Tunnel surrounding rock identification method and system

Abstract

The invention provides a tunnel surrounding rock identification method and system, and relates to the technical field of tunnel construction, and the method comprises the steps: obtaining a mixed frequency sound wave reflection signal collected by a sensor array disposed on a tunnel face, synchronously acquiring the hyperspectral image data of the tunnel face rock mass from the visible light to the short wave infrared band through a portable spectrometer; topological mode construction is carried out according to the mixed frequency sound wave reflection signals, and topological invariant features are obtained; performing cross-modal fusion according to the topology invariant features and the hyperspectral image data to obtain a fused topology-spectrum association graph; hidden manifold structure learning is carried out according to the fusion topology-spectrum correlation graph, and a low-dimensional feature vector is obtained; performing dynamic surrounding rock category division according to the low-dimensional feature vector to obtain a surrounding rock dynamic classification label; and performing identification result decision according to the surrounding rock dynamic classification label to obtain a tunnel surrounding rock identification result. According to the invention, high-precision and high-robustness dynamic identification of the tunnel surrounding rock state is realized.

Description

Technical Field

[0001] This invention relates to the field of wireless network control technology, and more specifically, to a method and system for identifying surrounding rock in tunnels. Background Technology

[0002] Tunnel surrounding rock identification is a crucial technical aspect of safe construction and dynamic design in tunnel engineering. For a long time, it has relied primarily on traditional methods such as geological logging and core drilling. While intuitive, these methods suffer from limitations including low efficiency, high subjectivity, and difficulty in providing real-time feedback. In recent years, with the development of geophysical exploration and intelligent algorithms, surrounding rock identification technologies based on single-sensor data such as acoustic waves and spectra have emerged. These technologies extract and classify data features through signal processing or machine learning models, improving the automation level of identification to some extent. However, these existing technologies typically isolate detection data with different physical properties, failing to achieve deep integration of multimodal information at the fundamental level. They struggle to capture the intrinsic correlation between rock mass structure, mineral composition, and stress state. Furthermore, most methods do not adequately consider the dynamic evolution of surrounding rock conditions over time, often relying on static snapshots for analysis. This ignores the temporal continuity of surrounding rock response during tunnel excavation, leading to inconsistent identification results and difficulty in reflecting evolutionary trends. In addition, at the decision-making level, they often rely on simple thresholds or preset rules, exhibiting weak modeling capabilities for the inherent uncertainties of the data and the coherence of the geological background. This results in a need to improve the reliability and robustness of identification results under complex geological conditions.

[0003] Based on the shortcomings of the existing technologies, there is an urgent need for a method and system for identifying surrounding rock in tunnels. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying surrounding rock in tunnels, thereby improving the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a method for identifying surrounding rock in tunnels, including: Acquire mixed-frequency acoustic wave reflection signals collected by a sensor array deployed at the tunnel face, and hyperspectral image data of the tunnel face rock mass in the visible to short-wave infrared bands simultaneously acquired by a portable spectrometer; Based on the mixed-frequency acoustic wave reflection signal, a topological pattern is constructed, and a persistent topological barcode of acoustic wave response is constructed by analyzing the multi-frequency response to obtain the topological invariant features characterizing the rock mass fracture and pore structure. Cross-modal fusion is performed based on the topological invariant features and the hyperspectral image data. By associating the persistence range of the topological barcode with the absorption bands of mineral features, a fused topological-spectral correlation map is obtained. Based on the fused topological-spectral correlation graph, implicit manifold structure learning is performed. By calculating the geodesic distance of the correlation graph nodes on the nonlinear manifold, a low-dimensional feature vector reflecting the potential instability mode of the rock mass is obtained. Dynamic surrounding rock classification is performed based on the low-dimensional feature vectors. Density clustering and the concept of sliding window in time series analysis are introduced to deduce the temporal evolution characteristics of the surrounding rock state during tunnel excavation, and dynamic classification labels of the surrounding rock are obtained. The identification result decision is made based on the dynamic classification label of the surrounding rock, and the final tunnel surrounding rock identification result is obtained by fusing multiple time-series classification results through confidence weighting.

[0005] Secondly, this application also provides a tunnel surrounding rock identification system, comprising: The acquisition module is used to acquire mixed-frequency acoustic wave reflection signals collected by a sensor array arranged on the tunnel face, as well as hyperspectral image data of the tunnel face rock mass in the visible to short-wave infrared bands, which are simultaneously acquired by a portable spectrometer. The construction module is used to construct a topology pattern based on the mixed frequency acoustic wave reflection signal, and to construct a persistent topology barcode of acoustic wave response by analyzing the multi-frequency response, thereby obtaining topological invariant features characterizing the rock mass fracture and pore structure. The fusion module is used to perform cross-modal fusion based on the topological invariant features and the hyperspectral image data. By associating the persistence interval of the topological barcode with the absorption band of the mineral feature, a fused topological-spectral correlation map is obtained. The learning module is used to learn the implicit manifold structure based on the fused topology-spectral correlation graph. By calculating the geodesic distance of the nodes in the correlation graph on the nonlinear manifold, a low-dimensional feature vector reflecting the potential instability mode of the rock mass is obtained. The segmentation module is used to dynamically classify the surrounding rock categories based on the low-dimensional feature vectors. It uses density clustering and introduces the concept of a sliding window from time series analysis to deduce the temporal evolution characteristics of the surrounding rock state during tunnel excavation and obtain dynamic classification labels for the surrounding rock. The output module is used to make identification decisions based on the dynamic classification labels of the surrounding rock, and obtain the final tunnel surrounding rock identification result by fusing multiple time-series classification results through confidence weighting.

