Multi-scale identification method and electronic device for rock mass of underground engineering long tunnel face

Abstract

This invention discloses a multi-scale identification method and electronic equipment for lithology at the working face of long tunnels in underground engineering, belonging to the field of lithology identification technology in underground engineering. This method addresses the problems of existing identification methods, such as single-scale operation, insufficient spatial continuity, and low predictive reliability. The key technical points are: acquiring multi-scale image data of rock cuttings, rock blocks, and the working face, and performing standardized preprocessing to construct a five-level cross-scale association chain of "interval—working face—segment—rock block—rock cuttings"; calculating the probability of pure lithology in rock cuttings and the probability of the main lithology at the rock block scale, thereby constraining the division of lithological regions at the working face and forming a standardized lithology database; and constructing and training a Transformer model based on this database to achieve interval-scale lithology prediction. This invention improves the accuracy and stability of lithology identification and prediction, and is mainly used for advanced geological forecasting and construction auxiliary decision-making in long tunnel construction.

Description

Technical Field

[0001] This application relates to the fields of underground engineering and intelligent geological identification technology, and in particular to a multi-scale identification method and electronic equipment for the lithology of the working face of long underground tunnels. Background Technology

[0002] Long underground tunnels are widely used in water conservancy, hydropower, transportation, and energy projects. Accurate acquisition of lithological information at the tunnel face during construction is crucial for surrounding rock stability analysis, support parameter design, and construction safety. Due to the tunnel's great depth and complex geological conditions, the lithology exhibits significant spatial heterogeneity and scale differences during construction, increasing the difficulty of lithological identification.

[0003] In the process of realizing this invention, the inventors discovered at least the following problems in the prior art: In existing technologies, lithology identification at the tunnel face largely relies on manual experience or single-scale image recognition methods, such as analyzing overall images of the tunnel face or local rock blocks and cuttings. While these methods can reflect local lithological characteristics to some extent, the lack of effective correlation between lithological information at different scales makes it difficult to achieve unified constraints and fusion of multi-scale lithological information, including rock cuttings, rock blocks, and the tunnel face. This results in insufficient stability and reliability of lithology identification results, failing to meet the needs for continuous lithology identification and prediction in long tunnel construction. Summary of the Invention

[0004] In view of this, embodiments of this application provide a multi-scale identification method and electronic device for lithology at the working face of long underground tunnels to solve the problems existing in the background art.

[0005] According to a first aspect of the embodiments of this application, a multi-scale identification method for the lithology of the working face of a long underground tunnel is provided, comprising: Acquire multi-scale lithology-related image data, which includes rock cuttings images, rock block images, and tunnel face images; The rock cuttings images, rock block images, and tunnel face images are preprocessed respectively, and the tunnel face is divided into intervals and segments along the tunnel axis. Then, cross-scale correlation is performed to obtain a five-level correlation chain of "interval-tunnel face-segment-rock block-rock cuttings". Based on the preprocessed rock cutting images, the probability of pure lithology at the rock cutting scale is calculated to achieve microscopic calibration; Based on the preprocessed rock block image, combined with the fifth-level correlation chain and the pure lithology probability at the rock fragment scale, the main lithology probability at the rock block scale is calculated to achieve mesoscopic constraints. Based on the preprocessed face image and combined with the main lithology probability at the block scale, the face-scale lithology region is divided to achieve macroscopic division; Based on the division results of the lithological regions at the working face scale, a standardized database is constructed; Based on the standardized database, a Transformer model is constructed and trained to perform interval-scale lithology prediction. The tunnel face image corresponding to the interval to be predicted and its associated lithological feature data are input into the trained Transformer model to perform interval-scale lithology prediction and obtain the prediction result. The lithological feature data consists of the pure lithology probability at the rock debris scale, the main lithology probability at the rock block scale, and the result of dividing the lithological region at the tunnel face scale.

[0006] Optionally, the method also includes acquiring prior geological data and using the prior geological data for the cross-scale correlation or interval-scale lithology prediction, wherein the prior geological data includes excavation direction, structural plane angle, and borehole core data.

[0007] Optionally, the rock cuttings images, rock block images, and tunnel face images are preprocessed respectively, and the tunnel face is divided into intervals and segments along the tunnel axis. Then, cross-scale correlation is performed to obtain a five-level correlation chain of "interval-tunnel face-segment-rock block-rock cuttings", including: The rock cuttings image, rock block image and tunnel face image are uniformly cropped, denoised and enhanced with lithological features to obtain lithological related feature parameters and intermediate results. Based on the spatial position corresponding to the preprocessed tunnel face image, multiple consecutive tunnel faces are divided into intervals along the tunnel axis, and the tunnel face is segmented within each interval. The lithology-related characteristic parameters and intermediate results are normalized. Based on the normalized lithological characteristic parameters and intermediate results, a cross-scale association table in CSV format is generated to describe the correspondence between intervals, working faces, segments, rock blocks and rock cuttings, thereby establishing a five-level association chain of "interval-working face-segment-rock block-rock cuttings".

[0008] Optionally, based on the preprocessed rock cuttings image, the probability of pure lithology at the rock cuttings scale is calculated, including: Feature extraction is performed on the preprocessed rock debris image to generate candidate regions, and rock debris bounding boxes and pixel-level mask images are output in parallel. The pixel-level mask images correspond to clean rock debris regions. The pixel-level masked image is divided into multiple image blocks, which are then converted into feature vectors and input into a Transformer encoder. The self-attention mechanism is used to capture the microscopic features of minerals, and the transfer learning parameters are adjusted based on the pre-trained weights, which are obtained from the ImageNet dataset. The output is a probability distribution of lithology categories. Calculate the probability of pure lithology at the lithological scale based on the probability distribution of the lithological categories.

