Tunnel point position change automatic observation method and system combined with AI algorithm

Abstract

The invention provides a tunnel point position change automatic observation method and system combined with an AI algorithm, and relates to the technical field of tunnel monitoring, and the method comprises the steps: employing a speckle projector to form an irregular speckle pattern on a tunnel wall surface, and carrying out the multi-angle collection through a polarized light camera array, and obtaining the speckle intensity data; pre-embedding a coding microstructure generated based on quantum state superposition in the speckle pattern; establishing a speckle-stress correlation analysis model according to the speckle intensity data, and calculating three-dimensional displacement data in combination with the deformation data; and inputting the environmental monitoring data, the stress distribution data and the displacement data into the self-evolution heterogeneous feature network for analysis and prediction. According to the method, the measurement precision is improved by adopting the quantum coding speckles, the dynamic correlation analysis of multi-source data is realized through the self-evolution heterogeneous feature network, and the environmental adaptability and prediction accuracy of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel monitoring, and particularly to an automatic observation method and system for tunnel point position changes combined with AI algorithms. Background Art

[0002] Monitoring the point position changes in tunnel engineering is crucial for ensuring construction and operation safety. Traditional monitoring methods mainly rely on devices such as laser rangefinders and displacement sensors for single-point measurement. With the development of computer vision and artificial intelligence technologies, vision-based non-contact measurement methods have gradually been applied. Currently, speckle interferometry and digital image correlation techniques are widely used for tunnel wall deformation monitoring, and these techniques achieve high-precision displacement measurement by analyzing the changes in surface features.

[0003] However, the existing technologies still have deficiencies. The traditional speckle measurement method is susceptible to factors such as light and temperature in a complex tunnel environment, and it is difficult to guarantee the measurement accuracy and stability; the existing data processing methods often separate stress analysis and displacement measurement, lacking the ability to perform collaborative analysis on multi-source data; a single neural network model is difficult to fully explore the correlation features between different types of monitoring data, and the reliability of the prediction results is limited; the existing systems have poor adaptability to environmental factors and are prone to misjudgment under complex working conditions.

[0004] In summary, the present invention proposes an automatic observation method for point position changes based on quantum-encoded speckles and self-evolving heterogeneous feature networks. By introducing polarized light speckle imaging and quantum state encoding microstructures, the anti-interference ability of the measurement system is improved; a self-evolving heterogeneous feature network is used to achieve dynamic correlation analysis of multi-source data, enhancing the environmental adaptability and prediction accuracy of the system. The present invention provides a new technical solution for point position change monitoring in tunnel engineering and has important practical value. Summary of the Invention

[0005] An embodiment of the present invention provides an automatic observation method and system for tunnel point position changes combined with AI algorithms, which can solve the problems in the existing technologies.

[0006] In the first aspect of the embodiment of the present invention,

[0007] An automatic observation method for tunnel point position changes combined with AI algorithms is provided, including:

[0008] Obtaining visual data of the points to be observed in the tunnel, including the irregular speckle patterns formed by a speckle projector on the tunnel wall;

[0009] Use a polarized light camera array to collect irregular speckle patterns from multiple angles, obtain speckle intensity data in different polarization directions, establish a speckle-stress correlation analysis model based on the speckle intensity data, and obtain the stress distribution data at the points through polarization transmission analysis and complex domain optimization iteration;

[0010] Pre-embed an encoded microstructure generated based on the superposition of quantum states in the irregular speckle pattern, obtain the deformation data of the encoded microstructure, and calculate the three-dimensional displacement data based on the deformation data and the stress distribution data;

[0011] Collect environmental monitoring data, and form an observation data set with the environmental monitoring data, the stress distribution data at the points, and the three-dimensional displacement data;

[0012] Input the observation data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts the data feature correlation through dynamic topology reconstruction and contrast learning to obtain the prediction result of the point change;

[0013] If the prediction result of the point change exceeds the preset change threshold, generate a warning message and send the warning message to the monitoring terminal.

[0014] In an alternative embodiment,

[0015] Establishing a speckle-stress correlation analysis model based on the speckle intensity data and obtaining the stress distribution data at the points through polarization transmission analysis and complex domain optimization iteration includes:

[0016] Normalize and perform Gaussian filtering on the speckle intensity data to obtain filtered intensity data;

[0017] According to the filtered intensity data, calculate the speckle intensity components in the main polarization direction respectively to obtain polarization transmission data, and calculate the second-order partial derivative of the polarization transmission data to obtain local stress direction data;

[0018] According to the local stress direction data, use the complex domain decomposition algorithm to establish a holographic spectrum analysis of the stress field, obtain the main stress components, construct a progressive optimization path through topological homotopy continuous deformation, and dynamically adjust the iteration step size through the path curvature to obtain the initial stress distribution data under the double constraints of elastic equilibrium and speckle intensity fitting;

[0019] Perform stress gradient constraint smoothing processing on the initial stress distribution data to obtain smoothed stress distribution data;

[0020] Calculate the weighted combined value of the speckle intensity fitting error and the stress balance deviation. When the weighted combined value is less than the preset first threshold, adjust the elastic equilibrium constraint weight and recalculate the stress distribution until the weighted combined value meets the first threshold requirement.

[0021] In an alternative embodiment,

[0022] Based on the local stress direction data, a holographic spectrum analysis of the stress field is established using a complex domain decomposition algorithm to obtain the principal stress components. An asymptotic optimization path is constructed through topological homotopy continuous deformation, and the iteration step size is dynamically adjusted by the path curvature. Under the double constraints of elastic equilibrium and speckle intensity fitting, the initial stress distribution data is obtained, including:

[0023] The local stress direction data is mapped to the complex plane through complex exponential transformation to obtain the complex domain representation of the stress field, and the complex domain representation of the stress field includes an amplitude term and a phase term;

[0024] Perform a two-dimensional Fourier transform on the complex domain representation of the stress field to obtain a frequency domain signal, and filter the frequency domain signal using a multi-order band-pass filter to obtain multiple groups of filtered signals;

[0025] Perform singular value decomposition on multiple groups of filtered signals respectively, extract the eigenvectors corresponding to the main eigenvalues, and construct the principal stress components;

[0026] Establish a topological homotopy mapping from the complex domain representation of the stress field to the principal stress components, calculate the path curvature of the topological homotopy mapping, and adaptively adjust the optimization step size according to the path curvature;

[0027] At each optimization step size, calculate the weighted combined objective function of the speckle intensity fitting error and the elastic equilibrium constraint. The speckle intensity fitting error characterizes the difference between the measured intensity and the simulated intensity, and the elastic equilibrium constraint includes the spatial partial derivative combination of the stress components;

[0028] Iteratively update the stress distribution along the topological homotopy mapping path until the difference between the weighted combined objective functions of two adjacent iterations is less than a preset convergence threshold.

[0029] In an alternative embodiment,

[0030] Pre-embed the encoded microstructures generated based on quantum state superposition in the irregular speckle pattern, obtain the deformation data of the encoded microstructures, and calculate the three-dimensional displacement data according to the deformation data and the stress distribution data, including:

[0031] Generate a quantum random sequence based on quantum state superposition, map it to an encoded microstructure through a Gaussian function, and embed the encoded microstructure in the irregular speckle pattern to obtain an encoded speckle pattern;

[0032] Extract the microstructure feature vectors in the encoded speckle pattern, including local gray moments, gradient direction histograms, and topological invariants;

[0033] Establish the mapping relationship of the microstructure feature vector before and after deformation to obtain the feature mapping function, construct and solve a robust matching objective function containing a regularization term based on the feature mapping function to obtain the microstructure deformation data;

[0034] Perform intensity analysis on the encoded speckle pattern to obtain speckle intensity distribution data, establish a speckle-stress correlation analysis model based on the speckle intensity distribution data, and calculate the stress distribution data;

[0035] Construct a displacement field reconstruction equation based on the microstructure deformation data and the stress distribution data. The displacement field reconstruction equation uses a radial basis function as the basic equation, includes a deformation fitting term and a displacement field gradient constraint term, and is subject to elastic equilibrium constraints and displacement boundary condition constraints;

[0036] Solve the displacement field reconstruction equation to obtain the displacement field weight coefficients, and substitute the displacement field weight coefficients into the radial basis function to obtain three-dimensional displacement data.

[0037] In an alternative embodiment,

[0038] Establish the mapping relationship of the microstructure feature vector before and after deformation to obtain the feature mapping function, construct and solve a robust matching objective function containing a regularization term based on the feature mapping function, and the obtained microstructure deformation data includes:

[0039] Calculate the topological stability index of each feature point in the microstructure feature vector, determine the feature point with the highest topological stability index as the main reference point, and divide the feature space into multiple adaptive deformation regions according to the local deformation gradient value centered on the main reference point;

[0040] Calculate the local curvature value for the adaptive deformation region, construct a second-order deformation descriptor, and combine the second-order deformation descriptor with the microstructure feature vector to form an enhanced feature vector;

[0041] Calculate the feature weight coefficients according to the local strain rate of the adaptive deformation region, and assign the feature weight coefficients to each component of the enhanced feature vector to obtain a dynamic feature weight function;

[0042] Use a hierarchical iteration method to calculate the deformation continuity parameter between the adaptive deformation regions;

[0043] Substitute the dynamic feature weight function and the deformation continuity parameter into the feature mapping function, and solve the robust matching objective function containing a regularization term to obtain the microstructure deformation data.