[0006] The beneficial effects of this invention are as follows: This invention constructs topological invariant features from multi-frequency acoustic wave and hyperspectral data, and fuses them across modes to learn the implicit manifold structure of the surrounding rock. Then, it combines time series analysis to perform dynamic classification and confidence optimization, thereby achieving high-precision and high-robust dynamic identification of the surrounding rock state of tunnels. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of the process of a tunnel surrounding rock identification method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a tunnel surrounding rock identification system according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a tunnel surrounding rock identification device according to an embodiment of the present invention; Figure 4 This is a diagram showing the arrangement of the sensor array.

[0009] The markings in the figure are as follows: 800, a tunnel surrounding rock identification device; 801, processor; 802, memory; 803, multimedia component; 804, I / O interface; 805, communication component; 901, acquisition module; 902, construction module; 903, fusion module; 904, learning module; 905, division module; 906, output module. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0011] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0012] Example 1: This embodiment provides a method for identifying the surrounding rock of a tunnel.

[0013] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0014] Step S100: Acquire mixed-frequency acoustic wave reflection signals collected by a sensor array arranged at the tunnel face, and hyperspectral image data of the tunnel face rock mass in the visible to short-wave infrared bands simultaneously acquired by a portable spectrometer. Understandably, the core of this step lies in simultaneously acquiring two fundamentally different types of physical data: acoustic wave reflection signals reflect the macroscopic mechanical structure and internal defects of the rock mass, while hyperspectral data reveals the mineral chemical composition of the rock mass surface. Specifically, an array of sensors deployed at the tunnel face is used for acoustic wave signal acquisition. This is necessary because a single sensor can only provide point-like information, while an array arrangement allows for simultaneous reception by multiple probes, capturing the response field distribution of acoustic waves propagating in the rock mass space. This spatial distribution data is essential for constructing subsequent topological features; preferably, such as... Figure 4 As shown, the sensor array can be arranged in a ring or grid pattern. This arrangement can effectively cover key areas such as the tunnel face arch and sidewalls, ensuring uniformity of spatial sampling and avoiding detection blind spots. The acquired mixed-frequency acoustic reflection signals should include a specifically designed low-frequency component (for penetrating the rock mass and reflecting macroscopic fracture structures) and a high-frequency component (more sensitive to microscopic pores and small defects), together forming a raw dataset reflecting the multi-scale structural characteristics of the rock mass. Simultaneously, a hyperspectral imager located near the bottom inside the tunnel face acquires hyperspectral image data of the rock mass in the visible to short-wave infrared bands. The hyperspectral image data needs to cover the visible to short-wave infrared bands, which have characteristic absorption bands for common alteration minerals such as hydrous minerals and clay minerals in the rock mass, thus providing detailed mineral composition information. The strict temporal and spatial synchronization of these two types of data ensures that the physical structure and chemical composition information correspond to the same rock mass region, providing a precisely matched data foundation for subsequent cross-modal deep fusion.

[0015] Step S200: Construct a topology pattern based on the mixed frequency acoustic wave reflection signal, and construct a persistent topology barcode of acoustic wave response by analyzing the multi-frequency response to obtain the topological invariant features characterizing the rock mass fracture and pore structure. It should be noted that this step uses topology as a mathematical tool to characterize the acoustic wave response. Unlike conventional signal feature extraction methods, topological pattern construction does not focus on specific signal amplitudes or frequency points, but rather analyzes the globally invariant features of the rock mass, such as the connectivity and shape of pores and fractures, revealed by different frequency responses. This approach makes the description of rock mass structure more robust, bypassing subtle interferences such as signal attenuation in complex media and directly capturing its inherent topological information.

[0016] Step S300: Perform cross-modal fusion based on topological invariant features and hyperspectral image data. By associating the persistence range of the topological barcode with the absorption bands of mineral features, a fused topological-spectral correlation map is obtained. Understandably, this step maps the "persistence" characteristics of the topological barcode representing the physical structure of the rock mass to the spectral absorption "bands" representing mineral composition. Essentially, this explores the intrinsic connection between the mechanical structural integrity of the rock mass and its key mineral composition (such as alteration minerals), thereby constructing a unified correlation map that simultaneously reflects the physical state and chemical properties of the rock mass.

[0017] Step S400: Based on the fused topology-spectral correlation graph, perform implicit manifold structure learning. By calculating the geodesic distance of the nodes in the correlation graph on the nonlinear manifold, obtain a low-dimensional feature vector reflecting the potential instability mode of the rock mass. It should be noted that this step is based on a fundamental assumption: although the state data of the surrounding rock exhibits a complex distribution in high-dimensional space, the essential factors determining its stability reside on a low-dimensional nonlinear manifold. Learning this implicit manifold structure through geodesic distance calculation aims to reduce the dimensionality of the high-dimensional, complex correlation graph data and map it to a space of essential features that more intuitively reflects the stability of the rock mass. This allows subsequent classification to ignore redundant information and directly focus on the core features most relevant to potential instability modes.

[0018] Step S500: Based on the low-dimensional feature vector, the surrounding rock is dynamically classified. Density clustering is used and the concept of sliding window in time series analysis is introduced to deduce the temporal evolution characteristics of the surrounding rock state during tunnel excavation, and the dynamic classification label of the surrounding rock is obtained. Understandably, this step introduces a dynamic perspective to address the engineering reality that the surrounding rock condition is not static but evolves over time during tunnel excavation. By combining temporal density clustering with a sliding window, this method not only classifies the surrounding rock condition at the current moment but also focuses on capturing the evolutionary patterns of state categories as the excavation progresses, such as gradual changes, abrupt changes, or continuations. This makes the identification results no longer isolated snapshots but dynamic sequences containing information on the trend of surrounding rock behavior, which better meets the dynamic needs of engineering decision-making.

