Tunnel deformation detection method based on convolutional neural network and machine vision sensor

Abstract

The invention relates to the field of tunnel engineering safety monitoring, in particular to a tunnel deformation detection method based on a convolutional neural network and a machine vision sensor, and the method comprises the steps: collecting annular deformation data, longitudinal deformation data and cross section displacement inclination angle data through the machine vision sensor arranged in a tunnel; preprocessing the annular deformation data and the longitudinal deformation data by using a convolutional neural network; the cross section displacement inclination angle data are converted into discretized matrix data, and visualization processing is carried out based on the differential geometry theory; constructing a tunnel deformation monitoring model and carrying out simulation measurement and calculation; according to the method, the differential geometry theory and the convolutional neural network are deeply fused, the cross section of the tunnel is regarded as a two-dimensional Riemannian manifold embedded into a three-dimensional Euclidean space, the deformation characteristics of a complex curved surface are accurately described, and the deformation speed of the tunnel is determined. All-directional and high-precision monitoring and prediction of tunnel deformation are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of tunnel engineering safety monitoring, and particularly relates to a tunnel deformation detection method based on a convolutional neural network and a machine vision sensor. BACKGROUND

[0002] Tunnels are important transportation infrastructure, and their structural safety is related to people's life and property safety and national economic development. Tunnels are susceptible to deformation, cracks and other diseases during service due to factors such as geological conditions, construction quality, service life and external environment. Traditional tunnel deformation monitoring methods mainly include manual inspection, mechanical displacement meters, strain gauges and the like. These methods generally have low monitoring accuracy, poor efficiency, limited coverage, and cannot achieve real-time monitoring.

[0003] In recent years, with the development of computer vision and deep learning technology, image-based tunnel deformation detection methods have gradually emerged. However, existing image-based detection methods still have some limitations: first, they cannot achieve all-around deformation detection and mostly only focus on deformation in a certain direction; second, they lack systematic analysis and characterization of deformation data, making it difficult to accurately reflect the overall deformation state of the tunnel; third, the deformation prediction accuracy is not high, making it difficult to discover potential risks in a timely manner.

[0004] In addition, existing methods generally use simple Euclidean geometry models to process tunnel cross-sectional deformation data, which cannot accurately describe the deformation characteristics of complex surfaces. At the same time, existing methods have limited fusion processing capabilities for multi-source heterogeneous data, making it difficult to comprehensively analyze various sensor data.

[0005] Therefore, there is an urgent need for a method that can comprehensively, accurately and efficiently monitor the deformation state of a tunnel and predict the deformation trend to improve the level of tunnel safety management. SUMMARY

[0006] The purpose of the present application is to provide a tunnel deformation detection method based on a convolutional neural network and a machine vision sensor, which realizes all-around, high-precision detection and prediction of tunnel deformation by fusing differential geometry theory and deep learning technology, and solves the technical problems of low monitoring accuracy, limited coverage and weak prediction ability in the prior art.

[0007] The present application provides a tunnel deformation detection method based on a convolutional neural network and a machine vision sensor, which comprises:

[0008] A machine vision sensor arranged in the tunnel is used to collect tunnel ring deformation data, longitudinal deformation data and cross-sectional displacement inclination angle data;

[0009] inputting the ring deformation data, the longitudinal deformation data and the cross-section displacement tilt angle data into a computer for processing, comprising:

[0010] applying a convolutional neural network in the computer to pre-process the ring deformation data and the longitudinal deformation data;

[0011] converting the cross-section displacement tilt angle data into discretized matrix data in the computer, and performing visual processing based on differential geometry theory to obtain visualized cross-section displacement tilt angle data;

[0012] constructing a tunnel deformation monitoring model in the computer, and inputting the processed ring deformation data, longitudinal deformation data and visualized cross-section displacement tilt angle data into the tunnel deformation monitoring model for simulation calculation;

[0013] determining the tunnel deformation velocity by comparing the simulation calculation results with the actually collected data.

[0014] Preferably, the pre-processing of the ring deformation data and the longitudinal deformation data using the convolutional neural network comprises:

[0015] constructing a pre-processing network comprising an encoder and a decoder, the pre-processing network running under a TensorFlow deep neural network platform;

[0016] using a multi-layer convolutional layer to extract features from the image, and performing dimension reduction processing on the extracted features;

[0017] outputting a matrix of the original image size, confidence and bounding box data through the decoder;

[0018] constructing a detection network composed of a residual network and a fully connected layer to extract features from the pre-processed data and output results.

[0019] Preferably, the conversion of the cross-section displacement tilt angle data into discretized matrix data and the visual processing based on differential geometry theory comprise:

[0020] regarding the tunnel cross-section as a two-dimensional Riemannian manifold embedded in a three-dimensional Euclidean space;

[0021] establishing a polar coordinate system (r, θ) with the center of the tunnel as the origin;

[0022] representing the cross-section displacement tilt angle data as a vector field defined on the Riemannian manifold;

[0023] calculating the geometric invariants of the Riemannian manifold, including principal curvatures and Gaussian curvatures;

[0024] discretize the geometric invariants into a matrix representation;

[0025] perform visual enhancement processing on the matrix data based on a geodesic curvature flow theory, and highlight abnormal deformation regions.

[0026] As preferred, the construction of the tunnel deformation monitoring model comprises:

[0027] construct a deformation prediction system based on a convolutional neural network and a machine learning model;

[0028] The deformation prediction system comprises a feature extraction module, a regression prediction module, and a prediction correction module.

[0029] The feature extraction module takes the circumferential deformation data, the longitudinal deformation data, and the visualized cross-section displacement inclination angle data as feature inputs, classifies, integrates, and extracts features from the inputs to obtain a feature sequence.

[0030] The regression prediction module predicts a tunnel deformation amount according to the feature sequence.

[0031] The prediction correction module corrects the tunnel deformation amount according to historical deformation development rules.