[0044] In an alternative embodiment,

[0045] Input the observation dataset into a preset self-evolving heterogeneous feature network. The self-evolving heterogeneous feature network extracts data feature associations through dynamic topology reconstruction and contrast learning, and the obtained point position change prediction results include:

[0046] Construct a feature association graph, map the point stress distribution data, three-dimensional displacement data, and environmental monitoring data in the observation dataset to the nodes in the graph, and establish dynamic connection weights based on the mutual information measure between the data.

[0047] Perform an adaptive pruning operation on the feature association graph, calculate the conditional mutual information between node pairs, and delete the connection relationships with conditional mutual information lower than the adaptive threshold, where the adaptive threshold is dynamically adjusted according to historical prediction errors.

[0048] Adopt a grouped contrast learning method, randomly sample multiple groups of subgraphs from the feature association graph, apply different intensities of data augmentation operations to each group of subgraphs to generate contrast samples, and train the feature extractor by maximizing the mutual information between the subgraphs.

[0049] Apply the trained feature extractor to the complete feature association graph, obtain the latent representations of the nodes in the graph, and calculate the dynamic influence coefficients between the nodes based on the attention mechanism.

[0050] Adaptively aggregate the latent representations of the nodes according to the dynamic influence coefficients, and use a non-linear transformer to map the aggregated features to the prediction space to obtain the point position change prediction results.

[0051] In the second aspect of the embodiments of the present invention,

[0052] Provide an automatic tunnel point position change observation system combined with an AI algorithm, including:

[0053] The first unit is used to obtain visual data of the points to be observed in the tunnel, including the irregular speckle patterns formed by the speckle projector on the tunnel wall surface.

[0054] The second unit is used to collect the irregular speckle patterns from multiple angles by using a polarized light camera array, obtain the speckle intensity data in different polarization directions, establish a speckle-stress correlation analysis model based on the speckle intensity data, and obtain the point stress distribution data through polarization transmission analysis and complex domain optimization iteration.

[0055] The third unit is used to pre-embed the coding microstructures generated based on the superposition of quantum states in the irregular speckle patterns, obtain the deformation data of the coding microstructures, and calculate the three-dimensional displacement data according to the deformation data and the stress distribution data.

[0056] The fourth unit is used to collect environmental monitoring data, and form an observation dataset with the point stress distribution data and the three-dimensional displacement data.

[0057] The fifth unit is used to input an observation data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts data feature associations through dynamic topology reconstruction and contrast learning to obtain a prediction result of point position change;

[0058] The sixth unit is used to generate a warning message if the prediction result of point position change exceeds a preset change threshold, and send the warning message to a monitoring terminal.

[0059] In the third aspect of the embodiments of the present invention,

[0060] There is provided an electronic device, including:

[0061] A processor;

[0062] A memory for storing instructions executable by the processor;

[0063] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0064] In the fourth aspect of the embodiments of the present invention,

[0065] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0066] In the embodiments of the present invention, automatic observation of tunnel point positions can be realized, the efficiency and accuracy of observation are improved, the need for manual intervention is reduced, thereby reducing human errors; through multi-angle acquisition and polarized light analysis, stress distribution data inside the tunnel can be obtained more comprehensively, enhancing the ability to evaluate the structural safety of the tunnel; by combining environmental monitoring data and the application of a self-evolving heterogeneous feature network, point position changes can be predicted in a timely manner and warning messages can be generated, which helps to discover potential safety hazards in advance and ensure the safe operation of the tunnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic flowchart of a method for automatic observation of tunnel point position changes in combination with an AI algorithm according to an embodiment of the present invention;

[0068] Figure 2 It is a diagram of eigen-decomposition of main components of a stress field;

[0069] Figure 3 It is a diagram of the topological homotopy mapping optimization process and stress field evolution;

[0070] Figure 4 It is a comparison diagram of displacement and strain measurement accuracies of a microstructure deformation analysis method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0072] The following will detail the technical solutions of the present invention with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0073] Figure 1 The following is a schematic flowchart of the automatic observation method for tunnel point position changes in combination with the AI algorithm in the embodiments of the present invention. As Figure 1 shown, the method includes:

[0074] Obtain visual data of the points to be observed in the tunnel, including the irregular speckle patterns formed by the speckle projector on the tunnel wall surface;

[0075] Use a polarization camera array to collect the irregular speckle patterns from multiple angles, obtain speckle intensity data in different polarization directions, establish a speckle-stress correlation analysis model based on the speckle intensity data, and obtain point stress distribution data through polarization transmission analysis and complex domain optimization iteration;

[0076] Pre-embed the encoded microstructures generated based on the superposition of quantum states in the irregular speckle patterns, obtain the deformation data of the encoded microstructures, and calculate three-dimensional displacement data based on the deformation data and stress distribution data;

[0077] Collect environmental monitoring data, and form an observation data set with the point stress distribution data and three-dimensional displacement data;

[0078] Input the observation data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts data feature correlations through dynamic topology reconstruction and contrast learning to obtain point change prediction results;

[0079] If the point change prediction result exceeds a preset change threshold, generate a warning message and send the warning message to the monitoring terminal.

[0080] In an alternative embodiment, establishing a speckle-stress correlation analysis model based on the speckle intensity data and obtaining point stress distribution data through polarization transmission analysis and complex domain optimization iteration includes:

[0081] Perform normalization and Gaussian filtering on the speckle intensity data to obtain filtered intensity data;

[0082] According to the filtering intensity data, calculate the speckle intensity components in the main polarization direction respectively to obtain polarization transmission data, and calculate the second-order partial derivative of the polarization transmission data to obtain local stress direction data;

[0083] According to the local stress direction data, establish a holographic spectrum analysis of the stress field by using a complex domain decomposition algorithm to obtain the principal stress components, construct an asymptotically optimized path through topological homotopy continuous deformation, and dynamically adjust the iteration step size through the path curvature to obtain the initial stress distribution data under the double constraints of elastic equilibrium and speckle intensity fitting;

[0084] Perform stress gradient constraint smoothing processing on the initial stress distribution data to obtain smoothed stress distribution data;

[0085] Calculate the weighted combined value of the speckle intensity fitting error and the stress balance deviation. When the weighted combined value is less than a preset first threshold, adjust the elastic equilibrium constraint weight and repeat the calculation of the stress distribution until the weighted combined value meets the requirements of the first threshold.

[0086] In a specific embodiment, normalize and perform Gaussian filtering on the acquired speckle intensity data. The collected original speckle intensity data usually has noise and uneven distribution. To ensure the accuracy of subsequent analysis, it is necessary to first normalize the original data. The normalization process maps all speckle intensity values to between 0 and 1. The specific operation is to divide the intensity value of each pixel point by the maximum intensity value in the entire image. For example, if the maximum intensity value in the original data is 4096 and the intensity value of a certain point is 2048, then the intensity value of this point after normalization is 0.5. After normalization, use a Gaussian filter to smooth the data. The standard deviation of the filter is set to 3 pixels, and the kernel size is 15×15 pixels. Through Gaussian filtering, high-frequency noise can be effectively suppressed while retaining the main structural features of the speckle image. In practical applications, for an 8-bit grayscale image, the standard deviation of the speckle intensity before filtering is about 42.5, and it drops to 11.8 after filtering, with a significant noise suppression effect.

[0087] Calculate the speckle intensity components in the main polarization direction according to the filtering intensity data to obtain polarization transmission data. Using the filtered intensity data, extract the intensity components in four polarization directions of 0°, 45°, 90°, and 135° respectively. Taking the 0° direction as an example, the intensity component in this direction is obtained by multiplying the filtered intensity value of the corresponding pixel point by a pre-calibrated polarization transmittance coefficient (usually a value between 0.85 and 0.95). After normalizing the intensity components in the four directions, construct a polarization transmission matrix. Each element of this matrix represents the light intensity transmittance at a specific position and a specific polarization direction. For example, in a test sample, the polarization transmittance in the 0° direction in the central region is 0.92, while it drops to 0.78 in the edge region, indicating the existence of stress distribution differences.

[0088] Calculate the second-order partial derivatives of the polarization transmission data to obtain the local stress direction data. The central difference method is used to calculate the first-order derivatives of the polarization transmission data in the x and y directions, and then the second-order partial derivatives are calculated based on the first-order derivatives. In numerical implementation, for the pixel at position (i, j) in the image, the calculation method of the second-order partial derivative in the x direction is: subtract the first-order derivative value at position (i - 1, j) from the first-order derivative value at position (i + 1, j), and then divide by 2. Similarly, calculate the second-order partial derivatives in the y direction and the cross direction. Through the analysis of the second-order partial derivatives, the local stress principal direction can be determined. In experimental verification, in the tensile test of the aluminum alloy specimen, the deviation between the stress direction and the loading direction is less than 3.2°, verifying the accuracy of this method.

[0089] Based on the local stress direction data, use the complex domain decomposition algorithm to establish the holographic spectrum analysis of the stress field and obtain the stress principal components. This step first maps the local stress direction data to the complex plane through complex exponential transformation to form a complex-valued stress field representation. In the complex domain, the stress field can be decomposed into the principal stress component and the secondary stress component, and the two correspond to different eigenvalues respectively. By performing singular value decomposition on the complex-valued representation, the eigenvector corresponding to the principal stress component is extracted. In the test case, the principal stress component usually accounts for more than 85% of the total stress, and for the uniformly loaded sample, this ratio can reach 94.7%.