[0019] Step S600: Make a decision on the identification result based on the dynamic classification label of the surrounding rock, and obtain the final tunnel surrounding rock identification result by weighted fusion of multiple time-series classification results with confidence.

[0020] It should be noted that due to the inherent uncertainty in geological identification, the classification result at a single moment may be unreliable. Therefore, by introducing confidence assessment and optimization, and fusing the decision results from multiple time-series windows, this method aims to improve the robustness and reliability of the final output. This weighted decision-making mechanism based on time-series context mimics the process by which experienced geological engineers synthesize information from multiple sources, thus achieving more cautious and reliable automatic identification at the algorithmic level.

[0021] Further, step S200 includes steps S210 to S230.

[0022] Step S210: Perform frequency domain feature separation processing on the mixed frequency acoustic wave reflection signal, and use the matching pursuit algorithm to adaptively decompose the signal into isolated frequency band components corresponding to macroscopic crack scattering and microscopic pore resonance respectively, to obtain multi-scale acoustic wave response modes. Step S220: Perform preliminary calculations of topological invariants based on the multi-scale acoustic response modes. Construct an alpha complex by analyzing the spatial distribution of the response amplitudes of each mode and calculate the variation of each Betti number with the distance parameter to obtain a multi-scale topological invariant sequence. Step S230: Based on the multi-scale topological invariant sequence, integrate and encode the topological features. By aligning and fusing the Betty number barcodes at different scales according to their persistence in the tunnel stress field direction, a persistent acoustic response topological barcode is formed, thus obtaining the topological invariant features.

[0023] Specifically, the process begins with frequency domain feature separation in step S210. The key lies in using a matching pursuit algorithm to adaptively decompose the mixed frequency signal into a series of isolated frequency band components. This processing is not a simple filtering process; rather, based on the physical mechanism of acoustic wave scattering at different scales in the rock medium, it separates the response signal into different modes corresponding to macroscopic fracture scattering effects and microscopic pore resonance effects, thus achieving targeted analysis of the multi-scale structure of the rock mass in principle. Based on the separated multi-scale acoustic response modes, step S220 performs preliminary calculations of topological invariants, constructing an alpha complex for the amplitude spatial distribution received by the sensor array under each mode. The alpha complex is a mathematical tool that constructs geometric shapes based on the spatial distance of point sets and attribute thresholds (acoustic wave amplitude in this embodiment). By systematically changing the distance parameter and calculating the changes in Betti numbers of each order (used to quantify the invariants of topological structure such as connected regions and the number of tunnel-like cavities) with the parameter, a multi-scale topological invariant sequence is obtained. This process transforms the complex spatial amplitude distribution into a quantitative topological descriptor characterizing the connectivity and cavity features of the rock mass structure. Step S230 then integrates features based on the characteristics of the stress environment in tunnel engineering. By aligning and fusing Betty number barcodes (charts recording the generation and disappearance of topological features with varying observation scales) obtained at different scales along the dominant direction of the stress field caused by tunnel excavation, a persistent acoustic response topological barcode is ultimately formed. This processing not only integrates multi-scale information but also ensures that the output topological invariant features contain the response characteristics of the rock mass structure to the engineering stress environment, enhancing the correlation between features and surrounding rock stability. The entire processing chain embodies a progressive approach from signal decomposition to geometric structure abstraction and then to feature fusion in the engineering context, aiming to extract topological descriptors that are sensitive to changes in rock mass structure and have clear engineering significance. The alpha complex construction formula is expressed as: ; In the formula, Indicated in scale Amplitude threshold Lower and distance scales The alpha complex constructed below; Indicates the first Acoustic response modes at various scales; Indicates modality The set of points formed by the response amplitudes at the spatial locations of the sensor array at the tunnel face; Point The amplitude of the acoustic response at that location; Indicated by point With the center, and A two-dimensional circle with radius . It is a threshold parameter based on the amplitude of the acoustic response; It represents a simplex (such as a point, line segment, triangle, etc.) in the alpha complex. Indicates the distance scale parameter; This condition must apply to the simplex. All vertices are true.

[0024] Formula for generating multiscale topological invariant sequences: ; In the formula, Indicated in scale Distance parameters Below, the complex's first Betty number; For complex The Homophonic groups; , For the first Rank and first Boundary operators of order; Boundary operators The core represents all Dimensional cycle; Boundary operators The image indicates that it is actually a certain Boundary of dimensional entities Dimensional cycle; Indicates dimension.

[0025] Further, step S300 includes steps S310 to S330.

[0026] Step S310: Perform quantitative inversion processing of mineral composition based on hyperspectral image data. Analyze the depth and area of ​​characteristic absorption bands of preset types of alteration minerals using a nonlinear spectral unmixing model to obtain a quantitative distribution map of mineral composition of the face rock mass. Step S320: Based on the quantitative distribution map of mineral composition and the topological invariant features, cross-modal correlation relationship is constructed. The nonlinear mapping relationship between different mineral combinations and the persistence interval of the topological barcode is modeled through graph attention network, and the tunnel axis direction is introduced as a spatial constraint to obtain the preliminary topological-spectral correlation feature tensor. Step S330: Based on the preliminary topological-spectral correlation feature tensor, perform correlation optimization and spectrum generation processing. Fuse multimodal information through tensor decomposition and use topological persistence as weight to optimize the correlation strength, generating a fused topological-spectral correlation graph.