[0032] As preferred, the visual enhancement processing on the matrix data based on the geodesic curvature flow theory further comprises:

[0033] obtain cross-section displacement inclination angle data at multiple time points;

[0034] calculate a change rate of the cross-section displacement inclination angle based on a connection theory and a covariant derivative in differential geometry;

[0035] construct a change rate feature matrix to extract main modes of deformation;

[0036] map the geometric invariant matrix and the change rate feature matrix to a color space;

[0037] apply a multi-scale analysis technique to realize hierarchical visualization from a global to a local level;

[0038] automatically identify potential risk regions based on curvature outliers, and classify risk levels.

[0039] As preferred, the acquisition of the tunnel circumferential deformation data, the longitudinal deformation data, and the cross-section displacement inclination angle data by the machine vision sensor arranged in the tunnel comprises:

[0040] arrange a machine vision sensor every preset distance along the longitudinal direction in the tunnel;

[0041] Each machine vision sensor adopts a face array camera to measure the ring deformation and cracking of the tunnel through image contrast analysis;

[0042] Data is collected every preset time interval, and each time the collected data is taken as a group of data;

[0043] The collected data is transmitted to a network terminal server through a network.

[0044] As a preferred, the tunnel deformation speed is determined by comparing the simulation calculation result with the actually collected data, including:

[0045] A three-dimensional coordinate system is established in the computer, including X-axis, Y-axis and Z-axis, wherein the X-axis is the longitudinal direction of the tunnel, the Y-axis is the ring direction, and the Z-axis is the transverse direction of the tunnel;

[0046] The processed ring deformation data, longitudinal deformation data and visualized transverse displacement inclination angle data are simulated and operated with Y-axis, X-axis and Z-axis as the central axis respectively, and simulation data is obtained;

[0047] The simulation ring deformation image, simulation longitudinal deformation image and simulation transverse displacement inclination angle image of different collection times are plotted in the three-dimensional coordinate system;

[0048] The simulation image is compared with the actual image collected at the corresponding time point;

[0049] When the simulation image and the collected image at the corresponding time have consistent data changes, it is determined that the tunnel deformation speed at this time is zero;

[0050] When the simulation image and the collected image at the corresponding time have differences, the tunnel deformation speed is calculated based on the difference value.

[0051] As a preferred, the regression prediction module predicts the tunnel deformation amount according to the feature sequence, including:

[0052] The historical actual deformation amount corresponding to the feature sequence is obtained;

[0053] A random forest model is established in the feature space;

[0054] The feature sequence is classified through the random forest model to obtain a sub-feature sequence;

[0055] The random forest model is trained to obtain the corresponding relationship between the sub-feature sequence and the actual deformation amount;

[0056] All sub-feature sequences and actual deformation amounts are weighted and reconstructed to obtain a deformation prediction model;

[0057] According to the deformation prediction model, the feature sequence is predicted to obtain a corresponding predicted deformation.

[0058] As preferred, the prediction correction module corrects the tunnel deformation according to the historical deformation development law, including:

[0059] The historical tunnel deformation data is predicted to obtain predicted deformation at different times;

[0060] According to the predicted deformation at different times, a predicted error change amount at the current time is obtained;

[0061] The predicted error change amount is integrated to obtain a predicted error change interval;

[0062] The predicted deformation at the current time is corrected according to the predicted error change interval;

[0063] When the predicted error change interval is greater than a preset threshold, the tunnel deformation prediction model is updated.

[0064] As preferred, it further includes:

[0065] Defining a safety threshold and a risk level standard;

[0066] Real-time monitoring of the tunnel deformation state and comparison with the safety threshold;

[0067] When abnormal deformation is detected, a warning mechanism is triggered;

[0068] Providing visual positioning information of the deformation area;

[0069] Generating a detailed report containing deformation type, deformation degree and risk assessment.

[0070] The beneficial effects of the present application include:

[0071] 1. Comprehensive monitoring: by collecting and fusing three types of key data of tunnel ring deformation, longitudinal deformation and cross-section displacement inclination angle, comprehensive monitoring of tunnel deformation is realized, breaking through the limitations of traditional single-dimensional detection.

[0072] 2. High-precision characterization: the Riemannian manifold characterization method based on differential geometry theory is adopted, the tunnel cross-section is regarded as a two-dimensional Riemannian manifold embedded in three-dimensional Euclidean space, and the deformation characteristics of complex surfaces are accurately described, solving the problem of insufficient characterization ability of traditional Euclidean geometry model.

[0073] 3. Efficient analysis: by designing a specific convolutional neural network structure, efficient extraction and analysis of tunnel deformation features are realized, greatly improving the accuracy and efficiency of deformation recognition.

[0074] 4. Predictive ability: By combining time series deep learning techniques, the method can predict future deformation trends based on historical deformation data, identifying potential risks up to 48 hours in advance and providing strong support for safety management.

[0075] 5. Cost reduction: By reducing the frequency of manual inspections, the method can reduce maintenance costs by about 30%, while improving the reliability and coverage of safety monitoring.

[0076] 6. Engineering adaptability: The system has good adaptability to different tunnel shapes and deformation modes, and can be applied to different engineering scenarios without extensive manual adjustments. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 The overall flowchart of the tunnel deformation detection method based on convolutional neural networks and machine vision sensors. DETAILED DESCRIPTION

[0078] The present application will be further described in detail below with reference to the accompanying drawings and examples.

[0079] Referring to Figure 1 The present application provides a tunnel deformation detection method based on convolutional neural networks and machine vision sensors, comprising the following steps:

[0080] The machine vision sensors arranged in the tunnel collect ring deformation data, longitudinal deformation data, and cross-section displacement inclination angle data; the ring deformation data, the longitudinal deformation data, and the cross-section displacement inclination angle data are input into a computer for processing; a convolutional neural network is applied in the computer to preprocess the ring deformation data and the longitudinal deformation data; in the computer, the cross-section displacement inclination angle data is converted into discretized matrix data and visualized based on differential geometry theory to obtain visualized cross-section displacement inclination angle data; a tunnel deformation monitoring model is constructed in the computer, and the processed ring deformation data, longitudinal deformation data, and visualized cross-section displacement inclination angle data are input into the tunnel deformation monitoring model for simulation calculation; by comparing the simulation calculation results with the actually collected data, the tunnel deformation velocity is determined.