[0009] Optionally, based on the preprocessed rock block image, and combining the fifth-order correlation chain and the pure lithology probability at the rock fragment scale, the main lithology probability at the rock block scale is calculated, including: Macroscopic structural features of the rock blocks are extracted from the preprocessed rock block images, and the initial probability distribution of lithology categories is output, represented in the form of a probability vector. Based on the five-level association chain and combined with the pure lithology probability at the rock fragment scale, the initial probability distribution of the lithology category is calibrated to obtain the main lithology probability at the rock block scale, which is represented in the form of a probability vector.

[0010] Optionally, based on the preprocessed face image and combined with the main lithology probability at the block scale, the face-scale lithological region is divided, including: Different receptive field features are extracted from the preprocessed face image. The shallow edge features and deep semantic features are fused by the decoder skip connection to output a pixel-level lithology probability distribution matrix. The probability vector corresponding to the main lithology probability at the block scale is multiplied element-wise with the pixel-level lithology probability distribution matrix. Lithology probabilities less than a set threshold are set to zero, and the effective lithology distribution is retained to obtain the probability of each pixel belonging to each lithology category after constraints. The lithology corresponding to the maximum probability of each pixel is selected as the lithology category of that pixel, and lithological regions are formed based on pixel connectivity. The output includes the face association ID, lithology probability distribution matrix, lithology region division color heat map, and lithology statistical results for each segment, which serve as the lithology region division results at the face scale.

[0011] Optionally, based on the lithological zone division results at the working face scale, a standardized database is constructed, including: The lithological probability distribution matrix, cross-scale correlation table, and geological prior data of the working face of each blasting cycle are indexed by combining timestamps and spatial coordinates and stored in a standardized database, which includes a historical sequence table, a geological parameter table, and a cross-scale correlation table. The data in the standardized database is processed to remove invalid data and fill in data breaks to ensure the continuity of the database in time and space, and to support retrieval by interval identifier and cyclic number; After the completion process is completed, the data in the standardized database is periodically verified, and the anomaly identification results are corrected.

[0012] Optionally, the Transformer model adopts an Encoder-Decoder architecture; The encoder contains multiple encoder layers. It uses a self-attention mechanism to calculate the correlation weights of each face and region in the historical sequence to capture the spatiotemporal evolution trend of lithology. A cross-attention mechanism is used to fuse prior geological data. The Decoder comprises multiple DecoderLayers. Based on the output features of the Encoder, it generates the lithology probability distribution matrix for the next blasting cycle through a masked self-attention mechanism; and calculates the reliable probability of lithology prediction during model inference.

[0013] According to a second aspect of the embodiments of this application, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0014] According to a third aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0015] The technical solutions provided by the embodiments of this application may include the following beneficial effects: As can be seen from the above embodiments, this application constructs a multi-scale collaborative identification and prediction system of "rock cuttings - rock blocks - working face - interval", which unifies the modeling and correlation analysis of the originally fragmented multi-source data. It forms a complete technical link from data collection, feature extraction, cross-scale fusion to interval prediction, effectively solving the technical problems of insufficient utilization of multi-scale information, unstable identification results and lack of reliable basis for interval scale prediction in the existing technology.

[0016] Specifically, at the microscopic level, pure lithology identification is performed on rock fragment images to obtain high-purity lithological feature information, thereby providing a reliable calibration benchmark for subsequent identification. At the mesoscopic level, a five-level correlation chain is introduced to map the pure lithology probability at the rock fragment scale to the corresponding rock block, thereby constraining and correcting the identification results at the rock block scale and reducing the impact of mixed lithology and surface interference on the identification results. At the macroscopic level, the lithological region of the tunnel face is divided by fusing the main lithology probability at the rock block scale, so that the identification results have both spatial continuity and structural rationality.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] Figure 1 This is a flowchart illustrating a multi-scale identification method for the lithology of the working face of a long underground tunnel, according to an exemplary embodiment.

[0020] Figure 2 This is a topology map illustrating the practical application of a multi-scale identification method for lithology at the working face of a long underground tunnel, according to an exemplary embodiment.

[0021] Figure 3 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0025] Explanation of terms: Debris Scale: A fine-scale identification method using millimeter- to centimeter-level rock fragments. Its core function is to determine the purity of lithology by extracting essential characteristics such as mineral composition and micro-texture of rock fragments, providing a benchmark calibration for rock block scale identification.

[0026] Rock Block Scale: A mesoscale that identifies intact rock blocks in the centimeter to decimeter range. It focuses on the probabilistic determination of the main lithology of the rock block and outputs an effective lithology category vector to constrain the lithology identification range at the working face scale.

[0027] Tunnel Face Scale: A macroscopic scale that uses the tunnel face of a meter-long tunnel as the identification object. It uses a semantic segmentation algorithm to achieve pixel-level division of lithological spatial distribution. Its continuous identification results are used to build a basic database for interval scale prediction.

[0028] Section Scale: A macroscopic continuous scale that takes the sequence of continuous working faces ranging from hundreds to thousands of meters as the research object. The core is based on the Transformer model to capture the spatiotemporal evolution trend of lithology and realize the prediction of lithology distribution of the working face in the next blasting cycle.

[0029] Multi-scale Recognition: A collaborative recognition system integrating four scales: rock cuttings, rock blocks, tunnel faces, and intervals. It improves the accuracy of lithology identification and prediction through hierarchical services and cross-scale information closed-loop integration, which includes "microscopic calibration of mesoscopic, mesoscopic constraint of macroscopic, and macroscopic support of interval prediction".

[0030] Pure lithology: A single lithology that does not contain interlayers or impurities (such as soil or broken particles) of other lithologies, and whose mineral composition and structure conform to the essential characteristics of that lithology. The determination result provides a benchmark anchor for subsequent scale identification.

[0031] Image classification algorithm: A machine learning algorithm (such as ResNet152, Vision Transformer (ViT)) that determines the target category by extracting image features. In this invention, it is used to determine the pure lithology at the rock debris scale and the probability identification of the main lithology at the rock block scale.