[0090] Construct an asymptotically optimized path through topological homotopy continuous deformation, establish a continuous deformation path from the initial estimate to the target stress distribution, and ensure that the topological structure of the stress field does not mutate during the optimization process. Use the parameter t ∈ [0, 1] to represent the degree of deformation, where t = 0 corresponds to the initial estimate and t = 1 corresponds to the target state. Each time the iteration is updated, the increment of t is dynamically adjusted according to the path curvature. The greater the curvature, the smaller the step size, to ensure the stability of the optimization. In practical applications, the initial step size is set to 0.05, the minimum step size is 0.01, and the maximum step size is 0.1. For the region with a more complex stress distribution, the step size will automatically decrease to 0.01 - 0.03, while in the region with a gentle stress gradient, the step size can increase to 0.08 - 0.1, effectively balancing the optimization speed and stability.

[0091] Iterative optimization is carried out under the double constraints of elastic equilibrium and speckle intensity fitting to obtain the initial stress distribution data. The elastic equilibrium constraint ensures that the stress distribution satisfies the mechanical equilibrium equation, and the speckle intensity fitting constraint ensures that the calculated stress field can reproduce the observed speckle image. The two constraints are balanced by a weight coefficient, and the initial weight ratio is set to 3:7 (elastic equilibrium: speckle fitting). During the iteration process, when the speckle fitting error is reduced to less than 30% of the initial error, the weight ratio is adjusted to 5:5 to further balance the two constraints. Experimental data show that after about 50 - 80 iterations, the speckle fitting error can be reduced to less than 15% of the initial error, and the stress balance residual is reduced to less than 0.05 MPa.

[0092] The initial stress distribution data is smoothed by stress gradient constraint to obtain the smoothed stress distribution data. The stress gradient constraint smoothing process uses the anisotropic diffusion filtering method, which smooths the noise and pseudo-structures in the stress distribution while maintaining the stress mutation boundary. The filtering intensity is controlled by the local stress gradient magnitude. The larger the gradient, the smaller the filtering intensity. In implementation, for regions with a gradient greater than 5 MPa / mm, the filtering coefficient is set to 0.2, while for regions with a gradient less than 1 MPa / mm, the filtering coefficient is 0.8. The smoothing process is usually carried out for 3 - 5 iterations, and the local gradient is recalculated and the filtering coefficient is adjusted after each iteration. After smoothing, the noise level of the stress distribution is reduced by about 70%, while maintaining the sharpness of the stress boundary.

[0093] Calculate the weighted combined value of the speckle intensity fitting error and the stress balance deviation. When the weighted combined value is less than the preset first threshold, adjust the elastic equilibrium constraint weight and recalculate the stress distribution. The weighted combined value is calculated as: the relative value of the speckle fitting error multiplied by 0.6 plus the relative value of the stress balance deviation multiplied by 0.4. The preset first threshold is set to 0.15, indicating that a 15% comprehensive error is allowed. When the weighted combined value is greater than the threshold, increase the elastic equilibrium constraint weight, gradually increasing from the initial value of 0.3 to 0.5 or 0.6, while decreasing the speckle fitting constraint weight. In the carbon fiber composite material test, the initial weighted combined value is 0.38. After about 15 weight adjustments and recalculations, the final weighted combined value is reduced to 0.13, meeting the preset threshold requirements.

[0094] In this embodiment, by normalizing the speckle intensity and performing Gaussian filtering, noise and background interference are effectively suppressed, and the signal-to-noise ratio is improved; by extracting the speckle component in the main polarization direction and calculating its second-order partial derivative, the accurate positioning of the local stress direction is achieved; by constructing a progressive optimization path using the topological homotopy continuous deformation algorithm based on the decomposition in the complex domain and dynamically adjusting the iteration step size according to the path curvature, the initial stress distribution converges rapidly under the double constraints of elastic equilibrium and speckle fitting; by performing constrained smoothing on the stress gradient and dynamically adjusting the weight based on the weighted combination threshold of the speckle fitting error and the stress balance deviation, the smooth transition and fine fitting of the stress field are realized, ensuring that the result is stable and close to the global optimum.

[0095] In an alternative embodiment, according to the local stress direction data, a holographic spectrum analysis of the stress field is established using the complex domain decomposition algorithm to obtain the stress principal components. A progressive optimization path is constructed through topological homotopy continuity, and the iteration step size is dynamically adjusted through the path curvature. The initial stress distribution data obtained under the double constraints of elastic equilibrium and speckle intensity fitting includes:

[0096] The local stress direction data is mapped to the complex plane through complex exponential transformation to obtain the complex domain representation of the stress field, and the complex domain representation of the stress field includes an amplitude term and a phase term;

[0097] Perform a two-dimensional Fourier transform on the complex domain representation of the stress field to obtain a frequency domain signal, and filter the frequency domain signal using a multi-order band-pass filter to obtain multiple groups of filtered signals;

[0098] Perform singular value decomposition on each group of filtered signals respectively, extract the eigenvectors corresponding to the main eigenvalues, and construct the stress principal components;

[0099] Establish a topological homotopy mapping from the complex domain representation of the stress field to the stress principal components, calculate the path curvature of the topological homotopy mapping, and adaptively adjust the optimization step size according to the path curvature;

[0100] At each optimization step size, calculate the weighted combined objective function of the speckle intensity fitting error and the elastic equilibrium constraint. The speckle intensity fitting error characterizes the difference between the measured intensity and the simulated intensity, and the elastic equilibrium constraint includes the spatial partial derivative combination of the stress components;

[0101] Iteratively update the stress distribution along the topological homotopy mapping path until the difference between the weighted combined objective functions of two adjacent iterations is less than a preset convergence threshold.

[0102] In a specific embodiment, the local stress direction data is mapped to the complex plane through complex exponential transformation to obtain the complex domain representation of the stress field; for the local stress direction angle θ(x, y) at the spatial position (x, y), it is mapped to the complex number Z(x, y) through complex exponential transformation, where the real part is cosθ(x, y) and the imaginary part is sinθ(x, y). This representation method encodes the angle information as points on the unit circle, making the stress field have the characteristics of continuity and periodicity. For example, for a point with a local stress direction angle of 30°, its complex representation is 0.866 + 0.5i; for a point with an angle of 120°, its complex representation is -0.5 + 0.866i. This representation includes an amplitude term and a phase term. The amplitude term is usually normalized to 1, representing a unit vector, while the phase term corresponds to the stress direction angle. In practical applications, for the tensile test of an aluminum alloy specimen, the complex representation of the stress direction in the central region is concentrated around 0.98 + 0.2i, indicating that the stress direction in this region is close to 0°, which is consistent with the expected loading direction.

[0103] Perform a two-dimensional Fourier transform on the complex domain representation of the stress field to obtain a frequency domain signal. This step converts the stress distribution in the spatial domain to a frequency domain representation, facilitating subsequent filtering and decomposition operations. The two-dimensional Fourier transform is implemented through the fast Fourier transform (FFT) algorithm. For a stress field image of 512×512 pixels, the FFT calculation time is approximately 0.08 seconds. In the frequency domain representation, the low-frequency part corresponds to the overall trend of the stress field, while the high-frequency part contains local details and noise information.

[0104] Filter the obtained frequency domain signal using a multi-order band-pass filter to obtain multiple sets of filtered signals. The multi-order band-pass filter is designed to capture stress characteristics at different scales respectively. Five sets of band-pass filters are used, and the center frequencies are set to 1 / 64, 1 / 32, 1 / 16, 1 / 8, and 1 / 4 of the image size respectively, and the bandwidth coefficient is 0.5. For example, for an image of 512×512 pixels, the center frequency of the first set of filters is 8, and the frequency band range is 4 - 12; the center frequency of the second set of filters is 16, and the frequency band range is 8 - 24. Through these filters, the characteristics of the stress field at different spatial scales can be separated. In the carbon fiber composite material test, the first set of filtered signals mainly captured the overall stress distribution trend, with the maximum amplitude being 0.92; the third set of filtered signals reflected the stress changes at a medium scale, and the amplitude dropped to 0.45; while the fifth set of filtered signals contained microscopic scale characteristics, and the amplitude further dropped to 0.12.

[0105] Perform singular value decomposition on multiple sets of filtered signals, extract the eigenvectors corresponding to the main eigenvalues, and construct the principal stress components. Singular value decomposition decomposes each set of filtered signals into the product of three matrices, where the diagonal matrix contains the singular values, sorted from largest to smallest. Usually, the eigenvectors corresponding to the largest 2-3 singular values are selected to form the principal components at that scale. For uniformly loaded materials, the first singular value usually accounts for more than 75% of the total energy; for complex loading conditions, more singular values may need to be considered. In the test of aluminum alloy specimens, the proportions of the first three singular values of the first set of filtered signals are 82.3%, 11.5%, and 4.7% respectively, indicating that the principal components have obvious dominance. The principal components at each scale are combined through weighted combination to form the complete principal stress components, and the weight coefficients are proportional to the corresponding singular values.