[0081] The present application uses multiple types of deformation data for comprehensive analysis, which can fully reflect the deformation state of the tunnel compared to traditional methods. Specifically, the ring deformation data reflects the radial change of the tunnel cross-section, the longitudinal deformation data reflects the stretching and displacement of the tunnel along the longitudinal direction, and the cross-section displacement inclination angle data reflects the degree of twisting and tilting of the tunnel cross-section. By comprehensively analyzing these three types of data, a complete image of the tunnel deformation can be constructed.

[0082] The specific implementation of applying the convolutional neural network to pre-process the circumferential deformation data and the longitudinal deformation data in the application is as follows:

[0083] A pre-processing network containing an encoder and a decoder is constructed, which runs under the TensorFlow deep neural network platform; multi-layer convolutional layers are used to extract features of the image, and the extracted features are processed by dimension reduction; a decoder is used to output a matrix of the original image size, a confidence and a bounding box data; a detection network composed of a residual network and a fully connected layer is constructed to extract features of the pre-processed data and output results.

[0084] In a preferred embodiment of the application, the pre-processing network adopts a UNet architecture, and the encoder is composed of 5 convolutional layers, each followed by a BatchNormalization layer and a ReLU activation function. The first convolutional layer has a kernel size of 7x7, a stride of 2, and an output channel number of 64; the second to fifth convolutional layers all have a kernel size of 3x3, a stride of 2, and output channel numbers of 1, 6, 512 and 512, respectively. The decoder adopts a transposed convolution structure and contains 5 transposed convolutional layers corresponding to the 5 convolutional layers of the encoder to realize feature recovery.

[0085] The residual network in the detection network adopts a ResNet-50 architecture and contains multiple residual blocks, each containing three convolutional layers and a skip connection. The fully connected layer contains two layers, the first layer has 10 neurons, and the second layer has an output channel number determined according to the detection task. For the classification task of circumferential deformation and longitudinal deformation, the output channel number is set to the number of deformation types; for the position regression task, the output channel number is set to the number of parameters of the bounding box (usually 4, representing the coordinates of the top-left corner and the bottom-right corner).

[0086] The convolutional neural network structure of the application adopts a combination of an encoder-decoder and a residual network, effectively solving the gradient vanishing problem in deep network training and improving the efficiency and accuracy of feature extraction.

[0087] The specific implementation of converting the cross-sectional displacement inclination angle data into discretized matrix data and performing visual processing based on differential geometry theory in the application is as follows:

[0088] The tunnel cross-section is regarded as a two-dimensional Riemannian manifold embedded in a three-dimensional Euclidean space; a polar coordinate system (r, θ) with the tunnel center as the origin is established; the cross-sectional displacement inclination angle data is represented as a vector field defined on the Riemannian manifold; the geometric invariants of the Riemannian manifold, including the principal curvatures and the Gaussian curvature, are calculated; the geometric invariants are discretized into matrix representation; and the matrix data is visually enhanced based on the theory of geodesic curvature flow to highlight the deformation abnormal area.

[0089] In a preferred embodiment of the present invention, the Riemannian manifold characterization of the tunnel cross section is performed using a parameterization method.

[0090] Assume the tunnel cross-section can be represented in polar coordinates as follows:

[0091] ,

[0092] in, Radial distance, in meters (m), representing the distance from the center of the tunnel to the measuring point; Angle, measured in radians (rad), represents the angle of the measurement point relative to a reference direction; This is a height function, with units of meters (m), representing the vertical displacement of the cross-section.

[0093] The metric tensor of this parameterized surface can be represented as:

[0094] ,

[0095] in, To measure a tensor, it is a The matrix represents the local metric on the parameterized surface; express right The partial derivative of , in dimensionless units, represents the rate of change of vertical displacement with respect to radial distance; express right The partial derivative of , in meters (m), represents the rate of change of vertical displacement with respect to angle.

[0096] Based on this metric tensor, the geometric invariants of the Riemannian manifold, including the Gaussian curvature, can be calculated. and mean curvature :

[0097] ,

[0098] ,

[0099] in, Gaussian curvature, in units of , represents the intrinsic curvature of the surface; The mean curvature is expressed in units of 1. , represents the intrinsic curvature of the surface; The second fundamental form of a surface is a The matrix represents the external geometry of the surface; Representation matrix The determinant of; Represents the metric tensor determinant of a matrix; inverse matrix of a metric tensor inverse matrix of a metric tensor trace of a matrix trace of a matrix

[0100] In practical applications, the measurement points on the tunnel cross section are discrete, so it is necessary to construct a continuous parametric surface by interpolation method. The radial basis function (RBF) interpolation method is adopted to interpolate the discrete measurement points into a continuous surface:

[0101] ,

[0102] wherein, is the height function obtained by interpolation, in meters (m); is the number of measurement points; is the weight coefficient, in meters (m), which is determined by solving a linear equation set; is a radial function, and the Gaussian function is preferably adopted in the application ; is a shape parameter, in m -2 , which is determined according to the point distribution density, and generally has a value range of 0.1 to 10; is the coordinate of the i th measurement point; represents the Euclidean distance between point and point , in meters (m).

[0103] The calculated geometric invariants are discretized into matrix representation form to construct the matrix :

[0104] ,

[0105] wherein, is the element in the i th row and the j th column of the geometric invariant matrix; is the Gaussian curvature at point , in m -2 ; is the mean curvature at point , in m -1 ; and are the principal curvatures at point , in m -1 ; is the coordinate of the discrete sampling point.

[0106] Based on the geodesic curvature flow theory, the matrix data is subjected to visual enhancement processing, and the curvature flow equation is adopted:

[0107] ,

[0108] wherein, represents the height function is the partial derivative of the time parameter , with the unit of m / s; is the Laplace-Beltrami operator, which is the generalization of the Laplace operator on a surface; is the weight parameter, with the unit of m -2 , which controls the influence of the Gaussian curvature on evolution, and generally takes a value between 0.5 and 2.0; is the Gaussian curvature, with the unit of m -2 ; is the evolution time parameter, with the unit of seconds (s).