[0027] Specifically, step S310 first involves the quantitative analysis of hyperspectral imaging data. The core of this process is the application of a nonlinear spectral unmixing model. This model overcomes the mineral spectral mixing effect and precisely quantifies key parameters such as the depth and area of ​​characteristic absorption bands of specific altered minerals (e.g., chlorite, kaolinite). It converts the spectral information of each pixel into the percentage of mineral components, thereby generating a quantitative map that accurately reflects the mineral types and spatial abundance distribution at the tunnel face. This provides a solid chemical composition data foundation for subsequent correlation with physical and mechanical characteristics. Building upon this, step S320 constructs a deep correlation between two heterogeneous features. Preferably, a graph attention network is used, treating pixels in the mineral distribution map and persistent intervals in the topological features as nodes in the graph. Through the network's attention mechanism, it adaptively learns the complex nonlinear mapping relationship between different mineral combinations and persistent intervals of topological barcodes characterizing rock mass structure features, rather than pre-setting simple rules. Crucially, this processing introduces the tunnel axis direction as a spatial constraint, enabling the model to consider the inherent anisotropy of the surrounding rock response in tunnel engineering when learning correlations. Ultimately, it outputs a feature tensor containing complex cross-modal relationships. Step S330 refines and visualizes this preliminary association. Tensor decomposition is used to reduce the dimensionality and denoise the aforementioned feature tensors, extracting the core components of the cross-modal association. During this process, the persistence of the topological features themselves is used as weights, assigning stronger association importance to topological structures that are stable across multiple scales, thereby optimizing the association strength. Ultimately, a fused topological-spectral association map with a clear structure, quantified relationships, and an intuitive representation of the intrinsic connections between the physicochemical properties of the rock mass is generated, providing a high-quality data foundation for subsequent manifold learning. The entire process embodies a progressive relationship from data quantification to nonlinear relationship modeling, and then to information condensation and visualization, aiming to deeply explore the coupling mechanism between mineral composition and rock mass structure.

[0028] Further, step S400 includes steps S410 to S430.

[0029] Step S410: Perform graph structure optimization processing based on the fused topology-spectral correlation graph. By introducing the difference between the tunnel axial and radial ground stresses as anisotropic weights into the edge weight calculation of the correlation graph, the optimized anisotropic correlation graph is obtained. Step S420: Perform diffusion process simulation based on the anisotropic correlation diagram. By simulating the diffusion process of physical quantities on the graph structure considering stress field constraints, construct a diffusion affinity matrix that reflects the correlation of the internal structure of the rock mass. Step S430: Extract manifold coordinates based on the diffusion affinity nucleus matrix. By performing eigenvalue decomposition on the matrix and selecting the eigenvector corresponding to the largest eigenvalue, a low-dimensional eigenvector representing the essential structure of the surrounding rock stability is obtained.

[0030] Specifically, step S410 optimizes the initial fusion graph in an engineering context by introducing the key physical factor inherent in tunnel engineering—the difference in axial and radial ground stress caused by excavation and unloading—as an anisotropic weight into the weight calculation of the connecting edges between nodes in the correlation graph. This makes the optimized graph structure no longer an abstract mathematical object that only reflects data similarity, but an "anisotropic correlation graph" that deeply integrates the constraints of the actual stress state of the rock mass. Its edge weights exhibit differences along the tunnel's axial and radial directions, more accurately simulating the actual impact of the ground stress field on the correlation of the internal structure of the rock mass. Based on this optimized graph structure, step S420 detects the potential connectivity or affinity between nodes by simulating the diffusion process of a virtual "physical quantity" (such as heat or matter) on the anisotropic graph. In this diffusion simulation, the edge weights under the ground stress constraint determine the ease or difficulty of "diffusion." The final constructed diffusion affinity kernel matrix reflects the essential correlation strength between any two nodes (representing different local features of the rock mass) in the graph under the comprehensive consideration of the correlation between physical structure and chemical composition and the ground stress field constraint. Step S430 aims to extract the low-dimensional representation that best represents the main structural features from this affinity kernel matrix containing complex relationships. By performing eigenvalue decomposition on this matrix and selecting the eigenvector corresponding to the largest eigenvalue, the high-dimensional, complex correlation data is projected into a low-dimensional subspace. The direction of this subspace retains the most important variation information in the data. The resulting low-dimensional eigenvector can capture the essential structural patterns hidden behind the original multimodal data that determine the stability state of the surrounding rock, providing a concise and highly representative feature input for subsequent dynamic classification. The entire processing chain embodies a progressive approach: embedding prior engineering knowledge (geostress) into the graph structure, mining deep correlations through physical process simulation (diffusion), and then using mathematical tools (eigenvalue decomposition) to extract essential information.

[0031] Further, step S500 includes steps S510 to S530.

[0032] Step S510: Based on the low-dimensional feature vectors, perform spatiotemporal feature sequence construction processing. By combining the tunnel face mileage information in the tunnel excavation direction, the continuously collected low-dimensional feature vectors are organized according to the excavation sequence to form a spatiotemporal feature sequence. Step S520: Perform density clustering processing based on the spatiotemporal feature sequence. By applying a density clustering algorithm that considers temporal reachability within a sliding window, a temporal density clustering tree that reflects the gradual change process of the surrounding rock state is generated. Step S530: Based on the temporal density clustering tree, the surrounding rock state is deduced and labels are generated. By analyzing the merging and splitting patterns of the clustering tree branches and detecting the evolution path of clusters crossing the sliding window, dynamic classification labels of the surrounding rock that characterize the evolution trend of surrounding rock stability are obtained.