[0106] Establish a topological homotopy mapping from the complex domain representation of the stress field to the principal stress components, calculate the path curvature of the topological homotopy mapping, and adaptively adjust the optimization step size according to the path curvature. The topological homotopy mapping is defined as the continuous deformation path S(t) from the initial state S0 to the target state S1, where t is a parameter with a value range of 0 to 1. The path curvature is calculated as the magnitude of the second derivative of S(t) with respect to the parameter t, reflecting the degree of bending of the path. Discretely calculate the curvature of three adjacent points and dynamically adjust the step size according to the curvature value. When the path curvature is less than 0.1, the step size is set to the maximum value of 0.1; when the curvature is between 0.1 and 0.5, the step size linearly decreases to 0.05; when the curvature is greater than 0.5, the step size further decreases to 0.02. This adaptive adjustment mechanism ensures rapid progress in the flat region of the path and cautious advancement in the sharply changing region, improving the efficiency and stability of the optimization. In a complete optimization process, the average value of the step size is about 0.063, the minimum value is 0.018, the maximum value is 0.1, and the total number of iterations is controlled within 40 times.

[0107] At each optimization step size, calculate the weighted combination objective function of the speckle intensity fitting error and the elastic equilibrium constraint. The speckle intensity fitting error is defined as the root mean square error between the measured intensity and the simulated intensity, and the initial value is usually between 0.3 - 0.5 after normalization. The elastic equilibrium constraint includes the combination of spatial partial derivatives of stress components, specifically manifested as the residual of the stress equilibrium equation. In the implementation, the finite difference method is used to calculate the partial derivatives of stress components with respect to spatial coordinates and calculate the residual of the resultant force balance. The two constraints are combined through weight coefficients α and β, with the initial setting of α = 0.6 and β = 0.4, which can be dynamically adjusted during the iteration process. The calculation of the objective function requires updating the stress distribution and recalculating the speckle intensity and elastic equilibrium residuals in each iteration, with a large computational amount. To improve efficiency, a graphics processing unit can be used to accelerate the calculation, reducing the calculation time from 1.2 seconds on the CPU to 0.15 seconds.

[0108] Iteratively update the stress distribution along the topological homotopy mapping path until the difference between the weighted combined objective functions of two adjacent iterations is less than the preset convergence threshold. Use the gradient descent method to update the stress distribution along the negative gradient direction of the objective function, and the step size is dynamically adjusted by the aforementioned path curvature. The preset convergence threshold is set to 0.001, indicating that convergence is considered when the relative change of the objective function is less than 0.1%. In the test, in most cases, the algorithm converges after 25 - 35 iterations, and finally the speckle intensity fitting error is reduced to less than 15% of the initial value, and the elastic equilibrium residual is reduced to less than 0.03 MPa. For carbon fiber composites with complex stress distributions, more iterations may be required for convergence, but usually no more than 50 times.

[0109] Traditional stress field reconstruction methods are mainly based on the photoelastic principle and phase analysis technology, and have problems such as low accuracy, poor noise resistance, and inability to handle complex stress distributions. Existing technologies usually adopt direct analytical methods or simple iterative methods, without considering the topological characteristics and multi-scale structures of the stress field, resulting in easy entrapment in local optimal solutions when dealing with complex materials. In view of these problems, this embodiment introduces complex domain representation and topological homotopy mapping technology, starting from the geometric essence of the stress field, and realizes high-precision reconstruction of the stress field through multi-scale analysis and adaptive optimization strategies. In particular, the introduction of topological homotopy mapping ensures the continuous change of the stress field during the optimization process, avoiding jumps and instabilities in traditional methods; while the path curvature adaptive step size adjustment mechanism greatly improves the convergence speed and robustness of the algorithm. Compared with traditional methods, the technical solution of this embodiment improves the stress reconstruction accuracy, especially reduces the reconstruction error in the complex stress distribution area, and at the same time has better noise suppression ability and computational efficiency, providing reliable technical support for material mechanical property analysis and structural safety assessment.

[0110] Such as Figure 2As shown, three main eigenvectors extracted by singular value decomposition of the stress field are presented. The first principal component (eigenvalue proportion 63.7%, eigenvalue 847.3) shows an overall gradient distribution, presenting a smooth transition from dark blue to light blue, reflecting the basic load structure of the stress field and mainly capturing the macroscopic change trend of the stress field. The second principal component (eigenvalue proportion 23.5%, eigenvalue 312.6) presents an annular stress concentration structure, with green as the main color tone, forming an obvious annular distribution in the center of the image, characterizing the stress concentration phenomenon in certain areas. This pattern usually corresponds to material boundaries or structural transition regions. The third principal component (eigenvalue proportion 8.4%, eigenvalue 111.8) shows a more complex local shear feature, with red as the main color tone, presenting an asymmetric distribution pattern and capturing local disturbances and directional changes in the stress field. These three principal components together account for 95.6% of the total energy of the stress field, fully demonstrating that the solution of this embodiment can effectively identify and separate the main structural features of the stress field through complex domain decomposition and feature extraction, providing a reliable feature basis for subsequent stress field reconstruction.

[0111] As Figure 3 shown, the whole process of stress field optimization based on topological homotopy mapping is presented. From left to right, there are four key stages: initial estimate, intermediate stage 1, intermediate stage 2, and final result. The initial estimate stage (root mean square error 7.62 MPa) shows a blurred and noisy stress distribution, with a curvature as high as 0.28. Therefore, a small step size of 0.05 is adopted for cautious optimization. In intermediate stage 1 (12 iterations, root mean square error reduced to 4.35 MPa), stress concentration areas begin to appear, the curvature drops to 0.17, and the step size increases to 0.14, accelerating the convergence process. Intermediate stage 2 (27 iterations, root mean square error further reduced to 2.78 MPa) shows a clearer multi-region stress concentration structure, the curvature drops to 0.10, and the step size increases to 0.21, and the optimization enters the fast convergence stage. The final result (43 iterations, root mean square error only 1.84 MPa) presents an accurate and clear stress distribution, including three obvious stress concentration areas, the curvature drops to 0.07, and the step size is adjusted to 0.08 for fine adjustment. The whole optimization process fully reflects the core advantages of topological homotopy mapping: avoiding the jump problem of traditional methods through continuous deformation paths, and at the same time dynamically adjusting the step size using path curvature, improving the optimization efficiency while ensuring convergence stability.

[0112] In an alternative embodiment, a coded microstructure generated based on quantum state superposition is pre-embedded in an irregular speckle pattern, deformation data of the coded microstructure is obtained, and three-dimensional displacement data is calculated according to the deformation data and stress distribution data, including:

[0113] Generate a quantum random sequence based on the superposition of quantum states, map it to an encoded microstructure through a Gaussian basis function, and embed the encoded microstructure in an irregular speckle pattern to obtain an encoded speckle pattern;

[0114] Extract the microstructure feature vectors from the encoded speckle pattern, including local gray moments, histogram of gradient directions, and topological invariants;

[0115] Establish the mapping relationship of the microstructure feature vectors before and after deformation to obtain a feature mapping function, construct and solve a robust matching objective function containing a regularization term based on the feature mapping function to obtain microstructure deformation data;

[0116] Conduct intensity analysis on the encoded speckle pattern to obtain speckle intensity distribution data, establish a speckle-stress correlation analysis model based on the speckle intensity distribution data, and calculate the stress distribution data;

[0117] Construct a displacement field reconstruction equation based on the microstructure deformation data and the stress distribution data. The displacement field reconstruction equation uses a radial basis function as the basic equation, includes a deformation fitting term and a displacement field gradient constraint term, and is subject to elastic equilibrium constraints and displacement boundary condition constraints;

[0118] Solve the displacement field reconstruction equation to obtain the displacement field weight coefficients, and substitute the displacement field weight coefficients into the radial basis function to obtain three-dimensional displacement data.

[0119] In a specific embodiment, a quantum random sequence is generated based on the superposition of quantum states, mapped into an encoded microstructure through a Gaussian function, and the encoded microstructure is embedded in an irregular speckle pattern to obtain an encoded speckle pattern. The generation of the quantum random sequence utilizes the superposition state characteristics of quantum bits. A 2×2×...×2 dimensional quantum state space is constructed through a quantum circuit, and a truly random sequence is obtained through measurement collapse. A 10-qubit system is used to generate a random sequence with a length of 1024, and the value range of each sequence element is between 0 and 1. The quantum random sequence has better randomness compared to traditional pseudo-random sequences. Its autocorrelation coefficient is less than 0.01, and the passing rate of the NIST randomness test reaches 98.5%. The generated random sequence is mapped into an encoded microstructure in the spatial domain through a Gaussian function. The standard deviation of the Gaussian function is set to 5 pixels, and the amplitude is controlled by the random sequence elements. For example, for a random sequence element value of 0.75, the corresponding Gaussian peak is 0.75 times the maximum gray value of 255, that is, 191. This mapping preserves the statistical characteristics of the random sequence and at the same time forms spatially continuous microstructure features. The encoded microstructure is embedded in the irregular speckle pattern at a ratio of 15%. The embedding rule is to replace specific regions in the speckle pattern with the encoded microstructure while maintaining the overall shape of the original speckle. The embedded encoded speckle pattern maintains the random distribution characteristics of the speckle macroscopically, while containing traceable encoded features microscopically. The entropy value of the gray distribution increases from 7.54 of the traditional speckle to 7.82, indicating an increase in information content.

