Intelligent evaluation system for tunnel surrounding rock stability based on multi-source image fusion

Abstract

The present application relates to the field of tunnel engineering, and particularly relates to a tunnel surrounding rock stability intelligent evaluation system based on multi-source image fusion, which comprises an image acquisition module, an image processing module, an image recognition module, a time series analysis module, a risk monitoring module and a surrounding rock evaluation module, the image processing module adopts a pyramid feature matching method and a feature similarity algorithm to realize registration and multi-scale superposition fusion of features of different image sources; the image recognition module identifies surrounding rock abnormal features by using a deep learning network model; the time series analysis module maps multi-time sequence surrounding rock state features to a high-dimensional manifold based on a Riemann manifold model, and calculates deformation rate and acceleration through geodesic theory and curvature analysis; the risk monitoring module calculates real-time monitoring indexes and sends early warning signals; the surrounding rock evaluation module evaluates surrounding rock grades through a surrounding rock stability index, and analyzes support safety risks.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering, specifically to an intelligent assessment system for tunnel surrounding rock stability based on multi-source image fusion, used for real-time monitoring and assessment of the stability status of tunnel surrounding rock under high stress conditions. Background Technology

[0002] Tunnel surrounding rock stability assessment is a crucial step in ensuring the safe construction and operation of tunnel projects. Traditional surrounding rock stability assessments mainly rely on manual on-site observation and limited monitoring data, which suffers from problems such as strong subjectivity, poor real-time performance, and low accuracy. With the development of sensor technology and artificial intelligence, image-based surrounding rock condition monitoring methods are gradually being applied to tunnel engineering. However, existing technologies still have the following shortcomings: first, a single data source cannot comprehensively reflect the surrounding rock condition; second, there is a lack of an effective fusion mechanism for multi-source heterogeneous data; third, traditional analysis methods struggle to capture the nonlinear dynamic characteristics of surrounding rock deformation; and fourth, the accuracy of surrounding rock stability assessment is not high, making it difficult to achieve precise early warning.

[0003] Currently, research mainly focuses on single-sensor monitoring, simple image processing, and empirical model prediction, lacking a comprehensive solution that organically combines multi-source image fusion, deep learning, and mathematical theory.

[0004] Therefore, there is an urgent need to develop a tunnel surrounding rock stability assessment system that can integrate multi-source image data, accurately analyze the dynamic deformation characteristics of the surrounding rock, and achieve intelligent assessment and early warning. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent assessment system for the stability of tunnel surrounding rock based on multi-source image fusion. By fusing visible light images, infrared images, and laser point cloud data, and combining deep learning and differential geometry theory, it can achieve accurate assessment and early warning of the stability of tunnel surrounding rock under high stress conditions.

[0006] This invention proposes an intelligent assessment system for tunnel surrounding rock stability based on multi-source image fusion, comprising:

[0007] The image acquisition module is used to acquire visible light images, infrared images, and laser point cloud data of the surrounding rock of the tunnel;

[0008] The image processing module is communicatively connected to the image acquisition module and is used to receive the visible light image, the infrared image and the laser point cloud data. It performs denoising and sharpening processing on the visible light image and the infrared image, performs registration using the pyramid feature matching method and feature similarity algorithm, and extracts features from the visible light image and the infrared image to achieve registration and multi-scale superposition and fusion of features from different image sources.

[0009] The image recognition module is communicatively connected to the image processing module and is used to perform feature recognition on the fused tunnel surrounding rock image using a trained deep learning network model, annotate abnormal features of the tunnel surrounding rock, and output data on cracks, seepage, peeling, and displacement on the surrounding rock surface.

[0010] The time series analysis module is connected to the image recognition module and is used to construct a Riemannian manifold model of surrounding rock deformation. It maps the multi-time series surrounding rock state characteristics to a high-dimensional Riemannian manifold, performs differential geometric analysis of the surrounding rock deformation trajectory based on geodesic theory and curvature analysis on the manifold, and calculates the rate and acceleration of surrounding rock deformation.

[0011] The risk monitoring module is communicatively connected to the image recognition module and the time series analysis module. It is used to calculate real-time monitoring indicators of the surrounding rock stability based on the abnormal characteristics of the surrounding rock, the deformation rate, and the deformation acceleration. When the real-time monitoring indicators approach or exceed the warning value, a warning signal is issued.

[0012] The surrounding rock assessment module is communicatively connected to the image recognition module, the time series analysis module, and the risk monitoring module. It is used to perform a comprehensive analysis of the abnormal characteristics of the surrounding rock and the deformation rate to obtain the surrounding rock stability index, assess the surrounding rock level through the surrounding rock stability index, and analyze the support safety risk based on the surrounding rock level.

[0013] Preferably, the time series analysis module includes:

[0014] Manifold building unit is used to map multi-temporal surrounding rock state characteristics to a high-dimensional Riemannian manifold, construct a surrounding rock deformation manifold model, and define the Riemannian metric and tangent space on the manifold.