[0032] Semantic Segmentation Algorithm: A machine learning algorithm (such as DeepLab V3+, UNet++) that divides the spatial distribution of targets through pixel-level classification. In this invention, it is used for lithological region division at the face of the tunnel and outputs a lithological probability distribution matrix.

[0033] Transformer Model: A deep learning model based on self-attention mechanism, which has the ability to capture long-distance spatiotemporal correlations. In this invention, it is used for interval-scale prediction. It takes a continuous working face lithology distribution sequence database as input and outputs the lithology distribution and reliability probability of the next cycle.

[0034] Cross-scale Correlation: Establish a five-level correlation system of "interval ID - face ID - segment ID - rock block ID - cuttings ID" to achieve spatiotemporal alignment and bidirectional information transmission of data at different scales, providing support for multi-scale collaborative identification.

[0035] Lithology Probability Distribution: At a certain scale (such as the working face), the set of probability values ​​for each spatial location belonging to a certain type of lithology is usually presented in matrix form, which intuitively reflects the spatial distribution characteristics of lithology.

[0036] Reliability Probability: The quantitative value of the predicted reliability of each lithological distribution area in the interval-scale prediction results. It is calculated based on the accuracy of the preceding multi-scale identification and the spatiotemporal correlation strength, and is used to guide the priority of construction decisions.

[0037] Spatio-temporal alignment: Using the completion time of the blasting cycle as the time reference and the tunnel construction coordinate system as the spatial reference, the timestamps and spatial coordinates of data at all scales are unified to ensure the accuracy of cross-scale information transmission.

[0038] Transfer learning: Transferring the weight parameters of a pre-trained model (such as ResNet or ViT models trained on the ImageNet dataset) to the target task, reducing the number of samples required for training the target task and accelerating the model's convergence speed.

[0039] Figure 1 This is a flowchart illustrating a multi-scale identification method for lithology at the working face of a long underground tunnel, according to an exemplary embodiment. Figure 2 This is a topology map illustrating the practical application of a multi-scale identification method for lithology at the working face of a long underground tunnel, according to an exemplary embodiment. Figure 1 and Figure 2 As shown, the method may include the following steps: S1: Acquire multi-scale lithology-related image data, which includes rock cuttings images, rock block images, and tunnel face images; Specifically, during tunnel construction, blasting cycles serve as the basic data acquisition unit. Construction personnel use data acquisition terminals to simultaneously collect lithological information at different scales and record the corresponding time and spatial location information of the images to ensure the correspondence between multi-scale data within the same construction cycle. The data acquisition terminal can be a smartphone, digital camera, or other device with image acquisition capabilities, and it also has data storage and transmission functions.

[0040] During data acquisition, rock cuttings, rock blocks, and tunnel faces were collected separately to obtain image data reflecting lithological characteristics at different scales. Specifically, rock cuttings images were used to characterize mineral composition and microstructure, rock block images were used to characterize macroscopic structure and assemblage, and tunnel face images were used to characterize overall spatial distribution. Simultaneously, corresponding association information was uniformly recorded for each type of image to facilitate subsequent cross-scale correlation processing.

[0041] Preferably, during the acquisition process, data at each scale are categorized and managed according to the blasting cycle number, and calibrated using timestamps and spatial coordinates to ensure consistency of data at different scales in time and space, thereby providing basic data support for subsequent multi-scale fusion analysis.

[0042] In one embodiment, the data acquisition specifications for each scale are as follows: Rock cuttings scale: Select 3-5 rock cuttings with a particle size of 5-20mm and no obvious impurities, and cover the five areas of the working face: left / middle / right / top / bottom; take pictures using the macro mode of a mobile phone, with a resolution of ≥1000×1000 pixels, and turn on the fill light (5-10cm away from the rock cuttings) to avoid shadows obscuring mineral features. Rock block size: Collect 1-2 complete rock blocks (grain size 20-50cm, surface free of cement / soil) from each of the 5 areas of the working face. Take front and side images with a camera, resolution ≥2000×1500 pixels. The front is perpendicular to the rock block surface, and the side shows the structural features (such as foliation, granular structure). Tunnel face dimensions: Small cross-section tunnel (≤10) Divide into 3 segments according to a 1×3 grid, large cross-section (>10) Divide the tunnel into 6 segments using a 2×3 grid, marking the segment lines with red paint; use a wide-angle camera mode for shooting, with a resolution of ≥4000×3000 pixels; place two 50W supplementary lights at a 45° angle on both sides of the tunnel face (distance 1.5-2m); first take a global image (including segment lines), then take segment images, with the tunnel face accounting for ≥80% of the total area. Interval scale: Record the excavation direction (azimuth) and structural plane angle (error ≤ 2°) for each blasting cycle. Summarize the borehole core data every 5 cycles. Use "blasting completion time" as the timestamp (accurate to the second). Use the tunnel construction coordinate system (starting point as origin, excavation direction as X-axis) to record spatial coordinates.

[0043] Based on the above requirements, images of rock cuttings, rock blocks, and the working face were acquired respectively.

[0044] S2: The rock cuttings image, rock block image, and tunnel face image are preprocessed respectively, and the tunnel face is divided into intervals and segments along the tunnel axis. Then, cross-scale correlation is performed to obtain a five-level correlation chain of "interval-tunnel face-segment-rock block-rock cuttings". This step includes the following sub-steps: S21: Perform unified cropping, noise reduction, and lithological feature enhancement processing on the rock cuttings image, rock block image, and tunnel face image to obtain lithological related feature parameters and intermediate results. Based on the spatial position corresponding to the preprocessed tunnel face image, divide multiple consecutive tunnel faces into intervals along the tunnel axis direction, and perform segmentation processing on the tunnel face within each interval. Specifically, when preprocessing the rock cuttings images, rock block images, and tunnel face images, a unified format conversion and standardization process can be performed on images of different scales through edge computing nodes or the server side to eliminate differences caused by different acquisition devices and environmental conditions. The preprocessing process includes image size unification, noise suppression, and contrast enhancement, thereby improving the recognizability of lithological features in the images.