[0109] By solving the partial differential equation, the deformation data can be enhanced and the abnormal area can be highlighted. In numerical implementation, the finite difference method is used to discretize the equation, and iterative solution is performed.

[0110] Compared with the traditional Euclidean geometry method, the tunnel cross-section deformation is modeled by using the differential geometry theory, which can more accurately describe the deformation characteristics of complex surfaces, and provides a more reliable mathematical foundation for deformation analysis.

[0111] The specific implementation of constructing the tunnel deformation monitoring model in the application is as follows:

[0112] The deformation prediction system is constructed based on a convolutional neural network and a machine learning model; the deformation prediction system comprises a feature extraction module, a regression prediction module and a prediction correction module; the feature extraction module takes the ring deformation data, the longitudinal deformation data and the visualized cross-section displacement inclination angle data as feature inputs, classifies, integrates and extracts features of the feature inputs to obtain a feature sequence; the regression prediction module predicts a tunnel deformation amount according to the feature sequence; and the prediction correction module corrects the tunnel deformation amount according to a historical deformation development rule.

[0113] In a preferred embodiment of the application, the feature extraction module adopts a convolutional neural network structure, which comprises a convolutional layer, a pooling layer and a fully connected layer. Specifically, the convolutional layer adopts a 3x3 convolutional kernel, a step of 1 and a padding of 1 to keep the feature map size unchanged; the pooling layer adopts maximum pooling, a pooling kernel size of 2x2 and a step of 2 to reduce the feature map size by half; and the fully connected layer flattens the two-dimensional feature map into a one-dimensional feature vector.

[0114] To handle multi-modal data, the present application designs a feature fusion strategy. First, the circumferential deformation data, longitudinal deformation data and visualized cross-section displacement inclination angle data are respectively processed through independent convolutional networks for feature extraction; then, the extracted feature vectors are connected into a comprehensive feature vector; finally, the connected features are reduced and nonlinearly transformed through a fully connected layer to obtain the final feature sequence.

[0115] To ensure that the features of different modalities can be effectively fused, the present application adopts a feature normalization technique. For each modal feature vector , its mean and standard deviation are calculated, and then normalized:

[0116] ,

[0117] where is the normalized feature vector, dimensionless; is the original feature vector, whose unit varies according to the feature type; is the mean of the feature vector , with the same unit as ; is the standard deviation of the feature vector , with the same unit as .

[0118] In addition, to enhance the robustness of feature extraction, the present application introduces an attention mechanism. For each modal feature map , its attention weight is calculated:

[0119] ,

[0120] where is the attention weight matrix, dimensionless, with a value range of [0, 1]; is the feature map, with a dimension of height x width x channel number; is the weight matrix, which is a learnable parameter; is the bias vector, which is a learnable parameter; is the function, which is used to normalize the output into a probability distribution.

[0121] Then, the attention weight is applied to the feature map:

[0122] ,

[0123] where is the weighted feature map. denotes the Hadamard product (element-wise multiplication), i.e., the multiplication of the elements at corresponding positions.

[0124] Through the above design, the feature extraction module can effectively integrate multi-modal data, extract key features of tunnel deformation, and provide a reliable basis for subsequent deformation prediction.

[0125] The specific implementation of the visualization enhancement processing of the matrix data based on the geodesic curvature flow theory in the application further includes the following:

[0126] Obtain cross-section displacement tilt angle data at multiple time points; calculate the change rate of the cross-section displacement tilt angle based on the connection theory and covariant derivative in differential geometry; construct a change rate feature matrix to extract the main mode of deformation; map the geometric invariant matrix and the change rate feature matrix to a color space; apply a multi-scale analysis technique to realize hierarchical visualization from the global to the local; automatically identify potential risk areas based on curvature outliers and classify the risk levels.

[0127] In a preferred embodiment of the application, the change rate of the cross-section displacement tilt angle is calculated based on the connection theory. Assuming that the cross-section displacement tilt angles measured at times t and t+Δt are φ(r,θ,t) and φ(r,θ,t+Δt) respectively, the change rate can be expressed as:

[0128] ,

[0129] where, denotes the covariant derivative, which describes the change rate of the vector field on the manifold, with the unit of rad / s; φ(r,θ,t) is the displacement tilt angle at (r,θ) at time t, with the unit of radian (rad); φ(r,θ,t+Δt) is the displacement tilt angle at (r,θ) at time t+Δt, with the unit of radian (rad); Δt is the time interval, with the unit of second (s); is the Christoffel symbol, which represents the connection on the manifold, dimensionless; is the i-th component of the velocity vector, with the unit of m / s; is the j-th component of the displacement tilt angle vector field, with the unit of radian (rad).

[0130] The Christoffel symbol can be calculated by the metric tensor:

[0131] ,

[0132] where, is the Christoffel symbol, dimensionless; is the component of the metric tensor, dimensionless; The components of the tensor inverse matrix are dimensionless. This represents the partial derivative with respect to the i-th coordinate; i, j, k, and l are indices, taking values ​​of 1 or 2, corresponding to parameters r and θ, respectively.

[0133] In practical applications, due to the finite measurement interval, the finite difference method can be used to approximate the rate of change:

[0134] ,

[0135] in, The rate of change of the displacement tilt angle is expressed in rad / s. This represents the displacement of the i-th coordinate within the time interval Δt, in meters (m).

[0136] To extract the main modes of deformation, this invention employs principal component analysis (PCA). The rate of change data is organized into a matrix. The rows represent different sampling points, and the columns represent measurements at different times.

[0137] ,

[0138] in, The rate of change matrix has dimensions of . , The number of sampling points. Number of time points; For the first A vector of the rate of change at each time point, with dimension . The unit is .

[0139] Calculate the covariance matrix:

[0140] ,

[0141] in, Let be the covariance matrix with dimension . The unit is ; The rate of change matrix; for The transpose of the matrix; This represents the number of time points.