[0015] The deformation trajectory analysis unit is communicatively connected to the manifold construction unit and is used to construct geodesics on the Riemannian manifold, calculate geodesic curvature, and analyze the dynamic characteristics of the surrounding rock deformation trajectory.

[0016] The stability prediction unit is communicatively connected to the deformation trajectory analysis unit and is used to predict the future deformation trend of the surrounding rock and evaluate the stability of the surrounding rock based on geodesic flow theory.

[0017] Preferably, the manifold building unit is further used for:

[0018] Normalization and dimensionality reduction preprocessing were performed on the multi-temporal characteristics of surrounding rock conditions.

[0019] Construct a mapping function to map the preprocessed feature vectors to points on the Riemannian manifold;

[0020] A Riemannian metric tensor is constructed based on the principle of minimizing the deformation energy of surrounding rock.

[0021] Calculate the connection coefficients and Riemann curvature tensors on the manifold.

[0022] Preferably, the deformation trajectory analysis unit is further used for:

[0023] Construct a discrete geodesic approximation between adjacent state points on the manifold;

[0024] Smoothing the geodesic lines yields a continuous deformation trajectory;

[0025] Calculate the geodesic curvature of the geodesic lines and analyze the acceleration characteristics of the surrounding rock deformation;

[0026] The stability of geodesics is analyzed based on the Jacobi field to assess the stability of surrounding rock deformation.

[0027] Preferably, the stability prediction unit is further used for:

[0028] Based on the final state and deformation trend of historical deformation trajectories, the initial conditions of the geodesic flow equation are set.

[0029] Solve the geodesic flow equations to predict the future deformation trajectory of the surrounding rock;

[0030] Calculate the curvature variation and the divergence of the Jacobian field on the predicted trajectory;

[0031] By comprehensively analyzing curvature, divergence, and Lyapunov index, a stability index is calculated to assess the risk of surrounding rock instability.

[0032] Preferably, the image recognition module includes:

[0033] The crack recognition model is constructed using the transfer learning method. Manually labeled crack contours and crack widths are substituted into the ResNet network to train and obtain the crack recognition model.

[0034] The seepage detection model is a trained PointNet neural network.

[0035] The surrounding rock displacement identification model is a trained PointNet neural network.

[0036] Preferably, the image processing module is further configured to:

[0037] Gaussian filtering noise reduction is performed on the visible light image and the infrared image;

[0038] The denoised image is processed by median filtering, then the difference operation is performed with the original image, the difference image is binarized, and then the sharpened image is obtained by dot product with the original image.

[0039] The SIFT algorithm is used to extract image feature points. The pyramid feature matching method is used to map the optimal feature points at the top of the pyramid to the next layer. The candidate feature points extracted in the next layer are coarsely matched, and candidate feature points that do not meet the threshold constraints are removed. Feature extraction is completed through iteration.

[0040] Preferably, the risk monitoring module is further used for:

[0041] The real-time monitoring index R is calculated based on the crack width, seepage depth and width, displacement, deformation rate and deformation acceleration on the surrounding rock surface.

[0042] An alarm message is issued when R is greater than or equal to 50%.

[0043] When R is greater than or equal to 70%, a shutdown and maintenance message is issued.

[0044] Preferably, the surrounding rock assessment module is further used for:

[0045] Calculate the surrounding rock stability index, including the crack index, water permeability index, displacement index, and thermal impact index;

[0046] The surrounding rock is classified into three grades based on the surrounding rock stability index: Grade a, Grade b, and Grade c.

[0047] Among them, level a indicates no safety risk of support, level b indicates low support risk, and level c indicates medium to high support risk.

[0048] Preferably, the surrounding rock assessment module optimizes the surrounding rock stability index based on a deep belief network and outputs the optimized surrounding rock grade and corresponding support recommendations.

[0049] The present invention has the following beneficial effects:

[0050] 1. By using multi-source image fusion technology, the limitations of incomplete information from a single data source are overcome, providing a comprehensive characterization of the surrounding rock condition;

[0051] 2. Based on differential geometry theory, a Riemannian manifold model of surrounding rock deformation is constructed, which accurately captures the nonlinear dynamic characteristics of surrounding rock deformation and significantly improves the analysis accuracy;

[0052] 3. By combining deep learning models, intelligent identification of surrounding rock anomaly features can be achieved, reducing human intervention and improving identification accuracy;

[0053] 4. The prediction framework based on geodesic flow theory can provide early warning of surrounding rock instability risk 12 to 24 hours in advance, which is about 3 times earlier than traditional methods;

[0054] 5. Through a multi-level risk early warning mechanism, a closed-loop management system is achieved, from monitoring to early warning to support decision-making, providing strong support for tunnel engineering safety;

[0055] 6. The false alarm rate of the system is reduced from 15% to 20% in traditional methods to 3% to 5%, which greatly reduces the waste of resources caused by false alarms;

[0056] 7. Accurately assess the surrounding rock grade and support requirements to avoid over-support, optimize resource allocation, and save 15% to 20% of project costs. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0058] Figure 2 This is a flowchart illustrating the workflow of the image processing module of the present invention.