[0033] Specifically, the spatiotemporal feature sequence construction in step S510 combines low-dimensional feature vectors reflecting the essential characteristics of the surrounding rock with tunnel face mileage information in the tunnel excavation direction, and organizes the feature vectors according to the time sequence and spatial location of the excavation. This process transforms a series of isolated feature points into a continuous feature sequence in the spatiotemporal dimension, so that each feature vector has a clear temporal context and spatial location information, laying a data foundation for analyzing the dynamic evolution of the surrounding rock state as the project progresses. Based on the sequence of the dynamic process at this moment, step S520 performs density clustering under temporal constraints. Within a sliding spatiotemporal window, an improved density clustering algorithm is applied. This algorithm not only considers the density distribution of feature points in the feature space, but also introduces the concept of "temporal reachability," that is, it prioritizes classifying points that are spatially and temporally adjacent and have similar features into one class. By continuously sliding windows and executing this clustering process, a temporal density clustering tree can be constructed. This tree structure records the complete phylogenetic relationship of different clusters (representing different types of surrounding rock conditions) as the tunneling process progresses, including their generation, continuation, merging, splitting, or disappearance, thus intuitively showing the gradual or abrupt change process of the surrounding rock condition. Step S530 parses the clustering tree to generate dynamic labels. By deeply analyzing the merging and splitting patterns of branches in the clustering tree, key nodes in the stability transition of the surrounding rock state can be identified. Simultaneously, by detecting the evolution paths of clusters that persist across continuous time windows, the dominant evolutionary trend of the surrounding rock state can be inferred. Finally, the dynamic classification label assigned to each time-series node not only contains the category information of the current state but also implies its evolutionary direction and stability trend information relative to previous states, achieving an improvement from static snapshot identification to dynamic process deduction. The entire process embodies the technical path of using spatiotemporal context information to capture and characterize the continuous evolutionary behavior of the surrounding rock state through temporal clustering tree analysis.

[0034] Further, step S600 includes steps S610 to S630.

[0035] Step S610: Confidence assessment is performed based on the dynamic classification labels of the surrounding rock. The initial confidence of the surrounding rock classification at each time node is calculated by analyzing the relative position density of the corresponding low-dimensional feature vector in its dynamic cluster. Step S620 performs confidence optimization processing based on the initial confidence level and the dynamic classification label sequence of the surrounding rock. By introducing geological coherence constraints, confidence propagation and smoothing are performed on classification results that are abrupt in time but adjacent in space to obtain the optimized integrated confidence level. Step S630: Perform multi-hypothesis decision fusion processing based on the optimized ensemble confidence. The optimal classification hypothesis of the current and previous sliding windows is fused through a weighted voting mechanism, and the category corresponding to the highest ensemble confidence is taken as the final tunnel surrounding rock identification result.

[0036] Specifically, step S610 quantifies the certainty of the classification by analyzing the relative position density of low-dimensional feature vectors representing the rock mass state within their assigned dynamic clusters. If the feature vector is located in the dense core region of the cluster, it indicates a high degree of consistency with the typical characteristics of that category, resulting in a high initial confidence level. If it is located in the sparse region at the edge of the cluster, it indicates greater uncertainty in the classification. This step provides a basic quantification of the credibility of each identification result. Based on this, step S620 optimizes the initial confidence level using spatiotemporal context information. By introducing the fundamental constraint principle of "geological coherence," the identification result sequence is examined, especially for classification results that show abrupt changes in the time series but are spatially adjacent (face mileage). By constructing spatial nearest neighbor relationships, the confidence values ​​of high-confidence nodes that are geographically adjacent and of the same category are propagated and smoothed to these low-confidence abrupt change nodes caused by local disturbances. This effectively suppresses the jumps in identification results caused by non-geological factors, thereby obtaining a more consistent integrated confidence level in the spatiotemporal dimension. Step S630 then performs the final integrated decision based on this. The mechanism no longer relies simply on the classification result at a single moment, but instead uses a weighted voting method to integrate the classification hypotheses (i.e., the optimal candidate categories) with the highest confidence after optimization within the current and previous sliding windows. The category with the highest integrated confidence is ultimately determined as the tunnel surrounding rock identification result. This multi-window, multi-hypothesis-based decision fusion mechanism simulates the approach of judging based on observation trends over a period of time, significantly improving the anti-interference ability and long-term stability of the final output result. The entire process embodies a progressive decision-making strategy from single-point uncertainty assessment to spatiotemporal context optimization, and then to multi-time-window information fusion, aiming to output a more reliable and practical identification conclusion.

[0037] Further, step S620 includes steps S621 to S623.

[0038] Step S621: Perform temporal abrupt change point detection processing based on the dynamic classification label sequence of the surrounding rock. By calculating the Jason-Shannon divergence of the category label distribution within adjacent temporal windows, identify non-geological temporal abrupt change points caused by brief local disturbances at the tunnel face, and obtain a set of potentially unreliable classification nodes. Step S623: Based on the potentially unreliable classification node set and the initial confidence, perform spatial context confidence propagation processing. By constructing a spatial nearest neighbor graph based on the tunnel face mileage coordinates, the confidence values ​​of high-confidence nodes that are geographically adjacent and have the same category label are propagated to low-confidence nodes, resulting in a confidence distribution corrected by spatial constraints. Step S623: Based on the confidence distribution corrected by spatial constraints, perform smoothing processing based on geological continuity priors. By introducing a Gaussian process regression model, with mileage as the input variable and the corrected confidence as the observation value, the geological continuity along the tunneling direction is strengthened, and the optimized integrated confidence is obtained.