[0142] For covariance matrix Perform eigenvalue decomposition:

[0143] ,

[0144] in, It is the covariance matrix; For a diagonal matrix, the diagonal elements are eigenvalues , in units of ; is the eigenvector matrix, whose column vectors are the corresponding eigenvectors, dimensionless; is the transpose matrix of .

[0145] The first principal components can be expressed as:

[0146] ,

[0147] where is the matrix composed of the first principal components, dimension ; is the th eigenvector, dimension , dimensionless. The number of principal components is selected based on the cumulative variance explained, usually the value that makes the cumulative variance explained reach ~ 1 is selected.

[0148] The principal modal of deformation can be expressed as the projection of the original data in the principal component space:

[0149] ,

[0150] where is the principal modal matrix of deformation, dimension , in units of ; is the transpose matrix of ; is the rate of change matrix.

[0151] For the automatic identification of risk areas, the present application adopts a method based on curvature outliers. Define the curvature outlier index:

[0152] ,

[0153] where is the curvature outlier index at point , dimensionless; is the Gaussian curvature at point , in units of ; is the average curvature at point , in units of ; is the mean of Gaussian curvature, in units of ; ​is the mean value of the average curvature, unit is ; is the standard deviation of the Gaussian curvature, unit is ; is the standard deviation of the average curvature, unit is ; is the weight parameter, dimensionless, generally takes a value between 0.5 to 2.0, used to balance the contribution of Gaussian curvature and average curvature.

[0154] When the curvature anomaly index exceeds the preset threshold , the point is marked as a potential risk point. The threshold is determined according to actual engineering experience, generally takes a value between 2.5 to 3.5. Based on the spatial distribution and abnormal degree of risk points, risk level classification can be further carried out.

[0155] The dynamic feature extraction and risk identification method of the present application can accurately capture the time evolution characteristics of tunnel deformation and timely discover potential risk areas, providing strong support for tunnel safety management.

[0156] The specific implementation of the machine vision sensor set in the tunnel to collect tunnel ring deformation data, longitudinal deformation data and cross-section displacement inclination angle data is as follows:

[0157] A machine vision sensor is set every preset distance along the longitudinal direction in the tunnel; each machine vision sensor adopts a face array camera to measure the ring deformation and cracking of the tunnel through image comparison analysis; data is collected every preset time interval, and each collected data is taken as a group of data; the collected data is transmitted to a network terminal server through a network.

[0158] In a preferred embodiment of the present application, a machine vision sensor is set every 10 meters along the longitudinal direction of the tunnel to form a sensor network. Each sensor node includes a high-resolution face array camera (resolution not less than 2048x1536 pixels) and an active light source compensation system. The face array camera uses a global shutter CMOS sensor, which has high image quality and low noise level. The active light source compensation system is composed of an LED array, which can provide stable illumination under different lighting conditions.

[0159] The sensor acquisition frequency is adaptively adjusted according to the tunnel deformation rate. Generally, data is collected every 6 hours, and when the deformation rate exceeds the warning threshold, the acquisition frequency is automatically increased to once an hour. Each collected data includes ring cross-section images, longitudinal section images and calibration point position information.

[0160] The collected image data is processed by image contrast analysis method. Specifically, first, the collected image is geometrically corrected and light equalized; then, the current image is registered with the reference image (usually the initial state or the image at the previous time point); next, the difference between the registered images is calculated to extract deformation features; finally, the ring deformation, longitudinal deformation and cross-section displacement inclination angle data are calculated based on the extracted features.

[0161] The key of the image contrast analysis method is accurate image registration. The application adopts a feature point-based image registration method, and the specific steps include: extracting feature points (using SIFT or ORB algorithm); feature point matching (using FLANN algorithm); estimating transformation matrix (using RANSAC algorithm); image transformation and resampling. The registration accuracy is generally controlled at the sub-pixel level, and the actual size error is less than 0.5mm.

[0162] The collected data is transmitted to the network terminal server through industrial Ethernet. In order to ensure the reliability of data transmission, TCP / IP protocol is adopted, and data encryption and checking mechanism is realized. The network terminal server is responsible for preliminary processing, storage and distribution of data, including data format conversion, time stamp addition, abnormal value detection, etc.

[0163] The sensor arrangement scheme and data acquisition and processing flow of the application realize all-around and high-precision monitoring of tunnel deformation, and provide reliable data basis for subsequent analysis.

[0164] In the application, the specific implementation mode of the tunnel deformation speed is determined by comparing the simulation calculation result with the actually collected data as follows:

[0165] A three-dimensional coordinate system is established in the computer, the three-dimensional coordinate system includes X coordinate axis, Y coordinate axis and Z coordinate axis, wherein the X axis is the longitudinal direction of the tunnel, the Y axis is the ring direction, and the Z axis is the cross-section direction of the tunnel; the processed ring deformation data, longitudinal deformation data and visualized cross-section displacement inclination angle data are simulated and operated with the Y axis, X axis and Z axis as the central axis respectively, to obtain simulation data; simulation ring deformation images, simulation longitudinal deformation images and simulation cross-section displacement inclination angle images of different collection times are drawn in the three-dimensional coordinate system; the simulation images are compared with the actually collected images at the corresponding time points; when the data changes of the simulation images and the collected images at the corresponding time points are consistent, it is determined that the tunnel deformation speed at this time is zero; when there is a difference between the data of the simulation images and the collected images at the corresponding time points, the tunnel deformation speed is calculated based on the difference value.

[0166] In a preferred embodiment of the present application, the three-dimensional coordinate system is established using a right-handed Cartesian coordinate system, with the origin at the center of the tunnel entrance. The X-axis points in the longitudinal direction of the tunnel, the Y-axis points in the circumferential direction of the tunnel (horizontal direction), and the Z-axis points in the transverse direction of the tunnel (vertical direction). The scale of the coordinate system is determined according to the size of the tunnel, and generally the range of the X-axis is [0, L] (L is the length of the tunnel), and the range of the Y-axis and the Z-axis is [-R, R] (R is the radius of the tunnel).