[0045] All rock cuttings images, rock block images, and tunnel face images were uniformly cropped to 512×512 pixels. The BM3D algorithm was used to remove dust noise from the tunnel environment. Adaptive histogram equalization was used to enhance lithological features (the rock cuttings images underwent additional threshold segmentation to screen for impurity regions), and lithological-related feature parameters and intermediate results were obtained.

[0046] After obtaining the preprocessed image data, further feature information related to lithology identification is extracted to form lithology-related feature parameters and intermediate results. The intermediate results may include image data after denoising and enhancement processing, preliminary segmentation regions, or feature response results, etc.

[0047] Based on this, using the spatial location information corresponding to the preprocessed tunnel face images, multiple consecutive tunnel faces are sorted according to the tunnel excavation direction and divided into several intervals in conjunction with blasting cycle information, so that each interval corresponds to a continuous construction area. Subsequently, within each interval, the tunnel face is segmented according to the geometric shape of the tunnel face or a preset division rule to obtain multiple segmented regions with spatial location identifiers.

[0048] S22: Normalize the lithology-related characteristic parameters and intermediate results; Specifically, when normalizing the lithology-related characteristic parameters and intermediate results, corresponding standardization methods can be adopted according to different parameter types to ensure they meet a unified data representation range. For numerical characteristic parameters, linear normalization can be used to map them to a preset interval; for image features or probability distribution data, normalization operations can be used to ensure they meet uniform scale requirements.

[0049] In one embodiment, auxiliary data used in subsequent analysis can be uniformly processed. For example, parameters such as angles and lengths in prior geological data can be uniformly converted to ensure consistent physical meaning between data from different sources. Abnormal data generated during the acquisition process, such as blurred images, abnormal exposure, or missing data, can be marked or removed to avoid interfering with subsequent identification results.

[0050] In this example, the lithological-related characteristic parameters and intermediate results are normalized to the [0,1] interval. Furthermore, the geological prior data (angle, size) can be uniformly converted to SI units (m, rad), and abnormal data (blurred images, overexposed / underexposed images) are marked as "invalid" and removed.

[0051] S23: Based on the normalized lithology-related characteristic parameters and intermediate results, generate a CSV format cross-scale association table to describe the correspondence between intervals, working faces, segments, rock blocks and rock cuttings, thereby establishing a five-level association chain of "interval-working face-segment-rock block-rock cuttings"; Specifically, after data normalization is completed, a unified data organization structure can be constructed based on the association identification information of data at each scale and the corresponding time and space attributes, which can be used to describe the correspondence between different scales.

[0052] In practice, data from intervals, working faces, segments, rock blocks, and rock cuttings can be assigned unique identifiers and matched according to their spatial affiliation within the same blasting cycle to establish a hierarchical mapping relationship from top to bottom. Specifically, intervals correspond to multiple working faces, working faces correspond to multiple segments, segments correspond to several rock blocks, and rock blocks further correspond to multiple rock cuttings data, thus forming a complete multi-scale correlation structure.

[0053] Furthermore, the aforementioned relationships can be stored in a structured data format, such as using tables or files to record the mapping relationships and attribute information between data at each level, in order to support subsequent rapid retrieval and retrieval.

[0054] In this embodiment, based on the normalized lithological related characteristic parameters and intermediate results, a CSV format association table is generated, and a five-level association chain is established: "Interval ID - Working Face ID - Segment ID - Rock Block ID - Rock Cutting ID". The fields include association ID, timestamp, spatial coordinates, data status (valid / invalid), and verification result to ensure spatiotemporal alignment of multi-scale data.

[0055] For example, a certain blasting cycle correlation chain is "QJ-202508-01 (interval) → ZF-08 (working face) → Section-02 (segment) → YK-08-04 (rock block) → YX-08-02 (rock cuttings)". Through this correlation chain, the source and correspondence of data at any scale can be traced.

[0056] S3: Based on the preprocessed rock cuttings image, calculate the probability of pure lithology at the rock cuttings scale to achieve microscopic calibration; this step includes the following sub-steps: S31: Extract features from the preprocessed rock debris image to generate candidate regions, and output the rock debris bounding box and pixel-level mask image in parallel. The pixel-level mask image corresponds to the clean rock debris region. Specifically, when extracting features from the preprocessed rock debris images, a deep learning-based instance segmentation method can be used to identify and separate the rock debris regions in the image. A feature extraction network performs multi-layer feature extraction on the input image and generates candidate regions. These candidate regions are then classified and their boundaries regressed to obtain the bounding box information of the rock debris.

[0057] Simultaneously, pixel-level mask images are generated for each candidate region to achieve fine segmentation of the rock debris area. These pixel-level mask images are used to characterize the clean areas of the rock debris, eliminating soil adhesion, broken edges, and other non-lithological interference information, thereby improving the accuracy of subsequent lithological identification.

[0058] In one embodiment, the preprocessed rock debris image is input, and features are extracted through the ResNet50 backbone network to generate candidate regions (RoI). The rock debris bounding boxes and pixel-level mask images are output in parallel. The pixel-level mask images correspond to clean rock debris areas (removing impurities such as soil and broken edges), and the annotation error is ≤1mm.

[0059] S32: Divide the pixel-level mask image into multiple image blocks, convert them into feature vectors, and input them into the Transformer encoder. Use the self-attention mechanism to capture the microscopic features of minerals, and adjust the transfer learning parameters based on the pre-trained weights. The pre-trained weights are obtained based on the ImageNet dataset, and the output is the probability distribution of lithology categories. Specifically, after obtaining the pixel-level mask image, it can be structured by dividing the image into multiple local regions and converting each local region into a feature vector to form input data suitable for sequence modeling.