[0039] Specifically, step S621, which detects outliers in the classification result sequence, is essentially about calculating the Jason-Shannon divergence of the dynamic classification label distribution of the surrounding rock within adjacent time windows. This divergence can effectively measure the degree of difference between two probability distributions. Through this calculation, nodes whose category distribution changes drastically in a short period of time can be identified. These abrupt changes are often not caused by actual changes in geological conditions, but rather by measurement or classification biases caused by brief local disturbances at the tunnel face (such as surface loosening or local water seepage). This allows for the identification of a set of potentially unreliable classification nodes that require confidence correction. Based on this identification result, step S622 focuses on local correction using spatial context. The method is to construct a spatial nearest neighbor map based on the mileage coordinates of the tunnel face. On this basis, the confidence values ​​of high-confidence nodes that are spatially adjacent and classified into the same surrounding rock category are propagated and assigned to low-confidence nodes that are identified as potentially unreliable. This process is based on the reasonable assumption that "spatial adjacent rock mass areas of the same category should have similar reliability". It effectively utilizes the spatial continuity of the tunnel surrounding rock and makes a preliminary correction to the low-confidence results caused by local non-geological factors, thus obtaining a confidence distribution corrected by spatial constraints. Step S623 further integrates fundamental geological principles at a macroscopic scale by introducing a Gaussian process regression model for global smoothing based on prior geological coherence. This model uses the mileage along the tunnel excavation direction as the input variable and the spatially corrected confidence score as the observed value for fitting. The inherent smoothing properties of the Gaussian process reinforce the prior knowledge that geological bodies along the tunnel axis typically possess continuity and gradual change, ultimately outputting an optimized integrated confidence score that is smoother in spatial distribution and more consistent with the understanding of geological laws. These three steps embody a progressive optimization strategy, from identifying unreliable points to utilizing spatial relationships for local correction, and finally to global smoothing based on geological principles.

[0040] Example 2:

[0041] like Figure 2 As shown, this embodiment provides a tunnel surrounding rock identification system, the system including: The acquisition module 901 is used to acquire mixed frequency acoustic wave reflection signals collected by a sensor array arranged on the tunnel face, and hyperspectral image data of the tunnel face rock mass in the visible to shortwave infrared bands, which are simultaneously acquired by a portable spectrometer. Module 902 is used to construct a topology pattern based on the mixed-frequency acoustic wave reflection signal, and to construct a persistent topology barcode of acoustic wave response by analyzing the multi-frequency response, thereby obtaining topological invariant features characterizing the fracture and pore structure of the rock mass. The fusion module 903 is used to perform cross-modal fusion based on topological invariant features and hyperspectral image data. By associating the persistence range of the topological barcode with the absorption bands of mineral features, a fused topological-spectral correlation map is obtained. Learning module 904 is used to learn the implicit manifold structure based on the fused topological-spectral correlation graph. By calculating the geodesic distance of the nodes in the correlation graph on the nonlinear manifold, a low-dimensional feature vector reflecting the potential instability mode of the rock mass is obtained. The partitioning module 905 is used to dynamically classify the surrounding rock categories based on low-dimensional feature vectors. It uses density clustering and introduces the concept of a sliding window from time series analysis to deduce the temporal evolution characteristics of the surrounding rock state during tunnel excavation and obtain dynamic classification labels for the surrounding rock. The output module 906 is used to make identification decisions based on the dynamic classification labels of the surrounding rock. It obtains the final tunnel surrounding rock identification result by fusing multiple time-series classification results through confidence weighting.

[0042] In one specific embodiment of this application, the construction module 902 includes: The first building unit is used to perform frequency domain feature separation processing based on the mixed frequency acoustic wave reflection signal. The signal is adaptively decomposed into isolated frequency band components corresponding to macroscopic crack scattering and microscopic pore resonance respectively through the matching pursuit algorithm to obtain multi-scale acoustic wave response modes. The second building unit is used to perform preliminary calculations of topological invariants based on the multi-scale acoustic response modes. It constructs an alpha complex by analyzing the spatial distribution of the response amplitudes of each mode and calculates the variation of each Betti number with the distance parameter to obtain a multi-scale topological invariant sequence. The third building unit is used to integrate and encode topological features based on the multi-scale topological invariant sequence. By aligning and fusing the Betty number barcodes at different scales according to their persistence in the tunnel stress field direction, a persistent topological barcode for acoustic response is formed, thus obtaining the topological invariant features.

[0043] In one specific embodiment of this application, the fusion module 903 includes: The first fusion unit is used to perform quantitative inversion processing of mineral composition based on hyperspectral image data. By analyzing the depth and area of ​​the characteristic absorption bands of preset types of altered minerals based on a nonlinear spectral unmixing model, a quantitative distribution map of mineral composition of the rock mass at the working face is obtained. The second fusion unit is used to construct cross-modal correlations based on the quantitative distribution map of mineral composition and topological invariant features. It models the nonlinear mapping relationship between different mineral combinations and the persistence interval of the topological barcode through graph attention network, and introduces the tunnel axis direction as a spatial constraint to obtain the preliminary topological-spectral correlation feature tensor. The third fusion unit is used to optimize the association relationship and generate the spectrum based on the preliminary topological-spectral association feature tensor. It fuses multimodal information through tensor decomposition and uses topological persistence as a weight to optimize the association strength, generating a fused topological-spectral association graph.

[0044] Example 3:

[0045] Corresponding to the above method embodiments, this embodiment also provides a tunnel surrounding rock identification device. The tunnel surrounding rock identification device described below and the tunnel surrounding rock identification method described above can be referred to in correspondence.

[0046] Figure 3 This is a block diagram illustrating a tunnel surrounding rock identification device 800 according to an exemplary embodiment. Figure 3 As shown, the tunnel surrounding rock identification device 800 may include: a processor 801 and a memory 802. The tunnel surrounding rock identification device 800 may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0047] The processor 801 controls the overall operation of the tunnel surrounding rock identification device 800 to complete all or part of the steps in the tunnel surrounding rock identification method described above. The memory 802 stores various types of data to support the operation of the tunnel surrounding rock identification device 800. This data may include, for example, instructions for any application or method operating on the tunnel surrounding rock identification device 800, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the tunnel surrounding rock identification device 800 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0048] In an exemplary embodiment, a tunnel surrounding rock identification device 800 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the tunnel surrounding rock identification method described above.