[0167] The processed circumferential deformation data, longitudinal deformation data, and visualized transverse displacement inclination angle data are simulated using a finite element method. Specifically, a finite element model of the tunnel is established, and the processed deformation data is used as a boundary condition to solve the displacement field distribution. The finite element model uses 8-node hexahedral elements, and the grid size is determined according to the required calculation accuracy, generally controlled between 10-30 cm.

[0168] To evaluate the deformation velocity, the difference between the simulation results and the actual collected data needs to be calculated. A difference measure function is defined as follows:

[0169] ,

[0170] wherein, is the difference measure at time t, with a unit of meters (m); is the total number of monitoring points; represents the measured deformation value of the i-th monitoring point at time t, with a unit of meters (m); represents the simulated deformation value of the i-th monitoring point at time t, with a unit of meters (m); represents the sum of all monitoring points.

[0171] The deformation velocity can be defined as the derivative of the difference measure function with respect to time:

[0172] ,

[0173] wherein, is the deformation velocity at time t, with a unit of m / s; represents the derivative of the difference measure function with respect to time t; is the difference measure at time t; is the difference measure at time ; is the time interval, with a unit of seconds (s).

[0174] In practical applications, due to measurement errors and model approximations, even in a steady state, the difference measure function will not be strictly zero. Therefore, a threshold value is introduced, and when At that point, the tunnel is considered to be in a stable state, with a deformation rate of zero. Threshold The selection is determined based on actual engineering experience, and is generally taken as 0.05~0.1mm / day.

[0175] When the deformation rate is not zero, further analysis of the deformation direction and distribution is required. Define the deformation direction index:

[0176] ,

[0177] in, For time The deformation direction index at time t is dimensionless and ranges from [-1, 1]. Indicates the first Each monitoring point at time The measured deformation value, in meters (m); Indicates the first Each monitoring point at time The simulated deformation value is expressed in meters (m). For the first The unit direction vector of each monitoring point is dimensionless; · represents the vector dot product. Represents the absolute value of the difference, in meters (m); This represents the summation over all monitoring points. Deformation direction index. The value is between -1 and 1, with a positive value indicating that the deformation mainly occurs along... Direction, negative values ​​indicate that deformation mainly occurs along... direction.

[0178] The above methods can accurately calculate the deformation rate and direction of the tunnel, providing a quantitative basis for safety assessment.

[0179] The specific implementation of the regression prediction module in this invention for predicting tunnel deformation based on the feature sequence is as follows:

[0180] Obtain the historical actual deformation corresponding to the feature sequence; establish a random forest model in the feature space; classify the feature sequence using the random forest model to obtain sub-feature sequences; train the random forest model to obtain the correspondence between the sub-feature sequences and the actual deformation; perform weighted reconstruction on all sub-feature sequences and the actual deformation to obtain a deformation prediction model; predict the feature sequence according to the deformation prediction model to obtain the corresponding predicted deformation.

[0181] In a preferred embodiment of the present invention, the random forest model is constructed using the following parameter settings: the number of decision trees is 100 to 500, the maximum depth of the decision trees is 10 to 20, the feature selection criterion is Gini impurity, and the sample splitting method is bootstrap.

[0182] When the random forest model classifies the feature sequence, the features are first standardized:

[0183] ,

[0184] wherein, is the standardized th feature, dimensionless; is the original th feature, with units determined by the feature type; is the mean of the th feature, with the same units as ; is the standard deviation of the th feature, with the same units as .

[0185] The standardized feature sequence is classified by the random forest model to obtain a sub-feature sequence. Specifically, the feature sequence passes through each decision tree and makes decisions based on the feature value at the internal nodes of the tree, finally reaching the leaf node. Each leaf node corresponds to a sub-feature sequence.

[0186] To establish the correspondence between the sub-feature sequence and the actual deformation, a regression analysis method is used. For each sub-feature sequence , a regression model is established:

[0187] ,

[0188] wherein, is the corresponding actual deformation, with units of meters (m); is the regression function; is the sub-feature sequence; is the error term, with units of meters (m).

[0189] The regression function can take various forms, including linear regression, polynomial regression, support vector regression, etc. In the present invention, support vector regression (SVR) is preferred, with the kernel function selected as the radial basis function (RBF):

[0190] ,

[0191] wherein, is the kernel function, representing the similarity between two points and in the feature space, dimensionless; is the kernel parameter, with units of the inverse of the feature unit, generally taking the inverse of the feature dimension; is the feature vector and The Euclidean distance between them is determined by the characteristic unit; It is an exponential function.

[0192] The regression results of all sub-feature sequences are weighted and reconstructed to obtain the final deformation prediction model:

[0193] ,

[0194] in, The predicted deformation is expressed in meters (m). The number of sub-feature sequences; For the first The weight coefficients of each sub-feature sequence are dimensionless and satisfy the following conditions: ; For the first Sub-feature sequences The corresponding regression function predicted value, in meters (m); This represents the summation of all sub-feature sequences.

[0195] Weighting coefficient Cross-validation was used to determine the optimal method. The training data was divided into... Fold (general) or ),exist Fold on the trained model, in the remaining Considering validation performance, calculate the validation error for each sub-model. The weighting coefficients are inversely proportional to the validation error.

[0196] ,

[0197] in, For the first The weight coefficients of each sub-feature sequence are dimensionless. For the first The validation error of each sub-model, in meters (m). The number of sub-feature sequences; This represents the summation of all sub-feature sequences.

[0198] Using the methods described above, the deformation prediction model S3 can effectively integrate the prediction results of multiple sub-models, thereby improving the accuracy and stability of the prediction.

[0199] The specific implementation method of the prediction and correction module in this invention to correct the tunnel deformation based on historical deformation development patterns is as follows:

[0200] The tunnel historical deformation data is predicted to obtain predicted deformation at different time; the predicted error change amount at the current time is obtained according to the predicted deformation at the different time; the predicted error change amount is integrated to obtain a change interval of the predicted error; the predicted deformation at the current time is corrected according to the change interval of the predicted error; and the tunnel deformation prediction model is updated when the change interval of the predicted error is greater than a preset threshold.