[0120] Extract the microstructure feature vectors in the encoded speckle pattern, including local gray moments, gradient direction histograms, and topological invariants. The feature extraction process uses a sliding window technique. The window size is set to 32×32 pixels, and the step size is 8 pixels. For each window, the first four-order local gray moments are calculated, including the mean, variance, skewness, and kurtosis. These four statistics respectively describe the central position, dispersion degree, symmetry, and sharpness of the gray distribution. In the test of aluminum alloy specimens, the third-order moment (skewness) in the encoded area is about 0.3 higher than that in the non-encoded area, indicating that the encoding enhances the asymmetry of the gray distribution. The gradient direction histogram forms an 8-dimensional feature vector by calculating the gradient direction of each pixel point and statistically counting the gradient occurrence frequencies in 8 direction intervals. The topological invariants include the number of connected regions, Euler number, and Betti number. These features have good stability during the deformation process. In a typical window, the number of connected regions is 24, and the Euler number is -5. These values change by no more than 10% under small deformations. Finally, the dimension of the feature vector extracted from each window is 16 (4 + 8 + 4). To improve the calculation efficiency, principal component analysis is used to reduce the feature vector to 10 dimensions, retaining 95.3% of the original information.

[0121] The mapping relationship of the microstructure feature vectors before and after deformation is established to obtain the feature mapping function. Based on the feature mapping function, a robust matching objective function containing a regularization term is constructed and solved to obtain the microstructure deformation data. The feature mapping function adopts the weighted combination form of neighboring feature points, and the weight coefficient is inversely proportional to the feature space distance. For each feature point before deformation, the most similar candidate point set is searched in the feature space after deformation, and the similarity threshold is set to 0.85. The robust matching objective function includes a feature distance term and a deformation smooth regularization term, and the regularization coefficient is dynamically adjusted according to the feature point density, usually set between 0.2 and 0.5. The objective function is solved using the alternating direction method of multipliers, and the convergence condition is that the relative residual is less than 0.001. In the tensile test of carbon fiber composite materials, the feature point matching accuracy reaches 94.6%, and the average position error is less than 0.3 pixels, which is significantly better than the 1.2-pixel error of the traditional speckle correlation method. The finally obtained microstructure deformation data includes the displacement vector field in three-dimensional space, and the number of sampling points is usually 1 / 64 of the image size. For example, for an image of 2048×2048 pixels, a 32×32 displacement sampling grid is formed.

[0122] The intensity analysis of the encoded speckle pattern is carried out to obtain the speckle intensity distribution data. Based on the speckle intensity distribution data, a speckle-stress correlation analysis model is established, and the stress distribution data is calculated. The speckle intensity analysis adopts a multi-scale decomposition method. The speckle image is decomposed into sub-bands of different scales through wavelet transform, and the spatial distribution characteristics of the speckle intensity are extracted. For an image with 10-bit precision, the range of speckle intensity values is usually between 0 and 1023, and after normalization, it is transformed into the relative intensity between 0 and 1. The speckle-stress correlation analysis model is based on the photoelastic principle to establish a quantitative relationship between the speckle intensity and the material stress state. The model considers the photoelastic coefficient of the material, the optical path length, and the polarization state of the incident light, and is solved by an iterative fitting algorithm. For a transparent epoxy resin specimen, the photoelastic coefficient is 3.6×10 -6 MPa -1 MPa, and the fitting accuracy reaches ±2.5 MPa. The stress distribution data includes the magnitude and direction of the principal stress, and the spatial resolution is the same as that of the speckle image. A smooth and continuous stress field distribution can be obtained through bicubic interpolation. In the test of metal materials, the surface stress distribution is indirectly measured by coating a transparent photoelastic coating on the surface, and the coating thickness is controlled between 0.2 - 0.3 mm to ensure the balance of sensitivity and spatial resolution.

[0123] A displacement field reconstruction equation is constructed based on microstructural deformation data and stress distribution data. The displacement field reconstruction equation uses radial basis functions as the basic equation, including a deformation fitting term and a displacement field gradient constraint term, and is subject to elastic equilibrium constraints and displacement boundary condition constraints. The radial basis function selects a Gaussian kernel function, and the support radius is set to 1 / 10 to 1 / 8 of the deformation region. For a 100mm×100mm test area, the support radius is set to 10mm. The deformation fitting term requires that the reconstructed displacement field is consistent with the microstructural deformation data at the sampling points, and the weight coefficient is set to 0.6. The displacement field gradient constraint term controls the smoothness of the displacement change in adjacent regions, and the weight coefficient is set to 0.4. The elastic equilibrium constraint ensures that the reconstructed displacement field conforms to the mechanical equilibrium equation, which is introduced by the Lagrange multiplier method. The initial value of the multiplier is set to 0.1 and is dynamically adjusted during the iteration process. The displacement boundary conditions are set according to the actual test situation. Usually, a zero displacement boundary is set at the fixed end of the specimen, and a known displacement or free boundary is set at the loading end. The displacement field reconstruction equation is finally transformed into a system of linear equations, the size of which is related to the number of radial basis function nodes. For a 32×32×3 displacement field, the dimension of the system of equations is about 3000×3000, and the solution time is controlled within 2 seconds on a modern computer.

[0124] The displacement field weight coefficients are obtained by solving the displacement field reconstruction equation, and the three-dimensional displacement data is obtained by substituting the displacement field weight coefficients into the radial basis function. The preconditioned conjugate gradient method is used to solve the displacement field reconstruction equation, and the preconditioned matrix selects incomplete Cholesky decomposition. The iterative convergence condition is that the relative residual is less than 10 -6 . For typical engineering problems, the number of iterations is usually between 200 and 300, and the solution accuracy reaches the micron level. The displacement field weight coefficient is a control parameter of the radial basis function, and its dimension is equal to the number of basis function nodes multiplied by the spatial dimension (in the X, Y, and Z directions). For a three-dimensional problem with a 32×32 node grid, the total number of weight coefficients is 3072. Substituting the obtained weight coefficients into the radial basis function, a continuous three-dimensional displacement field function can be obtained, and then the displacement value can be calculated at any position. For example, in the bending test of a composite laminate, the maximum displacement in the Z direction reaches 4.76mm, the error between the calculated value and the measured value by a laser displacement sensor is less than 0.08mm, and the relative error is 1.7%. The complete three-dimensional displacement data is stored in the form of a three-dimensional matrix, and the resolution can be improved to the original image resolution by interpolation, achieving a spatial resolution of the micron level and a displacement measurement accuracy of the sub-micron level. In the fatigue test of an aviation aluminum alloy component, this method successfully captured the displacement field singularity near the crack tip, providing key data support for crack propagation prediction.

[0125] In this embodiment, a random sequence is generated through quantum state superposition and mapped into a Gaussian-coded microstructure, realizing the high-entropy unpredictability of the coded speckle pattern; by extracting local gray moment, histogram of gradient directions, and topological invariant feature vectors, a robust description of microstructure deformation is realized; by establishing a feature mapping function and introducing a regularization term to construct a matching objective function, a high-precision solution of microstructure deformation data is realized; by combining the coded speckle intensity distribution with microstructure deformation data, the displacement field reconstruction equation based on radial basis function, and elastic equilibrium and boundary constraints, a synchronous high-precision reconstruction of the three-dimensional displacement field and stress distribution is realized.

[0126] In an alternative embodiment, a mapping relationship between microstructure feature vectors before and after deformation is established to obtain a feature mapping function, and a robust matching objective function containing a regularization term is constructed and solved based on the feature mapping function to obtain microstructure deformation data, including:

[0127] Calculate the topological stability index of each feature point in the microstructure feature vector, determine the feature point with the highest topological stability index as the main reference point, and divide the feature space into multiple adaptive deformation regions according to the local deformation gradient value centered on the main reference point;

[0128] Calculate the local curvature value for the adaptive deformation region, construct a second-order deformation descriptor, and combine the second-order deformation descriptor with the microstructure feature vector to form an enhanced feature vector;

[0129] Calculate the feature weight coefficient according to the local strain rate of the adaptive deformation region, and assign the feature weight coefficient to each component of the enhanced feature vector to obtain a dynamic feature weight function;

[0130] Adopt a hierarchical iteration method to calculate the deformation continuity parameter between the adaptive deformation regions;

[0131] Substitute the dynamic feature weight function and the deformation continuity parameter into the feature mapping function, and solve the robust matching objective function containing the regularization term to obtain the microstructure deformation data.