[0060] Subsequently, the feature vectors are input into an attention-based feature extraction model. A multi-layered coding structure is used to model the input features, capturing the fine details of mineral composition and texture in rock fragments. Through a self-attention mechanism, relationships between different local regions can be established, thereby enhancing the ability to represent complex mineral structures.

[0061] In one embodiment, the pixel-level mask image can be divided into 16×16 pixel image patches, converted into feature vectors, and then input into a 12-layer Transformer encoder. The self-attention mechanism is used to capture the microscopic features of minerals (such as the luster of quartz grains and the cleavage planes of feldspar). The weights are then fine-tuned by transfer learning using the ImageNet dataset to output the probability distribution of lithology categories.

[0062] S33: Calculate the probability of pure lithology at the lithology scale based on the probability distribution of the lithology categories; Specifically, after obtaining the probability distribution of lithological categories, the purity of rock fragments can be quantitatively assessed based on the model output. This purity characterizes whether the rock fragments are primarily composed of a single lithology, thus providing a reliable calibration basis for subsequent block-scale identification.

[0063] In one embodiment, statistical analysis can be performed on the lithological probability distribution output by each attention branch, and the probability of pure lithology at the lithic scale can be calculated by combining the consistency and concentration of the output results of each branch. The calculation process comprehensively considers the maximum probability value and the dispersion of its distribution in each branch, thereby obtaining an evaluation index reflecting the purity of lithology.

[0064] The probability of pure lithology at the cuttings scale is calculated using the following formula: In the formula: The probability of rock cutting purity (value 0-1); The number of attention heads in the ViT model (take 12); Let be the probability distribution of lithology categories output by the k-th attention head; for Standard deviation; (To avoid a denominator of 0); The maximum lithological probability at the cuttings scale is the output of the k-th attention head.

[0065] S4: Based on the preprocessed rock block image, combined with the fifth-order correlation chain and the pure lithology probability at the rock fragment scale, calculate the main lithology probability at the rock block scale to achieve mesoscopic constraints; this step includes the following sub-steps: S41: Extract the macroscopic structural features of the rock blocks from the preprocessed rock block image and output the initial probability distribution of lithology categories, represented in the form of a probability vector; Specifically, when extracting macroscopic structural features from preprocessed rock block images, deep learning-based image classification algorithms can be used to analyze the overall structural features of the rock blocks. Since the rock block scale can reflect the combined structure and tectonic features of the rock mass, more comprehensive lithological information can be obtained through joint analysis of front and side images of the rock blocks.

[0066] During the feature extraction process, the front and side images of the rock block can be used as inputs, and multi-layer feature learning can be performed on them through a feature extraction network to extract the macroscopic structural feature information of the rock block, including but not limited to grain structure features, bedding or foliation direction features, and overall texture distribution features.

[0067] In one embodiment, the preprocessed front and side images of the rock block are input, and macroscopic structural features (such as granular structure and foliation orientation) can be extracted using a 152-layer ResNet152 residual network, outputting an initial probability distribution of lithology categories. This probability distribution is a normalized probability vector for each lithology category, used to represent the preliminary judgment result that the rock block belongs to each lithology category.

[0068] S42: Based on the five-level association chain and combined with the pure lithology probability at the rock fragment scale, the initial probability distribution of the lithology category is calibrated to obtain the main lithology probability at the rock block scale, which is represented in the form of a probability vector. Specifically, after obtaining the initial probability distribution of lithology categories at the rock block scale, the rock block is further associated and matched with its corresponding rock cutting data based on the five-level association chain, thereby obtaining the pure lithology probability information at the rock cutting scale within the corresponding area of ​​the rock block.

[0069] Based on the above correlation, the results of rock cuttings identification are used as the calibration basis to correct the initial probability distribution of rock blocks. Specifically, when the rock cuttings identification results indicate that a certain lithology has a high probability of purity, the probability of the corresponding lithology in the initial probability distribution of rock blocks can be enhanced; when the rock cuttings results indicate the presence of mixed lithology or low purity, the initial probability distribution of rock blocks is conservatively adjusted to avoid misjudgment.

[0070] The formula for calculating the probability of the main lithology at the block scale is as follows: ; In the formula: The calibrated rock block belongs to the lithology The probability of; The initial probability distribution for lithology categories; Pure lithology for identifying rock cuttings; Let L be the purity probability of lithology L; Lithology The probability of purity.

[0071] S5: Based on the preprocessed tunnel face image and combined with the main lithology probability at the rock block scale, the lithological region at the tunnel face scale is divided to achieve macroscopic division; this step includes the following sub-steps: S51: Extract different receptive field features from the preprocessed face image, combine the decoder with skip connections to fuse shallow edge features and deep semantic features, and output a pixel-level lithology probability distribution matrix. Specifically, when dividing the preprocessed tunnel face image into lithological regions, a semantic segmentation algorithm can be used to perform pixel-level identification of the overall lithological distribution of the tunnel face. Since the tunnel face scale can reflect the spatial distribution characteristics and structural continuity of the rock mass, it is necessary to consider local boundary information while ensuring global semantic consistency.

[0072] During the feature extraction process, the face image can be input into the feature extraction network, and lithological information under different receptive fields can be obtained through a multi-scale feature extraction structure, thereby capturing both local detail features and overall distribution features simultaneously.

[0073] In one embodiment, the preprocessed face image is input, and different receptive field features are extracted using the ASPP (multi-scale dilated convolution) module. These features are then combined with a decoder skip connection to fuse shallow edge features and deep semantic features, outputting a pixel-level lithology probability distribution matrix. ), matrix dimension is (h is the image height in pixels,) The width in pixels. (Number of lithology categories).