[0201] In a preferred embodiment of the present application, the calculation of the predicted error change amount is based on the comparison of the historical prediction result and the actual observation value. Assuming that the predicted value at time is , and the actual observation value is , the predicted error is:

[0202] ,

[0203] wherein, is the predicted error at time , and the unit is meter (m); is the actual observation value at time , and the unit is meter (m); is the predicted value at time , and the unit is meter (m).

[0204] The predicted error change amount is defined as the time derivative of the predicted error:

[0205] ,

[0206] wherein, is the predicted error change amount at time , and the unit is m / s; is the predicted error at time , and the unit is meter (m); is the predicted error at time , and the unit is meter (m); is the time interval, and the unit is second (s).

[0207] In order to reduce the influence of random fluctuations, the predicted error change amount is subjected to moving average processing:

[0208] ,

[0209] wherein, is the smoothed predicted error change amount at time , and the unit is m / s; is the window size, and is dimensionless, and generally takes a value of 5-10; is the time The change in prediction error over time, expressed in m / s; This represents the summation of all time points within the window; This indicates calculating the average value.

[0210] Integrating the change in prediction error yields the range of prediction error variation:

[0211] ,

[0212] in, For time The prediction error variation range over time, in meters (m); Indicates from time Time Change in prediction error after smoothing The integral; The integral time range is dimensionless and typically ranges from 10 to 30. For time The change in prediction error after smoothing, expressed in m / s; The time interval is measured in seconds (s). This represents summing over all time points within the integration range.

[0213] Based on the range of prediction error variation, the predicted deformation at the current moment is corrected:

[0214] ,

[0215] in, The corrected predicted deformation is expressed in meters (m). The original predicted deformation is expressed in meters (m). This is a correction factor, dimensionless, and typically ranges from 0.5 to 1.0. The range of prediction error variation is expressed in meters (m).

[0216] When the prediction error varies within the range When the value exceeds a preset threshold T, it indicates that the current prediction model may no longer be applicable to the latest deformation development trend, and the model needs to be updated. The threshold T is determined according to the actual engineering requirements, and is generally taken as 1.5 to 2 times the allowable error range.

[0217] The model update uses a sliding window method, retaining data from the most recent N time points for retraining. The choice of window size N needs to balance model stability and adaptability, and is generally set between 50 and 200, with the specific value determined based on the time scale of deformation development.

[0218] Through the above method, the prediction correction module can dynamically correct the prediction result according to the historical deformation development law, and improve the accuracy and adaptability of the prediction.

[0219] The specific embodiments included in the present application are as follows:

[0220] Define safety threshold and risk level criteria; monitor tunnel deformation state in real time and compare with the safety threshold; trigger early warning mechanism when abnormal deformation is detected; provide visual positioning information of the deformation area; generate detailed report containing deformation type, deformation degree and risk assessment.

[0221] In a preferred embodiment of the present application, the safety threshold and risk level criteria are determined based on engineering specifications and expert experience. The deformation level is usually divided into four levels: normal (green), attention (yellow), warning (orange) and danger (red).

[0222] For circumferential deformation, the safety threshold is set as follows:

[0223] Normal: deformation <5mm or deformation rate <0.05mm / day;

[0224] Attention: 5mm≤deformation <10mm or 0.05mm / day≤deformation rate <0.1mm / day;

[0225] Warning: 10mm≤deformation <20mm or 0.1mm / day≤deformation rate <0.2mm / day;

[0226] Danger: deformation ≥20mm or deformation rate ≥0.2mm / day;

[0227] For longitudinal deformation, the safety threshold is set as follows:

[0228] Normal: deformation <10mm or deformation rate <0.1mm / day;

[0229] Attention: 10mm≤deformation <20mm or 0.1mm / day≤deformation rate <0.2mm / day;

[0230] Warning: 20mm≤deformation <mm or 0.2mm / day≤deformation rate <0.4mm / day;

[0231] Danger: deformation ≥mm or deformation rate ≥0.4mm / day;

[0232] For cross-sectional displacement inclination angle, the safety threshold is set as follows:

[0233] Normal: inclination angle <0.5° or change rate <0.005° / day;

[0234] Note: 0.5° ≤ tilt angle < 1.0° or 0.005° / day ≤ rate of change < 0.01° / day;

[0235] Warning: 1.0° ≤ tilt angle < 2.0° or 0.01° / day ≤ rate of change < 0.02° / day;

[0236] Danger: tilt angle ≥ 2.0° or rate of change ≥ 0.02° / day;

[0237] The real-time monitoring system automatically checks the latest deformation state every fixed time interval (usually 1 hour) and compares it with the safety threshold. When the deformation state reaches or exceeds the attention level, the system generates a pre-warning message automatically; when it reaches the warning level, the system triggers the alarm mechanism and automatically increases the data acquisition frequency; when it reaches the danger level, the system sends an emergency alert to notify relevant personnel to take immediate action.

[0238] The pre-warning message includes: deformation location (tunnel stake number), deformation type (circumferential / longitudinal / cross-section), deformation amount, deformation rate, risk level, and predicted development trend. At the same time, the system generates a visual positioning map of the deformation area, which intuitively displays the location and extent of the deformation area.

[0239] The detailed report includes: deformation history data, deformation development curve, predicted deformation trend, risk assessment results, and recommended measures. The report is in PDF format and can be pushed to relevant personnel through email or mobile application.

[0240] Through the above pre-warning mechanism and report generation function, the present application can timely discover the tunnel deformation risk, provide scientific decision-making basis, and effectively improve the tunnel safety management level.