[0132] In a specific embodiment, the topological stability index of each feature point in the microstructural feature vector is calculated, and the feature point with the highest topological stability index is determined as the main reference point. Centered on the main reference point, the feature space is divided into multiple adaptive deformation regions according to the local deformation gradient value. The topological stability index is quantified by analyzing the change degree of topological invariants during the deformation process of the feature point, specifically including three aspects: connectivity retention rate, homotopy equivalence, and topological entropy. The connectivity retention rate calculates the change ratio of the number of connected components within the neighborhood of the feature point. The smaller the change, the higher the stability. For a 32×32 pixel feature window, the number of connected components is usually between 15 and 30, and points with a retention rate higher than 90% are considered to have high connectivity stability. Homotopy equivalence measures whether the deformation of the neighborhood of the feature point can be achieved through continuous deformation, and is determined by calculating the difference in the Euler number before and after deformation. A difference of 0 indicates topological equivalence. The topological entropy calculates the change in the information entropy of the neighborhood of the feature point. Points with an entropy value change less than 5% have high information stability. These three indicators are weighted averaged to obtain the comprehensive topological stability index, and the weight ratio is 0.3:0.4:0.3. In the test of carbon fiber composite materials, the topological stability index is distributed between 0.4 and 0.95, and points with an index greater than 0.85 are screened as candidate main reference points, usually accounting for 10%-15% of the total feature points. The point with the highest index is selected from the candidate main reference points as the main reference point. For example, in a test, the topological stability index of the main reference point is 0.93, located at the feature space coordinates (127, 205). Centered on the main reference point, the local gradient value of the deformation field is calculated. The gradient value is calculated by dividing the modulus of the displacement difference between adjacent feature points by the point spacing. According to the magnitude of the gradient value, the K-means clustering algorithm is used to divide the feature space into 3-5 adaptive deformation regions, and the region division threshold is dynamically set according to the maximum gradient value. In the aluminum alloy tensile test, 4 regions are formed. The maximum gradient region is near the specimen notch, with a gradient value of 0.15, and the gradient value in the edge region drops to 0.02.

[0133] Calculate the local curvature value for the adaptive deformation region, construct a second-order deformation descriptor, and combine the second-order deformation descriptor with the microstructure feature vector to form an enhanced feature vector. The local curvature value describes the degree of non-linear deformation in the feature space and is calculated from the principal curvatures of the surface of the feature point distribution. For each deformation region, first construct a local coordinate system based on the feature point positions, with the origin at the center of the region and the coordinate axes along the principal deformation directions. In this coordinate system, fit a quadratic surface by the least squares method to obtain the coefficient matrix of the surface. Calculate the Gaussian curvature and mean curvature at each point from the coefficient matrix. The Gaussian curvature characterizes the intrinsic curvature of the surface, and the mean curvature reflects the extrinsic geometric properties of the surface. In large deformation regions, the Gaussian curvature typically reaches 0.05 - 0.1, while the mean curvature is in the range of 0.2 - 0.3; in small deformation regions, these two indicators drop to below 0.01 and 0.05 respectively. Construct a second-order deformation descriptor based on the curvature values, including 8 components such as the variance, skewness, kurtosis of the curvature distribution, and the principal curvature direction. Combine the second-order deformation descriptor with the original microstructure feature vector to form an enhanced feature vector. If the dimension of the original feature vector is 16, the dimension of the enhanced feature vector is 24. In the bending test of composite laminates, the matching accuracy of the enhanced feature vector is increased by 12.3% compared with the original feature vector. Especially in large deformation regions, the improvement is particularly obvious, and the false matching rate drops from 18.5% to 5.7%.

[0134] Calculate the feature weight coefficient according to the local strain rate of the adaptive deformation region, and assign the feature weight coefficient to each component of the enhanced feature vector to obtain a dynamic feature weight function. The local strain rate is calculated from the second invariant of the strain tensor, and this value reflects the intensity of the deformation. For different deformation regions, dynamically adjust the weight coefficients of each feature component according to the magnitude of the strain rate. The basic principle is: in large strain regions, the weights of topological features and second-order deformation descriptors increase, while the weights of gray scale moments and gradient direction histograms decrease; in small strain regions, it is the opposite. Specifically, when the strain rate is less than 0.02, the weight coefficients of gray scale moments and gradient direction histograms are set to 0.6, and the weight coefficients of topological features and second-order deformation descriptors are 0.4; when the strain rate is between 0.02 - 0.1, the weights of the four types of features are all set to 0.5; when the strain rate is greater than 0.1, the weights of gray scale moments and gradient direction histograms drop to 0.3, while the weights of topological features and second-order deformation descriptors rise to 0.7. The strain rate is estimated and calculated from the displacement field in the initial stage of the experiment and updated during the iteration process. The dynamic feature weight function is defined as the product of each component of the feature vector and the corresponding weight coefficient, forming an adaptive feature metric space. In the fatigue test of aviation aluminum alloy components, the dynamic weight adjustment improves the average accuracy of feature matching from 91.2% to 96.8%. Especially for the high strain region near the crack, the matching accuracy is improved by 8.5 percentage points.

[0135] The deformation continuity parameters between the adaptive deformation regions are calculated using a hierarchical iterative method. The deformation continuity parameters describe the smooth transition degree of the deformation field between adjacent deformation regions, and are measured by the continuity of the first-order derivative and the second-order derivative of the displacement field at the region boundary. The hierarchical iterative method starts from the global deformation and gradually refines to the local deformation. In the first layer of iteration, the entire feature space is regarded as a whole to calculate the global deformation parameters; in the second layer of iteration, the local deformation parameters of each adaptive deformation region are calculated respectively; in the third layer of iteration, the deformation parameters of the transition zone between regions are calculated. A transition zone with a width 3-5 times the feature point spacing is set at the region boundary to constrain the deformation field in the transition zone to satisfy at least C1 continuity (continuous first-order derivative). The deformation field of the transition zone smoothly connects the deformation parameters of adjacent regions through a weight fusion function. The weight function uses a cubic Hermite polynomial to ensure that both the function value and the first-order derivative are continuous at the boundary of the transition zone. For complex deformation fields, C2 continuity (continuous second-order derivative) can be further required. At this time, the weight function is upgraded to a quintic spline function. The hierarchical iteration is usually carried out for 3-5 rounds, and the convergence condition is that the difference between the displacement fields of adjacent two rounds is less than 1%. In the impact test of carbon fiber / epoxy resin composites, the deformation continuity parameters calculated using the hierarchical iterative method reduce the jump of the displacement field at the region boundary from the original 0.23 pixels to 0.04 pixels, significantly improving the smoothness of the reconstructed displacement field.

[0136] Substitute the dynamic feature weight function and the deformation continuity parameter into the feature mapping function to solve the robust matching objective function with a regularization term, and obtain the microstructural deformation data. The feature mapping function defines the correspondence between point pairs in the feature space, is constructed using a radial basis function network, the kernel function is selected as a Gaussian function, and the support radius is set to 2-3 times the average feature point distance. The dynamic feature weight function and the deformation continuity parameter serve as modulation factors of the mapping function, changing the metric structure of the feature space. The robust matching objective function includes a feature distance term, a deformation smoothness regularization term, and an outlier penalty term. The feature distance term measures the similarity between paired feature points and uses the Mahalanobis distance metric; the deformation smoothness regularization term controls the smoothness of the deformation field and is realized through a thin plate spline energy function; the outlier penalty term uses the Huber loss function to give a smaller weight to abnormal matches. The regularization coefficient is set according to the image noise level, usually between 0.1 and 0.5, and the greater the noise, the larger the coefficient. The solution of the objective function uses the alternating direction method of multipliers, which decomposes the original problem into a sequence of sub-problems and solves them iteratively. The initial iteration step size is set to 0.1, dynamically adjusted according to the residual change, and the maximum step size is limited to 0.5. The convergence condition is set to a relative residual less than 0.001 or the maximum number of iterations reaching 100. In practical applications, the solution process usually converges within 20-40 iterations. The obtained microstructural deformation data contains the three-dimensional displacement vector and the deformation gradient tensor of each feature point. In the large deformation test of aerospace titanium alloy components, the displacement measurement accuracy reaches ±0.05 pixels, and the strain measurement accuracy reaches ±200 microstrains, meeting the requirements of high-precision strain analysis.

[0137] The microstructural deformation analysis methods in the prior art are mainly based on digital image correlation technology and feature point matching algorithms, and there are difficulties in dealing with large deformations, discontinuous deformations, and complex materials. Traditional methods usually adopt a globally consistent feature extraction and matching strategy, which cannot adapt to the differences in local deformation characteristics. At the same time, the feature weights are fixed and it is difficult to dynamically adjust according to the degree of deformation. In addition, regional segmentation is often based on simple grid division, ignoring the physical characteristics of the deformation field. These problems lead to a significant decrease in the measurement accuracy in key regions such as large deformation regions, strain concentration regions, and material interfaces. In view of the above problems, this embodiment proposes an adaptive deformation analysis method based on topological stability. By calculating the topological stability index of feature points, the accurate identification of the main reference points with high stability during the deformation process is realized. Based on the adaptive regional division of local deformation gradient values, the regional segmentation is consistent with the physical characteristics of the actual deformation field. The introduction of the second-order deformation descriptor enhances the expression ability of the feature vector for non-linear deformations. The dynamic feature weight function realizes the adaptive adjustment of the importance of feature components. The hierarchical iteration method and the deformation continuity parameter ensure the smooth transition of the deformation field between regions. The high-precision analysis of complex deformation fields is realized. Especially in large deformation regions and strain concentration regions, the method proposed in this application has improved displacement measurement accuracy and strain measurement accuracy compared with traditional methods, providing more reliable data support for the evaluation of material mechanical properties and structural integrity.