[0059] Figure 3 This is a schematic diagram of the architecture of the time series analysis module of the present invention;

[0060] Figure 4 This is a flowchart of the surrounding rock stability prediction and risk assessment process of the present invention;

[0061] Figure 5 This is a flowchart for generating the surrounding rock grade assessment and support recommendations of the present invention. Detailed Implementation

[0062] Please refer to Figure 1 - Figure 5 The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0063] Reference Figure 1 The intelligent assessment system for tunnel surrounding rock stability based on multi-source image fusion provided by the present invention includes an image acquisition module 1, an image processing module 2, an image recognition module 3, a time series analysis module 4, a risk monitoring module 5, and a surrounding rock assessment module 6.

[0064] Image acquisition module 1 is used to acquire visible light images, infrared images, and laser point cloud data of the tunnel surrounding rock. Preferably, the visible light camera is an industrial-grade high-definition camera with a resolution of not less than 4 million pixels; the infrared camera is an uncooled infrared thermal imager with a temperature resolution of not more than 0.05℃; and the lidar is a high-precision three-dimensional laser scanner with an accuracy of not more than ±2mm. In one embodiment of the present invention, image acquisition module 1 synchronizes the timestamps of the acquired multi-source data to ensure that different data sources correspond to the surrounding rock state at the same time.

[0065] Image processing module 2 is communicatively connected to image acquisition module 1 and is used to receive visible light images, infrared images and laser point cloud data. It performs denoising and sharpening processing on visible light and infrared images, registration using pyramid feature matching method and feature similarity algorithm, and feature extraction on visible light and infrared images to achieve registration and multi-scale overlay fusion of features from different image sources.

[0066] The image recognition module 3 is connected to the image processing module 2 and is used to perform feature recognition on the fused tunnel surrounding rock image using the trained deep learning network model, mark the abnormal features of the tunnel surrounding rock, and output the crack, seepage, peeling and displacement data of the surrounding rock surface.

[0067] The time series analysis module 4 is connected to the image recognition module 3 to construct a Riemannian manifold model of surrounding rock deformation. It maps the multi-time series surrounding rock state characteristics to a high-dimensional Riemannian manifold, performs differential geometric analysis of the surrounding rock deformation trajectory based on geodesic theory and curvature analysis on the manifold, and calculates the rate and acceleration of surrounding rock deformation.

[0068] The risk monitoring module 5 is connected to the image recognition module 3 and the time series analysis module 4. It is used to calculate the real-time monitoring index of the surrounding rock stability based on the abnormal characteristics of the surrounding rock, as well as the deformation rate and deformation acceleration. When the real-time monitoring index approaches or exceeds the warning value, a warning signal is issued.

[0069] The surrounding rock assessment module 6 is connected to the image recognition module 3, the time series analysis module 4, and the risk monitoring module 5. It is used to conduct a comprehensive analysis of the abnormal characteristics and deformation rate of the surrounding rock to obtain the surrounding rock stability index, evaluate the surrounding rock level through the surrounding rock stability index, and analyze the support safety risk based on the surrounding rock level.

[0070] Reference Figure 2 The image processing module 2 is further used to perform Gaussian filtering noise reduction on the visible light image and the infrared image; the noise-reduced image is then subjected to median filtering, and then a difference operation is performed with the original image. The difference image is binarized and then multiplied with the original image to obtain the sharpened image; the SIFT algorithm is used to extract image feature points, and the pyramid feature matching method is used to map the optimal feature points at the top of the pyramid to the next layer. The candidate feature points extracted in the next layer are coarsely matched, and candidate feature points that do not meet the threshold constraints are removed. Feature extraction is completed through iteration.

[0071] In a preferred embodiment of the present invention, Gaussian filtering noise reduction uses a 5×5 Gaussian kernel with a standard deviation σ set to 1.5. This parameter setting effectively removes random noise from the surrounding rock image while preserving edge details. Binarization processing employs an adaptive thresholding method, with the threshold T calculated using the following formula: .

[0072] Where: μ is the average gray value of the local region, in gray levels, typically ranging from 0 to 255; σ is the standard deviation of the local region, in gray levels; k is an adjustment coefficient, dimensionless, ranging from 0.5 to 2.0, dynamically adjusted according to image contrast. When image contrast is low, k takes a smaller value; when image contrast is high, k takes a larger value.

[0073] The pyramid feature matching method constructs a four-layer pyramid structure, starting matching from the top layer (lowest resolution) and passing the matching results down layer by layer. The feature point matching threshold D is set as follows: .