[0074] S52: Perform element-wise multiplication between the probability vector corresponding to the main lithology probability at the block scale and the pixel-level lithology probability distribution matrix, set the lithology probability less than a set threshold to zero, retain the effective lithology distribution, and obtain the probability of each pixel belonging to each lithology category after constraint. Specifically, after obtaining the initial lithology probability distribution at the face of the tunnel, the main lithology probability at the block scale is introduced as constraint information to correct the pixel-level lithology probability distribution.

[0075] Based on the aforementioned five-level association chain, each region in the working face is associated and matched with its corresponding rock block to obtain the lithological probability information at the rock block scale for that region. Furthermore, the rock block-scale probability vector is fused with the pixel-level lithological probability distribution, and a weighted adjustment is applied to each lithological category.

[0076] Specifically, when a certain lithology has a high probability at the block scale, the corresponding pixel-level probability is enhanced; when a certain lithology has a low probability at the block scale, the corresponding pixel-level probability is suppressed, and probabilities below a set threshold are removed to reduce interference from low-confidence classification results.

[0077] In one embodiment, the probability vector output at the rock block scale is... With the initial matrix Perform element-wise multiplication and set the result to zero. The lithological probability is used to preserve the effective lithological distribution, and the constraint formula is as follows: In the formula: For the constrained first Pixels belong to lithology The probability of.

[0078] S53: Select the lithology corresponding to the maximum probability of each pixel as the lithology category of that pixel, and form a lithology region based on pixel connectivity, and output the lithology region division result at the face of the tunnel. Specifically, after completing the pixel-level lithology probability distribution calculation with constraints, the lithology category with the highest probability is selected as the final lithology label for each pixel location, thereby obtaining the preliminary lithology classification result of the working face.

[0079] Based on this, connectivity analysis can be performed on pixels with the same lithological category based on spatial adjacency, and spatially continuous pixel regions can be aggregated to form lithological region division results with practical engineering significance.

[0080] Furthermore, a corresponding lithological probability distribution matrix can be generated based on the formed lithological regions, and the corresponding face association ID can be obtained based on the five-level association chain to establish the association relationship between the face scale and other scale data. The lithological probability distribution matrix describes the probability distribution of different spatial locations within the face belonging to different lithological categories, providing input data for subsequent interval-scale predictions.

[0081] In one embodiment, visualization processing can also be performed based on the lithology probability distribution matrix, mapping different lithology categories to different colors, thereby generating a color heat map of lithology region division to intuitively reflect the spatial distribution characteristics of different lithologies in the tunnel face.

[0082] Furthermore, statistical analysis can be performed on each lithological region to calculate the area proportion, spatial distribution range, and corresponding ratio of different lithologies within each segment, thereby obtaining the lithological statistical results for each segment. These statistical results can be used to characterize the lithological composition of different regions of the tunnel face and provide data support for the formulation of construction support schemes and subsequent interval-scale predictions.

[0083] S6: Based on the division results of the lithological regions at the working face scale, construct a standardized database; this step includes the following sub-steps: S61: The lithological probability distribution matrix, cross-scale correlation table, and geological prior data of the working face of each blasting cycle are indexed by combining timestamps and spatial coordinates, and stored in a standardized database. The standardized database includes a historical sequence table, a geological parameter table, and a cross-scale correlation table. Specifically, after completing the lithological zoning at the face of the tunnel, the lithological identification results corresponding to each blasting cycle are structured and organized, and then uniformly organized and stored with cross-scale correlation information and geological prior data.

[0084] During the data import process, a unified indexing system can be established based on time and space dimensions. The time dimension uses the completion time of the blasting cycle as an identifier, while the space dimension is located based on the tunnel construction coordinate system, thereby achieving consistent alignment of multi-source data in time and space.

[0085] In one embodiment, the lithological probability distribution results, cross-scale correlations, and geological prior parameters of the tunnel face can be standardized and encapsulated, and written into a standardized database according to a unified data format. Different types of data can be managed by dividing them into multiple data tables according to their functions, including a historical sequence table for storing continuous historical identification results, a geological parameter table for recording geological parameter information, and a cross-scale correlation table for describing multi-scale mapping relationships.

[0086] S62: Process the data in the standardized database, remove invalid data and fill in the data breaks to ensure the continuity of the database in time and space, and support retrieval by interval identifier and cyclic number; Specifically, after the data is entered into the database, quality control and continuous optimization processes are performed on the data in the standardized database.

[0087] During the data cleaning process, abnormal data can be identified and removed based on preset rules. For example, samples with severe image occlusion, low recognition confidence, or missing data can be marked as invalid data (occlusion area ratio > 30%) and removed from the training dataset to avoid interfering with subsequent model training.

[0088] During the data completion process, data breaks that exist in continuous blasting cycles can be addressed by interpolating and reconstructing missing data using time series information. In one embodiment, an interpolation method based on adjacent cycle data can be used to complete missing areas, thereby restoring the continuous variation trend of lithology distribution over time and ensuring the spatiotemporal continuity of the database (missing rate ≤5%).

[0089] In addition, to improve data retrieval efficiency, an indexing mechanism based on interval identifiers and burst cycle numbers can be established in the database, enabling the system to quickly retrieve historical data sequences within a specified interval.

[0090] S63: After completing the completion process, periodically verify the data in the standardized database and correct the anomaly identification results; Specifically, after completing the data completion process, the data in the standardized database is periodically verified and corrected to further ensure the reliability and engineering applicability of the data.

[0091] In one embodiment, a geological engineer can sample and verify the lithology identification results in the database based on actual site conditions, and analyze the identification deviations in conjunction with the excavation and exposure results. When a significant deviation is found between the identification results and the actual lithology, the corresponding data can be corrected or relabeled. Simultaneously, the verification results can be fed back to the database to update the data status identifiers and provide high-quality labeled data for subsequent model training.