[0049] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the tunnel surrounding rock identification method described above. For example, the computer-readable storage medium may be the memory 802 including the program instructions described above, which may be executed by a processor 801 of a tunnel surrounding rock identification device 800 to complete the tunnel surrounding rock identification method described above.

[0050] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying surrounding rock in tunnels, characterized in that, include: Acquire mixed-frequency acoustic wave reflection signals collected by a sensor array deployed at the tunnel face, and hyperspectral image data of the tunnel face rock mass in the visible to short-wave infrared bands simultaneously acquired by a portable spectrometer; Based on the mixed-frequency acoustic wave reflection signal, a topological pattern is constructed, and a persistent topological barcode of acoustic wave response is constructed by analyzing the multi-frequency response to obtain the topological invariant features characterizing the rock mass fracture and pore structure. Cross-modal fusion is performed based on the topological invariant features and the hyperspectral image data. By associating the persistence range of the topological barcode with the absorption bands of mineral features, a fused topological-spectral correlation map is obtained. Based on the fused topological-spectral correlation graph, implicit manifold structure learning is performed. By calculating the geodesic distance of the correlation graph nodes on the nonlinear manifold, a low-dimensional feature vector reflecting the potential instability mode of the rock mass is obtained. Dynamic surrounding rock classification is performed based on the low-dimensional feature vectors. Density clustering and the concept of sliding window in time series analysis are introduced to deduce the temporal evolution characteristics of the surrounding rock state during tunnel excavation, and dynamic classification labels of the surrounding rock are obtained. The identification result decision is made based on the dynamic classification label of the surrounding rock, and the final tunnel surrounding rock identification result is obtained by fusing multiple time-series classification results through confidence weighting.

2. The tunnel surrounding rock identification method according to claim 1, characterized in that, Topology pattern construction based on the mixed-frequency acoustic wave reflection signal includes: Based on the mixed-frequency acoustic wave reflection signal, frequency domain feature separation processing is performed, and the signal is adaptively decomposed into isolated frequency band components corresponding to macroscopic crack scattering and microscopic pore resonance respectively through the matching pursuit algorithm to obtain multi-scale acoustic wave response modes. The topological invariants are initially calculated based on the multi-scale acoustic response modes. An alpha complex is constructed by analyzing the spatial distribution of the response amplitudes of each mode, and the variation of each Betti number with the distance parameter is calculated to obtain a multi-scale topological invariant sequence. Based on the multi-scale topological invariant sequence, the topological features are integrated and encoded. By aligning and fusing the Betty number barcodes at different scales according to their persistence in the tunnel stress field direction, a persistent acoustic response topological barcode is formed, thus obtaining the topological invariant features.

3. The tunnel surrounding rock identification method according to claim 1, characterized in that, Cross-modal fusion based on the topological invariant features and the hyperspectral image data includes: Based on the hyperspectral image data, a quantitative inversion of mineral composition is performed. The depth and area of ​​the characteristic absorption bands of preset types of alteration minerals are analyzed by a nonlinear spectral unmixing model to obtain a quantitative distribution map of mineral composition of the face rock mass. Based on the quantitative distribution map of mineral composition and the topological invariant features, cross-modal correlation relationships are constructed. The nonlinear mapping relationship between different mineral combinations and the persistence interval of the topological barcode is modeled through graph attention network, and the tunnel axis direction is introduced as a spatial constraint to obtain the preliminary topological-spectral correlation feature tensor. Based on the preliminary topological-spectral correlation feature tensor, correlation relationship optimization and spectrum generation are performed. Multimodal information is fused through tensor decomposition, and topological persistence is used as a weight to optimize the correlation strength, generating a fused topological-spectral correlation graph.

4. The tunnel surrounding rock identification method according to claim 1, characterized in that, Learning the implicit manifold structure based on the fused topology-spectral correlation graph includes: The graph structure is optimized based on the fused topology-spectral correlation graph. The difference between the tunnel axial and radial ground stresses is introduced as anisotropic weights into the edge weight calculation of the correlation graph to obtain the optimized anisotropic correlation graph. The diffusion process is simulated based on the anisotropic correlation diagram. By simulating the diffusion process of physical quantities on the diagram structure considering stress field constraints, a diffusion affinity matrix reflecting the correlation of the internal structure of the rock mass is constructed. Based on the diffusion affinity nucleus matrix, manifold coordinates are extracted. By performing eigenvalue decomposition on the matrix and selecting the eigenvector corresponding to the largest eigenvalue, a low-dimensional eigenvector characterizing the essential structure of the surrounding rock stability is obtained.

5. The tunnel surrounding rock identification method according to claim 1, characterized in that, Dynamic surrounding rock classification is performed based on the low-dimensional feature vectors, including: Based on the low-dimensional feature vectors, a spatiotemporal feature sequence is constructed. By combining the tunnel face mileage information in the tunnel excavation direction, the continuously collected low-dimensional feature vectors are organized according to the excavation sequence to form a spatiotemporal feature sequence. Density clustering is performed based on the spatiotemporal feature sequence. By applying a density clustering algorithm that considers temporal reachability within a sliding window, a temporal density clustering tree that reflects the gradual change process of the surrounding rock state is generated. Based on the temporal density clustering tree, the surrounding rock state is inferred and a label is generated. By analyzing the merging and splitting patterns of the clustering tree branches and detecting the evolution path of clusters crossing the sliding window, dynamic classification labels of the surrounding rock that characterize the evolution trend of surrounding rock stability are obtained.