[0074] in: Let i be the i-th descriptor component of the feature point in image A, which is dimensionless; Let be the i-th descriptor component of the feature point in image B, dimensionless; n is the descriptor dimension, SIFT descriptors are typically 128-dimensional and dimensionless; summation symbol This represents the summation of all n descriptor components; This represents the square root operation, calculating the Euclidean distance. When D is less than a threshold (empirically set to 0.6, dimensionless), the two feature points are considered a match.

[0075] Multi-scale overlay fusion employs a weighted average method, and the fused image F is calculated as follows:

[0076] .

[0077] in: This represents the pixel value of the visible light image at position (x, y), expressed in grayscale or RGB values. This represents the pixel value of the infrared image at position (x, y), in gray levels. This represents the pixel value of the point cloud projection image at position (x, y), in gray levels. , which are the weighting coefficients for visible light images, and are dimensionless; The weighting coefficients for the infrared image are dimensionless. The weight coefficients of the point cloud projection image are dimensionless and satisfy the following conditions: The weighting coefficients are dynamically adjusted based on image quality, typically... The value ranges from 0.4 to 0.5. The value ranges from 0.3 to 0.4. The value ranges from 0.1 to 0.3.

[0078] Image recognition module 3 includes a crack recognition model, a seepage recognition model, and a surrounding rock displacement recognition model.

[0079] The crack recognition model is constructed using transfer learning. Manually labeled crack contours and widths are fed into a ResNet network for training. Specifically, ResNet-50 is used as the base network, retaining the weights of the first 48 layers and modifying the last two fully connected layers to adapt to the crack recognition task. Model training employs a stochastic gradient descent optimizer with an initial learning rate of 0.001, dynamically adjusted using cosine annealing, a batch size of 32, and 100 training epochs. This model achieves a crack width measurement accuracy of ±0.1 mm and a detection recall rate exceeding 95%.

[0080] The seepage detection model is a trained PointNet neural network. PointNet directly processes point cloud data and can effectively identify seepage areas on the surface of the surrounding rock. The network input consists of n×3 point cloud coordinates and n×3 RGB color information (n being the number of points). After processing with T-Net, MLP, and max pooling, the output is the segmentation result of the seepage area. During training, the cross-entropy loss function, Adam optimizer, and learning rate of 0.0001 are used for 200 training epochs. This model can identify seepage traces as small as 2mm in width with an accuracy of over 90%.

[0081] The surrounding rock displacement identification model also uses the PointNet neural network, but it has been optimized for displacement detection tasks. The model compares point cloud data at different time points and calculates the displacement vector field. To improve the displacement detection accuracy, the model adopts a dual-time window strategy, simultaneously analyzing short-term (within 24 hours) and long-term (within 7 days) displacement changes, enabling the detection of minute displacements (accuracy up to ±0.5 mm).

[0082] Reference Figure 3 According to claim 2, the time series analysis module 4 includes a manifold construction unit 41, a deformation trajectory analysis unit 42, and a stability prediction unit 43.

[0083] Manifold building unit 41 is used to map multi-temporal surrounding rock state characteristics to a high-dimensional Riemannian manifold, construct a surrounding rock deformation manifold model, and define the Riemannian metric and tangent space on the manifold. This unit is one of the core innovations of this invention, applying differential geometry theory to surrounding rock deformation analysis, which can accurately characterize the nonlinear dynamic characteristics of surrounding rock deformation.

[0084] Manifold building unit 41 is further used to normalize and dimensionality-reducing preprocessing of multi-time series surrounding rock state characteristics; a mapping function is constructed to map the preprocessed feature vectors to points on the Riemann manifold; a Riemann metric tensor is constructed based on the principle of minimizing surrounding rock deformation energy; and the connection coefficient and Riemann curvature tensor on the manifold are calculated.

[0085] In a preferred embodiment of the present invention, the multi-time series surrounding rock state characteristics include feature data such as cracks, seepage, and displacement on the surrounding rock surface. The feature vector X(t) at time t is represented as:

[0086] .

[0087] in: The width parameter of the i-th crack is represented in mm; The parameter represents the seepage area at the j-th location, in mm; , , The three components of the displacement vector are represented in mm; the superscript T indicates the transpose of the vector; X(t) is a column vector whose dimensions depend on the number of cracks, seepage areas and displacement measurement points being monitored, and are usually 20-50 dimensions.

[0088] The eigenvector normalization uses the Min-Max method: .

[0089] in: The normalized eigenvectors are dimensionless, and the values ​​of each component range from [0,1]. This is the minimum value of the eigenvector, with the same unit as the original eigenvector; The maximum value of the eigenvector is represented by the same unit as the original eigenvector; subtraction and division operations are performed on each element of the vector separately.

[0090] Dimensionality reduction employs Principal Component Analysis (PCA), retaining the principal components (typically 10–20 dimensions) needed to explain 95% of the variance. The mapping function φ maps the normalized eigenvectors onto a d-dimensional Riemannian manifold M: .

[0091] Where: p(t) is a point on manifold M, representing the state of the surrounding rock at time t; p(t) is a d-dimensional vector, usually d takes the value of 10-20, which is dimensionless; φ is a mapping function, implemented by radial basis function (RBF) or kernel method.