[0138] As Figure 4 shown, it shows the comparison of displacement and strain measurement accuracies between the technical solution of the present invention and traditional DIC methods and feature point-based methods under different deformation conditions. It can be clearly seen from the chart that the technical solution of the present invention shows obvious advantages in all test scenarios. In the small deformation region (0 - 5%), the displacement measurement error of the technical solution of the present invention is only 1.2 μm, which is about 45% lower than 2.3 μm of the traditional DIC method and 2.1 μm of the feature point method. As the degree of deformation increases, the advantage becomes more obvious. In the large deformation region (15 - 30%), the displacement error of the technical solution of the present invention is 5.4 μm, while the traditional methods are as high as 18.7 μm and 21.3 μm, and the accuracy is improved by about 75%. Especially in the strain concentration region, the strain measurement error of the technical solution of the present invention is only 0.98%, while the traditional DIC method and the feature point method are 6.84% and 8.92% respectively, and the accuracy improvement reaches more than 85%. In the material interface region, the displacement error of the technical solution of the present invention is 3.5 μm and the strain error is 0.84%, which is also far better than other methods. These data fully prove that the technical solution of the present invention effectively solves the accuracy problem of traditional methods in complex deformation field analysis through innovative mechanisms such as topological stability analysis, adaptive regional division, and dynamic feature weight function.

[0139] In an alternative embodiment, the observed data set is input into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts data feature associations through dynamic topology reconstruction and contrast learning. The obtained point position change prediction result includes:

[0140] Construct a feature association graph, map the point stress distribution data, three-dimensional displacement data, and environmental monitoring data in the observed data set into nodes in the graph, and establish dynamic connection weights based on the mutual information measure between the data;

[0141] Perform an adaptive pruning operation on the feature association graph, calculate the conditional mutual information between node pairs, and delete the connection relationships with conditional mutual information lower than the adaptive threshold, where the adaptive threshold is dynamically adjusted according to the historical prediction error;

[0142] Adopt a grouped contrast learning method, randomly sample multiple groups of subgraphs from the feature association graph, apply different intensities of data augmentation operations to each group of subgraphs to generate contrast samples, and train the feature extractor by maximizing the mutual information between the subgraphs;

[0143] Apply the trained feature extractor to the complete feature association graph, obtain the latent representations of the nodes in the graph, and calculate the dynamic influence coefficients between the nodes based on the attention mechanism;

[0144] Adaptively aggregate the latent representations of the nodes according to the dynamic influence coefficients, and use a non-linear transformer to map the aggregated features to the prediction space to obtain the point position change prediction result.

[0145] In a specific implementation, a feature association graph is constructed. The point stress distribution data, three-dimensional displacement data, and environmental monitoring data in the observation dataset are mapped as nodes in the graph, and dynamic connection weights are established based on the mutual information measure between the data. The construction process of the feature association graph includes three steps: node definition, connection establishment, and weight calculation. The point stress distribution data usually includes the magnitude and direction of the principal stress. For example, the principal stress value at a certain bridge monitoring point is 42.5 MPa, and the direction is 35.2 degrees. The three-dimensional displacement data includes the displacement amounts in the X, Y, and Z directions. For example, the displacement amounts at the same monitoring point are (0.25 mm, 0.17 mm, -0.42 mm). The environmental monitoring data covers environmental factors such as temperature, humidity, and wind speed. For example, the temperature is 28.3 °C, the relative humidity is 65%, and the wind speed is 4.2 m / s. These multi-dimensional data of each monitoring point are respectively represented as independent nodes in the graph. For N monitoring points, a total of N×(n_stress + n_displacement + n_environment) nodes are formed, where n_stress, n_displacement, and n_environment respectively represent the dimensions of stress, displacement, and environmental parameters. The mutual information between nodes is quantified by calculating the probability distribution similarity of two data sequences. The higher the mutual information value, the stronger the correlation between the data. In a specific implementation, a method based on kernel density estimation is used to calculate the mutual information, and the bandwidth parameter is set to 0.15. The calculated mutual information value is used as the weight of the connection between nodes. For example, in the monitoring of a large-scale structure, the mutual information value between the principal stress node and the Z-direction displacement node is 0.78, indicating a strong correlation between the two, while the mutual information value between the principal stress and the environmental temperature node is only 0.23, indicating a weak correlation.

[0146] Perform an adaptive pruning operation on the feature association graph, calculate the conditional mutual information between node pairs, and delete the connection relationships with conditional mutual information lower than the adaptive threshold, where the adaptive threshold is dynamically adjusted according to historical prediction errors. Conditional mutual information measures the degree of information dependence between two nodes given the information of a third node, and it can filter out indirect correlations caused by common factors. The ternary kernel density estimation method is used to calculate the conditional mutual information. For each pair of nodes (i, j), calculate the conditional mutual information value given all other nodes k. In the high-speed railway bridge monitoring system, the mutual information value between the displacement nodes of two relatively distant monitoring points is 0.65 when the temperature effect is not considered, while the conditional mutual information drops to 0.12 under the given temperature condition, indicating that the correlation between the displacements of the two points mainly comes from the common influence of temperature. The adaptive threshold is initially set to 0.15 and then dynamically adjusted according to the prediction error. The prediction error is measured by the root mean square error. When the prediction error increases, the threshold is appropriately reduced to retain more potentially useful connections; when the prediction error decreases, the threshold is appropriately increased to simplify the graph structure. The specific adjustment strategy is as follows: when the prediction error increases by more than 10% compared to the previous cycle, the threshold is reduced by 0.02; when the prediction error decreases by more than 10%, the threshold is increased by 0.02; the threshold range is limited between [0.05, 0.3].

[0147] Adopt the grouped contrastive learning method. Randomly sample multiple groups of subgraphs from the feature correlation graph, apply different intensities of data augmentation operations to each group of subgraphs to generate contrastive samples, and train the feature extractor by maximizing the mutual information between subgraphs. Grouped contrastive learning constructs similar and dissimilar sample pairs to learn the intrinsic representation of data. When sampling subgraphs from the feature correlation graph, a random walk strategy is adopted. Starting from a randomly selected seed node, perform K-step random walks according to the connection weight probability, sample M subgraphs, and each subgraph contains L nodes. In practical applications, K is set to 3 - 5, M is set to 64 - 128, and L depends on the scale of the original graph, usually 10% - 20% of the total number of nodes. Apply data augmentation operations to the sampled subgraphs, including feature perturbation, topological perturbation, and masking operation. Feature perturbation is achieved by adding Gaussian noise, and the noise level is set to 10% of the standard deviation of the original features; topological perturbation randomly deletes or adds edges in the subgraph, and the perturbation rate is 15%; the masking operation randomly sets the features of 20% of the nodes in the subgraph to zero. For the same subgraph, generate two different augmented versions as positive sample pairs; the augmented versions of different subgraphs are used as negative sample pairs. The feature extractor adopts a graph neural network architecture, including 3 layers of graph convolutional layers, the hidden layer dimension is 128, the activation function is LeakyReLU, the learning rate is set to 0.001, and it is trained using the Adam optimizer. The contrastive loss function adopts the InfoNCE form, and the temperature parameter is set to 0.07. In the settlement monitoring system of a certain water conservancy project, the feature extractor trained by grouped contrastive learning improves the prediction accuracy compared with the traditional supervised learning method, especially in the identification of abnormal working conditions, and the accuracy rate is significantly improved.

[0148] Apply the trained feature extractor to the complete feature association graph to obtain the latent representations of the nodes in the graph, and calculate the dynamic influence coefficients between the nodes based on the attention mechanism. The trained feature extractor includes an encoder part. Apply it to the complete feature association graph to get the latent representation of each node. The latent representation is a vector in a low-dimensional space, and the dimension is usually set to 64 or 128, which captures the key feature information of the nodes. In the bridge health monitoring system, the original node feature dimension is about 15 - 20, and after encoding, it is reduced to 64 dimensions, but about 95% of the information is retained. The attention mechanism is used to calculate the dynamic influence coefficients between the nodes. A multi-head self-attention structure is adopted, and the number of heads is set to 8. Each attention head independently learns the node interaction relationships in different patterns. The attention scores are calculated by the dot product of the query vector and the key vector, and after softmax normalization, the influence coefficients between the nodes are obtained. In the specific implementation, for each node i, calculate its attention scores with all other nodes j to form an attention distribution. The attention mechanism can adaptively focus on important nodes. For example, in the monitoring of a certain high-rise building, when the wind speed suddenly increases, the nodes related to wind load obtain an attention score as high as 0.43, while this score is only 0.12 under static conditions. The dynamic influence coefficients reflect the intensity changes of the interaction between nodes under different working conditions, providing guidance for subsequent feature aggregation.

[0149] Adaptive aggregation of the latent representations of the nodes is performed according to the dynamic influence coefficients. A non-linear transformer is used to map the aggregated features to the prediction space to obtain the prediction results of the point position changes. Adaptive aggregation performs a weighted combination of the latent representations of all relevant nodes according to the dynamic influence coefficients to obtain the context-enhanced representation of the target node. A graph attention network structure is adopted, which includes residual connections to retain the original feature information. The dimension of the aggregated features is the same as the original latent representation, but it contains information from relevant nodes. For example, in the health monitoring system of a certain large bridge, the displacement prediction at the midpoint of the main span aggregates information from temperature, stress, and the displacements of adjacent points. When the temperature changes suddenly, the contribution weight of the temperature node increases to 0.38, while it is only 0.15 under normal temperature conditions. The aggregated features are mapped to the prediction space through a non-linear transformer. The transformer consists of a multi-layer perceptron, including two hidden layers with dimensions of 128 and 64 respectively, and the activation function is GELU. The activation function is not used in the last layer to support regression tasks. For classification tasks such as health status judgment, the softmax activation is used in the last layer. The transformer is trained using the mean squared error loss (for regression tasks) or the cross-entropy loss (for classification tasks), the learning rate is 0.0005, and the AdamW optimizer is used with a weight decay coefficient of 0.01. The prediction results are organized according to time windows, generating the change trends for the next 24 hours, 72 hours, and 7 days for each monitoring point, including the predicted values and 95% confidence intervals, supporting engineers in risk assessment and maintenance decision-making.