[0092] S7: Based on the standardized database, construct and train a Transformer model to perform interval-scale lithology prediction; Specifically, after the standardized database is constructed, a model for lithology prediction at the interval scale can be built based on the continuous historical data in the database, and the lithology distribution of the next blasting cycle can be predicted by learning from the historical data.

[0093] During model construction, historical lithological distribution data of the tunnel face in a standardized database can be organized into sequential data in chronological order, and combined with corresponding geological prior parameters to form the input samples required for model training. These input samples can simultaneously reflect the spatial distribution characteristics of lithology and its temporal changes as construction progresses.

[0094] In this embodiment, the Transformer model adopts an Encoder-Decoder architecture; The encoder contains multiple encoder layers. It uses a self-attention mechanism to calculate the correlation weights of each face and region in the historical sequence to capture the spatiotemporal evolution trend of lithology. A cross-attention mechanism is used to fuse prior geological data. The Decoder comprises multiple DecoderLayers. Based on the output features of the Encoder, it generates the lithology probability distribution matrix for the next blasting cycle through a masked self-attention mechanism; and calculates the reliable probability of lithology prediction during model inference.

[0095] Input data: A sequence of lithological probability distribution matrices corresponding to N historical cycles ( ) and geological prior parameter sequences ( The lithology probability distribution matrix sequence is constructed from the face lithology probability distribution matrix, the main lithology probability at the block scale, and the pure lithology probability at the cuttings scale; the geological prior parameter sequence includes information such as excavation direction, structural plane angle, and borehole core data.

[0096] During model training, the above data is organized in chronological order to form a training sample sequence; in the prediction stage, the corresponding lithology probability distribution matrix for the interval to be predicted is obtained and input into the model using the same data organization format as in the training stage.

[0097] Prediction logic: ①The Encoder calculates the correlation weights of each face and region in the historical sequence through a self-attention mechanism, capturing the spatiotemporal evolution trend of lithology (e.g., "the upper left section of the first 3 cycles is all granite, with high correlation weights"). ② Cross-attention mechanism integrates geological prior parameters (e.g., when the angle between structural planes is <30°, it enhances the correlation strength in areas with irregular lithological distribution). ③The Decoder uses the output features of the Encoder as conditions to generate the lithology probability matrix for the next cycle (N+1) through masked self-attention. ; Reliability probability calculation: In the formula: Let be the reliable probability of the (i,j)th pixel on the (N+1)th cyclic face; Let (i,j) be the attention weight for the (i,j)th pixel in the k-th and (N+1)-th cycles; Let be the lithology probability of the rock block in the region corresponding to the k-th cycle; Output: Lithology probability matrix of the working face in the next blasting cycle ( ), and the reliability probability of the corresponding region. .

[0098] S8: Input the face image of the interval to be predicted and its associated lithological feature data into the trained Transformer model to perform interval-scale lithological prediction and obtain the prediction result. The lithological feature data consists of the pure lithological probability at the rock debris scale, the main lithological probability at the rock block scale, and the result of dividing the lithological region at the face of the tunnel. Specifically, after completing the Transformer model training, the face image corresponding to the prediction interval is first preprocessed, and then combined with the associated pure lithology probability at the rock debris scale, the main lithology probability at the rock block scale, and the face lithology region division results, to construct corresponding multi-scale lithological feature data, so as to realize the prediction of lithological distribution in the next blasting cycle or target interval.

[0099] In one embodiment, the above-mentioned multi-scale feature data can be uniformly encoded to map the rock cutting scale information and rock block scale information to the spatial distribution of the face, and then fused with the face lithological region division results to construct the corresponding lithological probability distribution matrix.

[0100] Subsequently, the lithology probability distribution matrix is ​​organized according to time sequence and spatial location to form a lithology probability distribution matrix sequence, which is then input into the trained Transformer model, which predicts and calculates the lithology distribution in the target interval.

[0101] In this embodiment, the method may further include acquiring prior geological data and using the prior geological data for the cross-scale correlation or interval-scale lithology prediction, wherein the prior geological data includes excavation direction, structural plane angle, and borehole core data.

[0102] Specifically, prior geological data can be introduced as constraint information during model input or prediction to further improve the rationality of the prediction results. After standardizing the prior geological data, it is fused with multi-scale lithological characteristic data and used as part of the model input in the prediction process.

[0103] As can be seen from the above embodiments, this application constructs a multi-scale collaborative identification and prediction system of "rock cuttings - rock blocks - working face - interval", which unifies the modeling and correlation analysis of the originally fragmented multi-source data. It forms a complete technical link from data collection, feature extraction, cross-scale fusion to interval prediction, effectively solving the technical problems of insufficient utilization of multi-scale information, unstable identification results and lack of reliable basis for interval scale prediction in the existing technology.

[0104] Specifically, at the microscopic level, pure lithology identification is performed on rock fragment images to obtain high-purity lithological feature information, thereby providing a reliable calibration benchmark for subsequent identification. At the mesoscopic level, a five-level correlation chain is introduced to map the pure lithology probability at the rock fragment scale to the corresponding rock block, thereby constraining and correcting the identification results at the rock block scale and reducing the impact of mixed lithology and surface interference on the identification results. At the macroscopic level, the lithological region of the tunnel face is divided by fusing the main lithology probability at the rock block scale, so that the identification results have both spatial continuity and structural rationality.