[0092] The Riemannian metric tensor g is constructed based on the principle of minimizing the deformation energy of the surrounding rock, and its representation at a point p on the manifold is a d×d matrix: .

[0093] in: Let be the dimensionless element in the i-th row and j-th column of the Riemannian metric tensor g at point p; Representation of features manifold coordinates The partial derivatives of represent the mapping relationship between the characteristic space and the manifold space; Features The weighting coefficients are dimensionless and reflect their importance in the assessment of surrounding rock stability; N is the total number of features; the summation symbol is... This represents a weighted summation of all N features. Typical weight values ​​are 0.4-0.5 for crack width, 0.3-0.4 for seepage parameters, and 0.2-0.3 for displacement parameters.

[0094] Connectivity Calculations based on Riemannian metric tensors: .

[0095] in: The connection coefficient represents the parallel transport characteristics on the manifold and is dimensionless. The elements of the inverse matrix of the Riemannian metric tensor satisfy the following condition: ,in is the Kronecker function, which is 1 when k=m and 0 otherwise; Represents the Riemannian metric tensor elements manifold coordinates The partial derivatives of l; according to Einstein's summation convention, the repeating index l is implicitly summed.

[0096] The Riemann curvature tensor R is calculated as follows: .

[0097] in: is a component of the Riemann curvature tensor, which is dimensionless; Represents the connection coefficient manifold coordinates The partial derivatives; The product of the two connection coefficients is summed over the index m; the components of the curvature tensor reflect the geometric characteristics of the surrounding rock deformation manifold, and the larger the value, the more severe or irregular the deformation.

[0098] The deformation trajectory analysis unit 42 is communicatively connected to the manifold construction unit 41 and is used to construct geodesics on the Riemannian manifold, calculate geodesic curvature, and analyze the dynamic characteristics of the surrounding rock deformation trajectory.

[0099] The deformation trajectory analysis unit 42 is further used to construct discrete geodesic approximations between adjacent state points on the manifold; smooth the geodesics to obtain continuous deformation trajectories; calculate the geodesic curvature of the geodesics and analyze the acceleration characteristics of the surrounding rock deformation; and analyze the stability of the geodesics based on the Jacobian field to evaluate the stable state of the surrounding rock deformation.

[0100] In a preferred embodiment of the present invention, the discrete geodesic is approximated by adjacent state points on the manifold. and The shortest path between them. Continuous geodesic γ(t) satisfies the geodesic equation: .

[0101] in: This represents the component of the geodesic γ in the i-th coordinate direction, and is dimensionless. express The second derivative with respect to time t, in units of 1 / time²; express The first derivative with respect to time t, in units of 1 / time; The connection coefficient is dimensionless; according to Einstein's summation convention, the repetition indices j and k are implicitly summed. The geodesic describes the shortest path between two points on the manifold, corresponding to the minimum energy path for surrounding rock deformation.

[0102] The geodesic curvature κ is calculated as follows: .

[0103] Where: κ is geodesic curvature, which is dimensionless; The covariant acceleration vector component of the geodesic is calculated using the following formula: The unit is 1 / time²; D represents the covariant derivative, taking into account the geometric properties of the manifold; The components of the Riemannian metric tensor are dimensionless; according to Einstein's summation convention, the repeated indices i and j are implicitly summed. This represents the square root operation. Geodesic curvature directly reflects the acceleration characteristics of surrounding rock deformation. The larger the κ value, the greater the deformation acceleration and the worse the stability of the surrounding rock. Experience shows that when κ > 0.05, close attention is needed; when κ > 0.1, it indicates that the surrounding rock is in a rapid deformation stage; when κ > 0.2, the surrounding rock is already in a critical state of instability.

[0104] The Jacobian field J(t) describes the motion state of points around a geodesic and satisfies the Jacobian equation:

[0105] .

[0106] in: Let be the component of the Jacobian field in the i-th coordinate direction, which is dimensionless; for The second covariant derivative, in units of 1 / time²; The components of the Riemann curvature tensor are dimensionless. ,in The inverse matrix elements of the Riemannian metric tensor; The components of the geodesic tangent vector are represented by 1 / time. According to Einstein's summation convention, the repeating indices j, k, and l are implicitly summed. The divergence div(J) of the Jacobian field reflects the stability of the geodesic; a positive divergence indicates divergence around the geodesic (the surrounding rock deformation tends to be unstable), while a negative divergence indicates convergence around the geodesic (the surrounding rock deformation tends to be stable).

[0107] The stability prediction unit 43 is communicatively connected to the deformation trajectory analysis unit 42, and is used to predict the future deformation trend of the surrounding rock and evaluate the stability of the surrounding rock based on geodesic flow theory.