6. The tunnel surrounding rock identification method according to claim 1, characterized in that, The identification result decision is made based on the dynamic classification label of the surrounding rock, including: The confidence level is evaluated based on the dynamic classification label of the surrounding rock. The initial confidence level of the surrounding rock classification at each time node is calculated by analyzing the relative position density of the corresponding low-dimensional feature vector in its dynamic cluster. The confidence level is optimized based on the initial confidence level and the dynamic classification label sequence of the surrounding rock. By introducing geological coherence constraints, the confidence level is propagated and smoothed for classification results that are abrupt in time but adjacent in space, and the optimized integrated confidence level is obtained. Based on the optimized integrated confidence, a multi-hypothesis decision fusion process is performed. The optimal classification hypothesis of the current and previous sliding windows is fused through a weighted voting mechanism, and the category corresponding to the highest integrated confidence is taken as the final tunnel surrounding rock identification result.

7. The tunnel surrounding rock identification method according to claim 6, characterized in that, Based on the initial confidence level and the dynamic classification label sequence of the surrounding rock, confidence level optimization processing is performed, including: Based on the dynamic classification label sequence of the surrounding rock, a temporal abrupt change point detection process is performed. By calculating the Jason-Shannon divergence of the category label distribution within adjacent temporal windows, non-geological temporal abrupt change points caused by brief local disturbances at the tunnel face are identified, and a set of potentially unreliable classification nodes is obtained. Based on the set of potentially unreliable classification nodes and the initial confidence, spatial context confidence propagation processing is performed. By constructing a spatial nearest neighbor graph based on the tunnel face mileage coordinates, the confidence values ​​of high-confidence nodes that are geographically adjacent and have the same category label are propagated to low-confidence nodes, resulting in a confidence distribution corrected by spatial constraints. Based on the spatially constrained confidence distribution, a smoothing process based on geological continuity priors is performed. By introducing a Gaussian process regression model, mileage is used as the input variable and the corrected confidence is used as the observation value for fitting, thereby enhancing the geological continuity along the tunneling direction and obtaining the optimized integrated confidence.

8. A tunnel surrounding rock identification system, characterized in that, include: The acquisition module is used to acquire mixed-frequency acoustic wave reflection signals collected by a sensor array arranged on the tunnel face, as well as hyperspectral image data of the tunnel face rock mass in the visible to short-wave infrared bands, which are simultaneously acquired by a portable spectrometer. The construction module is used to construct a topology pattern based on the mixed frequency acoustic wave reflection signal, and to construct a persistent topology barcode of acoustic wave response by analyzing the multi-frequency response, thereby obtaining topological invariant features characterizing the rock mass fracture and pore structure. The fusion module is used to perform cross-modal fusion based on the topological invariant features and the hyperspectral image data. By associating the persistence interval of the topological barcode with the absorption band of the mineral feature, a fused topological-spectral correlation map is obtained. The learning module is used to learn the implicit manifold structure based on the fused topology-spectral correlation graph. By calculating the geodesic distance of the nodes in the correlation graph on the nonlinear manifold, a low-dimensional feature vector reflecting the potential instability mode of the rock mass is obtained. The segmentation module is used to dynamically classify the surrounding rock categories based on the low-dimensional feature vectors. It uses density clustering and introduces the concept of a sliding window from time series analysis to deduce the temporal evolution characteristics of the surrounding rock state during tunnel excavation and obtain dynamic classification labels for the surrounding rock. The output module is used to make identification decisions based on the dynamic classification labels of the surrounding rock, and obtain the final tunnel surrounding rock identification result by fusing multiple time-series classification results through confidence weighting.

9. The tunnel surrounding rock identification system according to claim 8, characterized in that, The building module includes: The first building unit is used to perform frequency domain feature separation processing on the mixed frequency acoustic wave reflection signal, and adaptively decompose the signal into isolated frequency band components corresponding to macroscopic crack scattering and microscopic pore resonance respectively through the matching pursuit algorithm to obtain multi-scale acoustic wave response modes. The second construction unit is used to perform preliminary calculations of topological invariants based on the multi-scale acoustic response modes. By constructing an alpha complex based on the spatial distribution of the response amplitudes of each mode, and calculating the variation of each order Betti number with the distance parameter, a multi-scale topological invariant sequence is obtained. The third construction unit is used to integrate and encode topological features based on the multi-scale topological invariant sequence. By aligning and fusing the Betty number barcodes at different scales according to their persistence in the tunnel stress field direction, a persistent acoustic response topological barcode is formed, thus obtaining the topological invariant features.

10. The tunnel surrounding rock identification system according to claim 8, characterized in that, The fusion module includes: The first fusion unit is used to perform quantitative inversion processing of mineral composition based on the hyperspectral image data. By analyzing the depth and area of ​​the characteristic absorption bands of preset types of alteration minerals based on a nonlinear spectral unmixing model, a quantitative distribution map of mineral composition of the face rock mass is obtained. The second fusion unit is used to construct cross-modal correlations based on the quantitative distribution map of mineral components and the topological invariant features. It models the nonlinear mapping relationship between different mineral combinations and the persistence interval of the topological barcode through a graph attention network, and introduces the tunnel axis direction as a spatial constraint to obtain a preliminary topological-spectral correlation feature tensor. The third fusion unit is used to optimize the association relationship and generate the spectrum based on the preliminary topological-spectral association feature tensor. It fuses multimodal information through tensor decomposition and uses topological persistence as a weight to optimize the association strength, thereby generating a fused topological-spectral association graph.

Images