[0105] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the multi-scale identification method for lithology at the working face of a long underground tunnel as described above. Figure 3 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is part of a multi-scale identification device for lithology at the working face of a long underground tunnel provided in an embodiment of the present invention. (Except for...) Figure 3 In addition to the processor, network interface, memory, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0106] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the multi-scale identification method for lithology at the working face of a long underground tunnel as described above. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0107] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0108] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A multi-scale identification method for lithology at the working face of a long underground tunnel, characterized in that, include: Acquire multi-scale lithology-related image data, which includes rock cuttings images, rock block images, and tunnel face images; The rock cuttings images, rock block images, and tunnel face images are preprocessed respectively, and the tunnel face is divided into intervals and segments along the tunnel axis. Then, cross-scale correlation is performed to obtain a five-level correlation chain of "interval-tunnel face-segment-rock block-rock cuttings". Based on the preprocessed rock cutting images, the probability of pure lithology at the rock cutting scale is calculated to achieve microscopic calibration; Based on the preprocessed rock block image, combined with the fifth-level correlation chain and the pure lithology probability at the rock fragment scale, the main lithology probability at the rock block scale is calculated to achieve mesoscopic constraints. Based on the preprocessed face image and combined with the main lithology probability at the block scale, the face-scale lithology region is divided to achieve macroscopic division; Based on the division results of the lithological regions at the working face scale, a standardized database is constructed; Based on the standardized database, a Transformer model is constructed and trained to perform interval-scale lithology prediction. The tunnel face image corresponding to the interval to be predicted and its associated lithological feature data are input into the trained Transformer model to perform interval-scale lithology prediction and obtain the prediction result. The lithological feature data consists of the pure lithology probability at the rock debris scale, the main lithology probability at the rock block scale, and the result of dividing the lithological region at the tunnel face scale.

2. The method according to claim 1, characterized in that, It also includes acquiring geological prior data and using the geological prior data for the cross-scale correlation or interval-scale lithology prediction, wherein the geological prior data includes excavation direction, structural plane angle and borehole core data.

3. The method according to claim 1, characterized in that, The rock cuttings images, rock block images, and tunnel face images are preprocessed respectively. The tunnel face is then divided into intervals and segments along the tunnel axis. Cross-scale correlation is then performed to obtain a five-level correlation chain: "interval-tunnel face-segment-rock block-rock cuttings," including: The rock cuttings image, rock block image and tunnel face image are uniformly cropped, denoised and enhanced with lithological features to obtain lithological related feature parameters and intermediate results. Based on the spatial position corresponding to the preprocessed tunnel face image, multiple consecutive tunnel faces are divided into intervals along the tunnel axis, and the tunnel face is segmented within each interval. The lithology-related characteristic parameters and intermediate results are normalized. Based on the normalized lithological characteristic parameters and intermediate results, a cross-scale association table in CSV format is generated to describe the correspondence between intervals, working faces, segments, rock blocks and rock cuttings, thereby establishing a five-level association chain of "interval-working face-segment-rock block-rock cuttings".

4. The method according to claim 1, characterized in that, Based on the preprocessed rock cuttings images, the probability of pure lithology at the rock cuttings scale is calculated, including: Feature extraction is performed on the preprocessed rock debris image to generate candidate regions, and rock debris bounding boxes and pixel-level mask images are output in parallel. The pixel-level mask images correspond to clean rock debris regions. The pixel-level masked image is divided into multiple image blocks, which are then converted into feature vectors and input into a Transformer encoder. The self-attention mechanism is used to capture the microscopic features of minerals, and the transfer learning parameters are adjusted based on the pre-trained weights, which are obtained from the ImageNet dataset. The output is a probability distribution of lithology categories. Calculate the probability of pure lithology at the lithological scale based on the probability distribution of the lithological categories.

5. The method according to claim 1, characterized in that, Based on the preprocessed rock block image, combined with the fifth-order correlation chain and the pure lithology probability at the rock fragment scale, the main lithology probability at the rock block scale is calculated, including: Macroscopic structural features of the rock blocks are extracted from the preprocessed rock block images, and the initial probability distribution of lithology categories is output, represented in the form of a probability vector. Based on the five-level association chain and combined with the pure lithology probability at the rock fragment scale, the initial probability distribution of the lithology category is calibrated to obtain the main lithology probability at the rock block scale, which is represented in the form of a probability vector.

6. The method according to claim 1, characterized in that, Based on the preprocessed tunnel face image and combined with the main lithology probability at the rock block scale, the lithological regions at the tunnel face scale are divided, including: Different receptive field features are extracted from the preprocessed face image. The shallow edge features and deep semantic features are fused by the decoder skip connection to output a pixel-level lithology probability distribution matrix. The probability vector corresponding to the main lithology probability at the block scale is multiplied element-wise with the pixel-level lithology probability distribution matrix. Lithology probabilities less than a set threshold are set to zero, and the effective lithology distribution is retained to obtain the probability of each pixel belonging to each lithology category after constraints. The lithology corresponding to the maximum probability of each pixel is selected as the lithology category of that pixel, and lithological regions are formed based on pixel connectivity, outputting the lithological region division results at the face-of-face scale.

7. The method according to claim 1, characterized in that, Based on the lithological zone division results at the working face scale, a standardized database is constructed, including: The lithological probability distribution matrix, cross-scale correlation table, and geological prior data of the working face of each blasting cycle are indexed by combining timestamps and spatial coordinates and stored in a standardized database, which includes a historical sequence table, a geological parameter table, and a cross-scale correlation table. The data in the standardized database is processed to remove invalid data and fill in data breaks to ensure the continuity of the database in time and space, and to support retrieval by interval identifier and cyclic number; After the completion process is completed, the data in the standardized database is periodically verified, and the anomaly identification results are corrected.

8. The method according to claim 1, characterized in that, The Transformer model adopts an Encoder-Decoder architecture; The encoder contains multiple encoder layers. It uses a self-attention mechanism to calculate the correlation weights of each face and region in the historical sequence to capture the spatiotemporal evolution trend of lithology. A cross-attention mechanism is used to fuse prior geological data. The Decoder comprises multiple DecoderLayers. Based on the output features of the Encoder, it generates the lithology probability distribution matrix for the next blasting cycle through a masked self-attention mechanism; and calculates the reliable probability of lithology prediction during model inference.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-8.

Images