[0108] According to claim 5, the stability prediction unit 43 is further used to set the initial conditions of the geodesic flow equation based on the endpoint state and deformation trend of the historical deformation trajectory; solve the geodesic flow equation to predict the future deformation trajectory of the surrounding rock; calculate the curvature change and the divergence of the Jacobian field on the predicted trajectory; and comprehensively analyze the curvature, divergence and Lyapunov index to calculate the stability index and assess the risk of instability of the surrounding rock.

[0109] In a preferred embodiment of the present invention, based on the historical deformation trajectory γ(t), t∈[ , ], Set the initial conditions for the geodesic flow equation:

[0110] .

[0111] .

[0112] in: To predict the trajectory at time... The initial position, and the historical trajectory in The positions at the same time are dimensionless. To predict the trajectory at time... The initial velocity, and the historical trajectory in The speed is the same at any given time, and the unit is 1 / time.

[0113] The numerical solution of the geodesic flow equation adopts the fourth-order Runge-Kutta method, with a time step of 0.1 hours. The prediction time window is dynamically adjusted according to the surrounding rock type and deformation rate, usually ranging from 12 to 72 hours.

[0114] In predicted trajectory Calculate geodesic curvature And Jacobi field divergence Lyapunov Index The calculation is as follows:

[0115] .

[0116] in: The Lyapunov exponent is represented by 1 / time. It represents the limit as time approaches infinity; Represent the natural logarithm function; Let denote the norm of the Jacobian field at time t, which is dimensionless; denoted by the norm of the Jacobian field at the initial moment, dimensionless; t represents time, in units of time.

[0117] In practical calculations, the average value of a finite time window T (usually 24–48 hours) is taken: .

[0118] Where: T is a finite time window, and the unit is time; This is an approximation of the Lyapunov exponent, expressed in units of 1 / time. Lyapunov exponent This reflects the degree of chaos in the system. >0 indicates that the system has chaotic characteristics, the deformation process of the surrounding rock is unstable and sensitive to initial conditions.

[0119] The overall stability index S is calculated as follows: .

[0120] Where: S is the comprehensive stability index, which is dimensionless; α, β and γ are weighting coefficients, which are dimensionless and take values ​​of 0.4, 0.3 and 0.3 respectively, and satisfy α+β+γ=1; The maximum value of geodesic curvature within the predicted time window is dimensionless. The maximum value of the Jacobian field divergence within the predicted time window is given in units of 1 / time. To predict the maximum value of the Lyapunov exponent within the time window, the unit is 1 / time. All parameters were normalized before calculation to ensure consistent dimensions.

[0121] The threshold settings for the stability index S are as follows: S<0.1 indicates that the surrounding rock is stable; 0.1≤S<0.3 indicates that the surrounding rock is potentially unstable; 0.3≤S<0.5 indicates that the surrounding rock is moderately unstable; and S≥0.5 indicates that the surrounding rock is severely unstable and immediate measures are required.

[0122] The risk monitoring module 5 is further used to calculate the real-time monitoring index R based on the crack width, seepage depth and width, displacement, deformation rate and deformation acceleration of the surrounding rock surface; when R is greater than or equal to 50%, an alarm message is issued; when R is greater than or equal to 70%, a shutdown equipment message and maintenance message are issued.

[0123] In a preferred embodiment of the present invention, the real-time monitoring index R is calculated as follows: .

[0124] in: For real-time monitoring indicators, the unit is percentage (%); The crack width is in mm. is the crack width threshold in mm, usually set to 5 mm; s is the seepage depth in mm. d represents the threshold for seepage depth and width, in mm, typically set to 20 mm; d represents the displacement, in mm. This is the displacement threshold, in mm, and is usually set to 10 mm. The deformation rate is expressed in mm / h. This is the rate threshold, measured in mm / h, and is typically set to 0.5 mm / h. This is the deformation acceleration, in mm / h. 2 ; Acceleration threshold, in mm / h 2 It is usually set to 0.05 mm / h. 2 ; , , , and Let be the weighting coefficients, dimensionless, taking values ​​of 0.2, 0.15, 0.2, 0.25, and 0.2 respectively, and satisfying the following conditions: + + + + =1; multiply by 100% to convert the result to a percentage.

[0125] Threshold settings for real-time monitoring indicator R: R<30% indicates a safe state; 30%≤R<50% indicates a need for attention; 50%≤R<70% indicates a need for warning and an alarm message should be issued; R≥70% indicates a dangerous state and an order to shut down the equipment and provide maintenance information should be issued.

[0126] The surrounding rock assessment module 6 is further used to calculate the surrounding rock stability index, including the crack index, seepage index, displacement index, and thermal impact index. Based on the surrounding rock stability index, the module classifies the surrounding rock into three levels: Level A, Level B, and Level C. Level A represents no safety risk without support, Level B represents low risk with support, and Level C represents medium to high risk with support. The surrounding rock assessment module 6 optimizes the surrounding rock stability index based on a deep belief network and outputs the optimized surrounding rock level and corresponding support recommendations.