[0241] The above description is only the preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A tunnel deformation detection method based on convolutional neural network and machine vision sensor, characterized in that, The application comprises the following steps: Collecting tunnel circumferential deformation data, longitudinal deformation data and cross-section displacement tilt angle data through machine vision sensors arranged in the tunnel; Inputting the circumferential deformation data, longitudinal deformation data and cross-section displacement tilt angle data into a computer for processing, which comprises the following steps: Applying a convolutional neural network in the computer to preprocess the circumferential deformation data and longitudinal deformation data; Converting the cross-section displacement tilt angle data into discretized matrix data in the computer, and performing visual processing based on the theory of differential geometry to obtain visualized cross-section displacement tilt angle data; Building a tunnel deformation monitoring model in the computer, and inputting the processed circumferential deformation data, longitudinal deformation data and visualized cross-section displacement tilt angle data into the tunnel deformation monitoring model for simulation calculation; Determining the tunnel deformation speed by comparing the simulation calculation results with the actually collected data. 2.The tunnel deformation detection method based on a convolutional neural network and a machine vision sensor according to claim 1, wherein, The application of the convolutional neural network to preprocess the circumferential deformation data and longitudinal deformation data comprises the following steps: Building a preprocessing network comprising an encoder and a decoder, which runs under the TensorFlow deep neural network platform; Using multiple convolutional layers to extract features from images, and performing dimension reduction processing on the extracted features; Outputting a matrix with the size of the original image, confidence and bounding box data through the decoder; Building a detection network composed of a residual network and a fully connected layer to extract features and output results from the preprocessed data. 3.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 1, wherein, The conversion of the cross-section displacement tilt angle data into discretized matrix data and the visual processing based on the theory of differential geometry comprise the following steps: Regarding the tunnel cross-section as a two-dimensional Riemannian manifold embedded in a three-dimensional Euclidean space; Establishing a polar coordinate system (r, θ) with the center of the tunnel as the origin; Expressing the cross-section displacement tilt angle data as a vector field defined on the Riemannian manifold; Calculating the geometric invariants of the Riemannian manifold, including the principal curvatures and the Gaussian curvature; Discretizing the geometric invariants into matrix representation; Performing visual enhancement processing on the matrix data based on the theory of geodesic curvature flow to highlight the deformation abnormal areas. 4.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 1, wherein, The building of the tunnel deformation monitoring model comprises the following steps: Building a deformation prediction system based on convolutional neural networks and machine learning models; The deformation prediction system comprises a feature extraction module, a regression prediction module and a prediction correction module; The feature extraction module inputs the circumferential deformation data, longitudinal deformation data and visualized cross-section displacement tilt angle data as features, classifies, integrates and extracts the features to obtain a feature sequence; The regression prediction module predicts the tunnel deformation based on the feature sequence; The prediction correction module corrects the tunnel deformation based on the historical deformation development law. 5.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 3, wherein, The visual enhancement processing of the matrix data based on the theory of geodesic curvature flow further comprises the following steps: Obtaining cross-section displacement tilt angle data at multiple time points; Calculating the change rate of the cross-section displacement tilt angle based on the connection theory and covariant derivative in differential geometry; Building a change rate feature matrix to extract the main mode of deformation; mapping the geometric invariants matrix and the rate of change features matrix to a color space; applying a multi-scale analysis technique to realize hierarchical visualization from global to local; automatically identifying potential risk areas based on curvature outliers and classifying risk levels. 6.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 1, wherein, The tunnel annular deformation data, longitudinal deformation data and cross-section displacement inclination angle data are collected by machine vision sensors arranged in the tunnel, and the tunnel annular deformation data, longitudinal deformation data and cross-section displacement inclination angle data include: A machine vision sensor is arranged in the tunnel at a preset distance along the longitudinal direction; Each machine vision sensor adopts a face array camera to measure the annular deformation and cracking of the tunnel by image contrast analysis; Data is collected at a preset time interval, and each set of collected data is used as a group of data; The collected data is transmitted to a network terminal server through a network. 7.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 1, wherein, The tunnel deformation speed is determined by comparing the simulation calculation results with the actually collected data, and the tunnel deformation speed includes: A three-dimensional coordinate system is established in the computer, including an X-axis, a Y-axis and a Z-axis, wherein the X-axis is the longitudinal direction of the tunnel, the Y-axis is the annular direction, and the Z-axis is the cross-section direction of the tunnel; The processed annular deformation data, longitudinal deformation data and visualized cross-section displacement inclination angle data are simulated and calculated with the Y-axis, X-axis and Z-axis as the central axes respectively, and simulation data is obtained; Simulation annular deformation images, simulation longitudinal deformation images and simulation cross-section displacement inclination angle images of different collection times are drawn in the three-dimensional coordinate system; The simulation images are compared with the actual images collected at the corresponding time points; When the simulation images and the collected images at the corresponding time points have consistent data changes, it is determined that the tunnel deformation speed at this time is zero; When the simulation images and the collected images at the corresponding time points have differences, the tunnel deformation speed is calculated based on the difference values. 8.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 4, characterized in that, The regression prediction module predicts the tunnel deformation amount according to the feature sequence, and the regression prediction module includes: A historical actual deformation amount corresponding to a feature sequence is obtained; A random forest model is established in a feature space; The feature sequence is classified by the random forest model to obtain a sub-feature sequence; A corresponding relationship between the sub-feature sequence and the actual deformation amount is obtained by training the random forest model; All sub-feature sequences and actual deformation amounts are weighted and reconstructed to obtain a deformation prediction model; The feature sequence is predicted according to the deformation prediction model to obtain a corresponding predicted deformation amount. 9.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 4, characterized in that, The prediction correction module corrects the tunnel deformation amount according to the historical deformation development law, and the prediction correction module includes: The tunnel historical deformation data is predicted to obtain predicted deformation amounts at different times; A prediction error change amount at the current time is obtained according to the predicted deformation amounts at different times; The prediction error change amount is integrated to obtain a change interval of the prediction error; The predicted deformation amount at the current time is corrected according to the change interval of the prediction error; When the change interval of the prediction error is greater than a preset threshold, the tunnel deformation prediction model is updated. 10.The tunnel deformation detection method based on convolutional neural network and machine vision sensor according to claim 1, wherein, Further including: Defining a safety threshold and a risk level standard; Real-time monitoring of the tunnel deformation state and comparison with the safety threshold; When abnormal deformation is detected, triggering an early warning mechanism; providing visualized localization information of the deformation area; generating a detailed report containing the type of deformation, the degree of deformation and the risk assessment.

Images