattering lamp source of 10000 Lux is respectively distributed at the position which is 10-12 m away from the face and is 1.5m away from the center line of the tunnel, image acquisition equipment with pixels not lower than 2000 ten thousand is used for shooting, multiple image acquisition is carried out on the same face, and finally one piece with the highest image quality is selected for storage and is used as final image input.
3. The intelligent judging and identifying method for the bimodal three-layer multi-scale multi-task fusion of the geological features of the tunnel face is characterized in that in the step S2, the original while-drilling parameters comprise indexes of propelling pressure, impact pressure, turning moment and feeding speed, the indexes are used for reflecting hardness and crushing degree of surrounding rock, and the geological sketch information comprises information of lithology category, joint crack development condition and groundwater condition and is used as a source of a subsequent judging and identifying label.
4. The method for intelligent judgment by double-mode three-layer multi-scale multi-task fusion of tunnel face geological features according to claim 1 is characterized in that in step S3, original while-drilling parameters automatically collected by an intelligent drilling jumbo are integrated, coordinates, feeding speed, impact pressure, propelling pressure and revolving pressure of each blast hole are extracted, and the parameters are unified and standardized to be of fixed length; meanwhile, extracting mileage information and geological features corresponding to the face from geological sketch; And constructing a multi-mode data set consisting of the face image, the while-drilling parameters and surrounding rock grading information by taking mileage and blasthole numbers as index labels.
5. The bimodal three-layer multi-scale multi-task fusion intelligent judgment method of tunnel face geological features according to claim 1, wherein the GeoDeformNet network comprises: (1) Focusing on abrupt change areas of parameters while drilling between adjacent blastholes by crossing the attention layer; firstly, selecting a blast hole similar to the current blast hole by calculating the Euclidean geometric distance ; Wherein the method comprises the steps of Represent the first Drilling and the first Euclidean distance of the individual drill holes, Representing two different borehole coordinates; Then calculating the statistical characteristics of the selected blast hole and the current blast hole, and specifically calculating the mean value difference, the standard deviation and the maximum absolute value difference, wherein the calculation formula is as follows: ; ; ; Wherein the method comprises the steps of Represent the first The mean value difference, standard deviation and maximum absolute value difference of the drill holes are calculated, and k represents the number of the drill holes involved in calculation; Representing an original parameter matrix while drilling; the final output characteristics are: ; Wherein the method comprises the steps of Represent the first Coded borehole characteristics of individual boreholes, Sequentially represent the first The original parameter characteristics, the mean value difference, the standard deviation and the maximum absolute value difference of each drilling hole, Representing vector concatenation, R representing a real number domain, and d representing a dimension; (2) The deformable convolution layer is used for extracting local features of lithology texture mutation in the image; wherein the deformable convolution layer is specifically defined as follows: ; ; ; ; ; Wherein the method comprises the steps of Representing the image of the face being input, Representing an activated feature extracted from the convolution layer of the original image, Representing a basis A learnable offset is defined, for constructing the deformable layer, Representing an activated feature extracted from a deformable convolution layer of the input feature, Representing a feature derived from the input feature via maximum pooling, Is a feature derived from the input feature via ReLU activation; (3) The regional attention mechanism is used for realizing the fine correspondence of the image characteristics and the blasthole drilling parameters according to the pixel coordinates, and firstly, the drilling coordinates are required to be normalized, and then, the linear mapping is carried out to project the drilling coordinates to a unified dimension for alignment: ; ; Wherein, the Representing coordinates normalized to [ -1, 1], The size after the image processing is represented, Representing a sampling layer for extracting the characteristics of the image at the corresponding position according to the coordinates; the extracted while-drilling parameters and the image features are aligned, so that subsequent fusion is facilitated, and the specific alignment steps are as follows: ; ; Wherein the method comprises the steps of Representing the entered borehole coordinate matrix, Representing the original while-drilling parameters of the input, Representing the aligned while drilling parameter characteristics, Representing the characteristics of the while-drilling parameters after processing across the attention layer, Is a self-defined cross-attention layer used for extracting the differentiation characteristics of adjacent blastholes while drilling, Representing a trainable projection weight matrix, Representing a bias matrix; finally, the aligned while-drilling features and the image features are fused through a multi-layer perceptron, ; Wherein the method comprises the steps of As the feature matrix after the fusion, Representing pre-fusion image features and while drilling features, respectively.
6. The method for intelligent judgment and identification by bimodal three-layer multi-scale multi-task fusion of geological features of tunnel face according to claim 1, wherein GeoPyramidGAT network performs structural face preliminary division on the face image based on a super-pixel segmentation method, and records a pixel coordinate set of each region: ; ; ; Wherein the method comprises the steps of Representing an original image by a superpixel segmentation algorithm Divided into The block is provided with a plurality of channels, For recording the edge pixel coordinates of each image block, For mapping the recorded edge coordinates with the borehole locations to demarcate the area of the borehole, Representing the pixel abscissa and ordinate of each point on the super pixel block; And calculating the statistical characteristics of the blasthole while drilling parameters in the area, wherein the statistical characteristics comprise mean value, standard deviation, minimum value and maximum value: ; ; ; ; Wherein the method comprises the steps of Mean, standard deviation, minimum, maximum values representing the while-drilling parameters in the respective region, Representing the while-drilling parameters of the ith blast hole in the area; Then, the image block needs to be subjected to feature extraction, which is specifically performed as follows: ; Wherein the method comprises the steps of Representing pairs of image blocks Pyramid pooling is carried out to convert the image feature pyramid into a three-layer image feature pyramid; Finally, constructing a graph structure, wherein each node in the graph represents an image area, and the node characteristics comprise image area characteristics and parameter statistic information while drilling, and the specific method is as follows: ; ; ; ; Wherein the method comprises the steps of Representing the statistics while drilling of the current region obtained through the calculation, Features representing the image blocks of the current region obtained through the above calculation, Representing the spliced statistical diagnosis while drilling and the image characteristics, Representing the pixel coordinate locations of the nodes of the map calculated, Representing the set of edges between the constructed nodes, Two different graph nodes are represented, The method comprises the steps of representing the number of edges which should be adjacent to a current node according to a K neighbor algorithm, wherein K is a super parameter of the K neighbor algorithm; and finally, taking the constructed graph structure as input to be sent to a graph neural network, and learning fusion characteristics at the region level.