[0127] In a preferred embodiment of the present invention, the surrounding rock stability index I is calculated as follows: .

[0128] Where: I is the surrounding rock stability index, which is dimensionless; The crack index is dimensionless. Permeability index, dimensionless; is the displacement index, dimensionless; is the thermal influence index, dimensionless; , , and are weight coefficients, dimensionless, with values of 0.35, 0.25, 0.3, and 0.1 respectively, and satisfy + + + = 1.

[0129] Calculation of the crack index : The crack image is cropped using the threshold method to obtain the crack image. The crack width w is obtained by image analysis. If w ≤ 3 mm, then = 1; if w > 3 mm, then = w / 3. Where w is the crack width, in mm; is the crack index, dimensionless.

[0130] Calculation of the water seepage index : The water seepage area is cropped using the threshold method. The pixel width of the image gray value greater than the threshold is the water seepage width s; if 2 mm ≤ s ≤ 5 mm, then = 1; if 5 mm < s ≤ 20 mm, then = 2; if s > 20 mm, then = 3. Where s is the water seepage width, in mm; is the water seepage index, dimensionless.

[0131] Calculation of the displacement index : If the surrounding rock displacement is d, then = d / 5, where d is the displacement amount, in mm; is the displacement index, dimensionless. [[ID=B]]

[0132] Calculation of the thermal influence index : If the thermal influence range is S, then = S / 100, where S is the thermal influence area, in ; is the thermal influence index, dimensionless.

[0133] Threshold setting of the surrounding rock stability index I: I < 1.5 indicates that the surrounding rock is stable, rated as grade a, with no support safety risk; 1.5 ≤ I < 3 indicates that the surrounding rock is relatively stable, rated as grade b, with a low support risk; I ≥ 3 indicates that the surrounding rock is unstable, rated as grade c, with a medium to high support risk.

[0134] Deep Belief Networks (DBNs) were used to optimize the surrounding rock stability index. The DBN consisted of three layers of Restricted Boltzmann Machines (RBMs) with 256, 128, and 64 hidden nodes, respectively. Training employed a contrastive divergence algorithm with a learning rate of 0.01, a momentum coefficient of 0.9, and 1000 training epochs. The input consisted of the surrounding rock stability index and its constituent indicators; the output was the optimized surrounding rock grade and support recommendations.

[0135] Support recommendations are generated based on the surrounding rock grade: Grade A recommends routine monitoring without special support; Grade B recommends enhanced monitoring, with appropriate increases in anchor bolt density and shotcrete thickness; Grade C recommends immediate action, including adding steel arches, increasing anchor bolt density, increasing shotcrete thickness, and, if necessary, implementing advanced support and grouting reinforcement.

[0136] The various modules of this invention and their working principles have been described in detail above. This invention, through multi-source image fusion technology and differential geometry theory, achieves intelligent assessment of the stability of tunnel surrounding rock, providing strong support for tunnel engineering safety. It should be noted that the above embodiments are merely preferred embodiments of this invention and are not intended to limit the scope of protection of this invention. Those skilled in the art can make various equivalent modifications or variations to this invention without departing from its essence, and such modifications or variations should be considered to fall within the scope of protection of this invention.

Claims

1. A tunnel surrounding rock stability intelligent evaluation system based on multi-source image fusion, characterized in that, include: The image acquisition module is used to acquire visible light images, infrared images, and laser point cloud data of the surrounding rock of the tunnel; The image processing module, communicatively connected to the image acquisition module, is used to receive the visible light image, the infrared image, and the laser point cloud data; to perform denoising and sharpening processing on the visible light image and the infrared image; to perform registration using a pyramid feature matching method and a feature similarity algorithm; to extract features from the visible light image and the infrared image; to project the laser point cloud data into a point cloud projection image; and to perform weighted fusion of the visible light image, the infrared image, and the point cloud projection image to achieve registration and multi-scale overlay fusion of features from different image sources. The image recognition module is communicatively connected to the image processing module and is used to perform feature recognition on the fused tunnel surrounding rock image using a trained deep learning network model, annotate abnormal features of the tunnel surrounding rock, and output data on cracks, seepage, peeling, and displacement on the surrounding rock surface. The image recognition module includes: a crack recognition model, which is constructed using a transfer learning method, by substituting manually labeled crack outlines and crack widths into a ResNet network to train and obtain the crack recognition model; a seepage recognition model, which is a trained PointNet neural network; and a surrounding rock displacement recognition model, which is a trained PointNet neural network. A time series analysis module, communicatively connected to the image recognition module, is used to receive surrounding rock anomaly feature data output by the image recognition module at different time points to construct multi-time series surrounding rock state features, build a Riemannian manifold model of surrounding rock deformation, map the multi-time series surrounding rock state features to a high-dimensional Riemannian manifold, and perform differential geometric analysis of the surrounding rock deformation trajectory based on geodesic theory and curvature analysis on the manifold to calculate the deformation rate and deformation acceleration of the surrounding rock. The time series analysis module includes: a manifold construction unit, used to map the multi-time series surrounding rock state features to a high-dimensional Riemannian manifold, construct a surrounding rock deformation manifold model, and define Riemannian metrics and tangent space on the manifold; a deformation trajectory analysis unit, communicatively connected to the manifold construction unit, used to construct geodesics on the Riemannian manifold, calculate geodesic curvature, and analyze the dynamic characteristics of the surrounding rock deformation trajectory; and a stability prediction unit, communicatively connected to the deformation trajectory analysis unit, used to predict the future deformation trend of the surrounding rock and evaluate the stability of the surrounding rock based on geodesic flow theory. The risk monitoring module is communicatively connected to the image recognition module and the time series analysis module. It is used to calculate real-time monitoring indicators of the surrounding rock stability based on the abnormal characteristics of the surrounding rock, the deformation rate, and the deformation acceleration. When the real-time monitoring indicators approach or exceed the warning value, a warning signal is issued. The real-time monitoring indicators are used to characterize the risk level of the surrounding rock at the current moment and trigger a real-time warning. The surrounding rock assessment module is communicatively connected to the image recognition module, the time series analysis module, and the risk monitoring module. It is used to perform a comprehensive analysis of the abnormal characteristics and deformation rate of the surrounding rock to obtain a surrounding rock stability index, assess the surrounding rock grade through the surrounding rock stability index, and analyze the support safety risk based on the surrounding rock grade. The surrounding rock stability index is used to comprehensively assess the overall grade of the surrounding rock and generate support recommendations. 2.The multi-source image fusion based tunnel surrounding rock stability intelligent evaluation system according to claim 1, characterized in that, The manifold building unit is further used for: Normalization and dimensionality reduction preprocessing were performed on the multi-temporal characteristics of surrounding rock conditions. Construct a mapping function to map the preprocessed feature vectors to points on the Riemannian manifold; A Riemannian metric tensor is constructed based on the principle of minimizing the deformation energy of surrounding rock. Calculate the connection coefficients and Riemann curvature tensors on the manifold. 3.The multi-source image fusion based tunnel surrounding rock stability intelligent evaluation system according to claim 1, characterized in that, The deformation trajectory analysis unit is further used for: Construct a discrete geodesic approximation between adjacent state points on the manifold; Smoothing the geodesic lines yields a continuous deformation trajectory; Calculate the geodesic curvature of the geodesic lines and analyze the acceleration characteristics of the surrounding rock deformation; The stability of geodesics is analyzed based on the Jacobi field to assess the stability of surrounding rock deformation. 4.The multi-source image fusion based tunnel surrounding rock stability intelligent evaluation system according to claim 3, characterized in that, The stability prediction unit is further used for: Based on the final state and deformation trend of historical deformation trajectories, the initial conditions of the geodesic flow equation are set. Solve the geodesic flow equations to predict the future deformation trajectory of the surrounding rock; Calculate the curvature variation and the divergence of the Jacobian field on the predicted trajectory; By comprehensively analyzing curvature, divergence, and Lyapunov index, a stability index is calculated to assess the risk of surrounding rock instability. 5.The multi-source image fusion based intelligent tunnel surrounding rock stability evaluation system according to claim 1, characterized in that, The image processing module is further used for: Gaussian filtering noise reduction is performed on the visible light image and the infrared image; The denoised image is processed by median filtering, then the difference operation is performed with the original image, the difference image is binarized, and then the sharpened image is obtained by dot product with the original image. The SIFT algorithm is used to extract image feature points. The pyramid feature matching method is used to map the optimal feature points at the top of the pyramid to the next layer. The candidate feature points extracted in the next layer are coarsely matched, and candidate feature points that do not meet the threshold constraints are removed. Feature extraction is completed through iteration. 6.The multi-source image fusion based intelligent tunnel surrounding rock stability evaluation system according to claim 1, characterized in that, The risk monitoring module is further used for: The real-time monitoring index R is calculated based on the crack width, seepage depth and width, displacement, deformation rate and deformation acceleration on the surrounding rock surface. An alarm message is issued when R is greater than or equal to 50%. When R is greater than or equal to 70%, a shutdown and maintenance message is issued. 7.The multi-source image fusion based intelligent tunnel surrounding rock stability evaluation system according to claim 1, characterized in that, The surrounding rock assessment module is further used for: Calculate the surrounding rock stability index, including the crack index, water permeability index, displacement index, and thermal impact index; The surrounding rock is classified into three grades based on the surrounding rock stability index: Grade a, Grade b, and Grade c. Among them, level a indicates no safety risk of support, level b indicates low support risk, and level c indicates medium to high support risk.

8. The intelligent assessment system for tunnel surrounding rock stability based on multi-source image fusion according to claim 7, characterized in that, The surrounding rock assessment module optimizes the surrounding rock stability index based on a deep belief network and outputs the optimized surrounding rock grade and corresponding support recommendations.

Images