7. The method for intelligent judgment by bimodal three-layer multi-scale multi-task fusion of tunnel face geological features according to claim 1, wherein the GAV-Mamba network maps pixel coordinates of an image with while-drilling parameters at corresponding positions one by one to form the while-drilling parameters into a two-dimensional structure consistent with the image size, and specifically comprises the following steps: Defining network coordinates of an image: ; Wherein the method comprises the steps of The width of the face picture is represented, Representing the height of the face picture; For each channel where the parameter while drilling is located, interpolating sparse parameter values on a two-dimensional image domain: ; Wherein the method comprises the steps of Representing the two-dimensional interpolation method adopted; finally splicing the normalized image and the interpolation parameter graph along the channel dimension, ; Representing the image of the face being input, And obtaining tensor F for extracting global semantic features and performing model training.
8. A system of a bimodal three-layer multi-scale multi-task fusion intelligent judgment method according to the geological features of the tunnel face of any one of claims 1-7, characterized by comprising the following modules: The image acquisition module (110) is used for acquiring images of the tunnel face after the tunnel site blasting is completed; the data collection module (120) collects and collates construction site data, including original while-drilling parameters and geological sketch information of the corresponding face; The multi-mode data set construction module (130) is used for integrating and cleaning original while-drilling parameters and constructing a multi-mode data set by combining the face image and the geological feature information extracted from geological sketch; The intelligent judging module (140) is used for constructing a surrounding rock grade intelligent judging model based on the multi-mode data set and carrying out intelligent judging on the surrounding rock grade of the face.

Description

Dual-mode three-layer multi-scale multi-task fusion intelligent judgment method and system for tunnel face geological features Technical Field The invention relates to the technical field of tunnel engineering, in particular to a bimodal three-layer multi-scale multi-task fusion intelligent judgment method and system for geological features of tunnel face. Background The drilling and blasting method is one of the main methods of the current rock tunnel construction, and the construction quality and the safety of the method are highly dependent on accurate judgment of stratum conditions. In the construction process, the face is the area which most directly reflects lithology characteristics and blasting effects, and is traditionally observed and analyzed mainly by means of manual experience. However, this approach is greatly affected by subjective factors, and it is difficult to achieve quantitative evaluation. In recent years, with the development of computer vision technology, intelligent analysis of a face using image processing and deep learning methods has become possible. Through the technology of semantic segmentation, target detection, texture analysis and the like, the rock information of the face, joint fracture characteristics and blasting effect can be extracted, and data support is provided for construction optimization. However, there is a limitation in relying on image analysis, such as illumination conditions, rock surface pollution, etc., which may affect the recognition accuracy, so that it is necessary to combine other data sources to perform multi-mode information fusion, so as to improve the accuracy and stability of stratum recognition. The while-drilling parameter is another important stratum characterization data in the construction process of the drilling and blasting method, and mainly comprises drilling speed, torque, propelling force, impact frequency and the like. Formations of different lithology exhibit different parametric characteristics during drilling, e.g., a fast soft rock drilling rate but a low torque, and a slow hard rock drilling rate but a high torque. Based on the parameters, researchers classify the stratum by adopting statistical analysis, machine learning and deep learning methods, and a certain recognition effect is obtained. However, the while-drilling parameters can only reflect the stress condition of the drilling machine in the drilling process, and it is difficult to directly provide the spatial distribution information or structural characteristics of the stratum. Therefore, the analysis of a single data source has a limitation, and how to effectively integrate the face image and the while-drilling parameters so as to realize more accurate stratum identification and construction optimization becomes the key direction of the current research. Multimodal learning provides a new technical approach to solve this problem. By fusing the face image and the while-drilling parameters, the complementary characteristics of visual information and drilling data can be fully utilized, and the accuracy of stratum identification is improved. The currently mainstream multi-modal fusion method comprises feature level fusion, decision level fusion and depth fusion, wherein the depth fusion usually adopts a multi-modal neural network, such as image feature extraction by combining with CNN, LSTM or transducer processes time series data, and captures the association of the two through a self-attention mechanism. The fusion mode not only can improve lithology recognition precision, but also can be further used for blasting parameter optimization, construction quality evaluation and intelligent decision. In the future, technologies such as 3D visual reconstruction, real-time monitoring and automatic control are combined, and the tunnel construction of the drilling and blasting method is promoted to develop towards a more intelligent and more efficient direction. Disclosure of Invention The invention aims to solve the problems of single data utilization, insufficient information fusion, low automation degree and the like in the prior tunnel surrounding rock identification technology, and provides a bimodal three-layer multi-scale multi-task fusion intelligent judgment method and system for tunnel face geological features, which comprehensively utilize spatial texture information of images and physical feedback information of drilling processes, the depth fusion model with local, regional and global multi-scale perceptibility is constructed, high-precision intelligent identification of geological features of tunnel face is realized, multiple application requirements of surrounding rock classification identification, blasting parameter optimization, construction risk early warning and the like are considered, and the problems mentioned in the background art are solved. In order to achieve the purpose, the invention provides the following technical scheme that the bimodal three-layer