[0150] Traditional structural health monitoring and prediction methods mainly rely on statistical models and shallow machine learning algorithms, such as time series analysis, regression trees, and support vector machines. These methods usually process various types of monitoring data separately and are difficult to effectively capture the complex interaction relationships between different physical quantities. At the same time, traditional methods have limited ability to integrate multi-source heterogeneous data, resulting in insufficient prediction accuracy under complex working conditions. In addition, although existing deep learning methods have powerful representation capabilities, they often adopt fixed network structures and parameters, lack the adaptive ability to changes in data correlation, and have poor stability during long-term monitoring. In this embodiment, a multi-source data fusion prediction method based on a graph structure is introduced. Different types of monitoring data are uniformly represented as nodes in a graph through a feature correlation graph, realizing the natural integration of multi-source heterogeneous data. The adaptive pruning operation identifies true causal relationships based on conditional mutual information, effectively reducing data redundancy and computational complexity. The grouped contrast learning method does not rely on a large amount of labeled data and learns the internal structure of the data in a self-supervised manner, improving the generalization ability of the model in the case of data sparsity. The attention mechanism and dynamic influence coefficients enable the model to dynamically adjust the importance of relationships between nodes according to changes in working conditions, enhancing the adaptability of the model to complex working conditions.

[0151] The automatic tunnel point change observation system combining AI algorithms in the embodiments of the present invention includes:

[0152] A first unit for acquiring visual data of points to be observed in the tunnel, including irregular speckle patterns formed by a speckle projector on the tunnel wall surface;

[0153] A second unit for collecting the irregular speckle patterns from multiple angles using a polarized light camera array to obtain speckle intensity data in different polarization directions, establishing a speckle-stress correlation analysis model based on the speckle intensity data, and obtaining point stress distribution data through polarization transmission analysis and complex domain optimization iteration;

[0154] A third unit for pre-embedding an encoded microstructure generated based on quantum state superposition in the irregular speckle pattern, acquiring deformation data of the encoded microstructure, and calculating three-dimensional displacement data based on the deformation data and the stress distribution data;

[0155] A fourth unit for collecting environmental monitoring data and forming an observation data set with the point stress distribution data and the three-dimensional displacement data;

[0156] A fifth unit for inputting the observation data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts data feature correlations through dynamic topology reconstruction and contrast learning to obtain point change prediction results;

[0157] The sixth unit is configured to generate a warning message if the prediction result of the point position change exceeds a preset change threshold, and send the warning message to the monitoring terminal.

[0158] In a third aspect of the embodiments of the present invention,

[0159] a provided electronic device includes:

[0160] a processor;

[0161] a memory for storing instructions executable by the processor;

[0162] wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0163] In a fourth aspect of the embodiments of the present invention,

[0164] a provided computer-readable storage medium has computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0165] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions for performing various aspects of the present invention loaded thereon.

[0166] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic observation method for tunnel point position changes combined with AI algorithms, characterized in that, Including: Obtain the visual data of the points to be observed in the tunnel, including the irregular speckle pattern formed by the speckle projector on the tunnel wall surface; Use a polarized light camera array to collect the irregular speckle pattern from multiple angles, obtain the speckle intensity data in different polarization directions, establish a speckle-stress correlation analysis model based on the speckle intensity data, and obtain the point stress distribution data through polarization transmission analysis and complex domain optimization iteration; Pre-embed the coding microstructure generated based on quantum state superposition in the irregular speckle pattern, obtain the deformation data of the coding microstructure, and calculate the three-dimensional displacement data according to the deformation data and the stress distribution data; Collect the environmental monitoring data, and form an observation data set with the environmental monitoring data, the point stress distribution data, and the three-dimensional displacement data; Input the observation data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts the data feature correlation through dynamic topology reconstruction and contrast learning to obtain the point change prediction result; If the point change prediction result exceeds the preset change threshold, generate a warning message and send the warning message to the monitoring terminal.

2. The method according to claim 1, characterized in that, Establish a speckle-stress correlation analysis model based on the speckle intensity data, and obtain the point stress distribution data through polarization transmission analysis and complex domain optimization iteration, including: Perform normalization and Gaussian filtering on the speckle intensity data to obtain the filtered intensity data; According to the filtered intensity data, calculate the speckle intensity components in the main polarization direction respectively to obtain the polarization transmission data, and calculate the second-order partial derivative of the polarization transmission data to obtain the local stress direction data; According to the local stress direction data, use the complex domain decomposition algorithm to establish the holographic spectrum analysis of the stress field, obtain the stress principal components, construct an asymptotic optimization path through topological homotopy continuous deformation, and dynamically adjust the iteration step size through the path curvature to obtain the initial stress distribution data under the double constraints of elastic equilibrium and speckle intensity fitting; Perform stress gradient constraint smoothing on the initial stress distribution data to obtain the smoothed stress distribution data; Calculate the weighted combined value of the speckle intensity fitting error and the stress balance deviation. When the weighted combined value is less than the preset first threshold, adjust the elastic equilibrium constraint weight and repeat the calculation of the stress distribution until the weighted combined value meets the requirements of the first threshold.

3. The method according to claim 2, wherein According to the local stress direction data, use the complex domain decomposition algorithm to establish the holographic spectrum analysis of the stress field, obtain the stress principal components, construct an asymptotic optimization path through topological homotopy continuous deformation, and dynamically adjust the iteration step size through the path curvature to obtain the initial stress distribution data under the double constraints of elastic equilibrium and speckle intensity fitting, including: Map the local stress direction data to the complex plane through complex exponential transformation to obtain the complex domain representation of the stress field, and the complex domain representation of the stress field includes an amplitude term and a phase term; Perform two-dimensional Fourier transform on the complex domain representation of the stress field to obtain a frequency domain signal, and use a multi-order band-pass filter to filter the frequency domain signal to obtain multiple groups of filtered signals; Perform singular value decomposition on multiple groups of filtered signals respectively, extract the eigenvectors corresponding to the main eigenvalues, and construct the stress principal components; Establish a topological homotopy mapping from the complex domain representation of the stress field to the principal stress components, calculate the path curvature of the topological homotopy mapping, and adaptively adjust the optimization step size according to the path curvature; At each optimization step size, calculate the weighted combined objective function of the speckle intensity fitting error and the elastic equilibrium constraint. The speckle intensity fitting error characterizes the difference between the measured intensity and the simulated intensity, and the elastic equilibrium constraint includes the spatial partial derivative combination of the stress components; Iteratively update the stress distribution along the topological homotopy mapping path until the difference between the weighted combined objective functions of two adjacent iterations is less than a preset convergence threshold.

4. The method according to claim 1, wherein Pre-embed an encoded microstructure generated based on quantum state superposition in the irregular speckle pattern, obtain the deformation data of the encoded microstructure, and calculate the three-dimensional displacement data according to the deformation data and the stress distribution data, including: Generate a quantum random sequence based on quantum state superposition, map it to an encoded microstructure through a Gaussian function, and embed the encoded microstructure in the irregular speckle pattern to obtain an encoded speckle pattern; Extract the microstructure feature vectors from the encoded speckle pattern, including local gray moments, histogram of gradient directions, and topological invariants; Establish the mapping relationship of the microstructure feature vectors before and after deformation to obtain a feature mapping function, construct and solve a robust matching objective function containing a regularization term based on the feature mapping function to obtain the microstructure deformation data; Conduct intensity analysis on the encoded speckle pattern to obtain speckle intensity distribution data, establish a speckle-stress correlation analysis model according to the speckle intensity distribution data, and calculate the stress distribution data; Construct a displacement field reconstruction equation based on the microstructure deformation data and the stress distribution data. The displacement field reconstruction equation uses a radial basis function as the basic equation, includes a deformation fitting term and a displacement field gradient constraint term, and is constrained by the elastic equilibrium constraint and the displacement boundary condition; Solve the displacement field reconstruction equation to obtain the displacement field weight coefficients, and substitute the displacement field weight coefficients into the radial basis function to obtain the three-dimensional displacement data.

5. The method according to claim 4, wherein Establish the mapping relationship of the microstructure feature vectors before and after deformation to obtain a feature mapping function, construct and solve a robust matching objective function containing a regularization term based on the feature mapping function to obtain the microstructure deformation data, including: Calculate the topological stability index of each feature point in the microstructure feature vector, determine the feature point with the highest topological stability index as the main reference point, and divide the feature space into multiple adaptive deformation regions according to the local deformation gradient value centered on the main reference point; Calculate the local curvature value for the adaptive deformation region, construct a second-order deformation descriptor, and combine the second-order deformation descriptor with the microstructure feature vector to form an enhanced feature vector; Calculate the feature weight coefficients according to the local strain rate of the adaptive deformation region, and assign the feature weight coefficients to each component of the enhanced feature vector to obtain a dynamic feature weight function; Use a hierarchical iterative method to calculate the deformation continuity parameter between the adaptive deformation regions; Substitute the dynamic feature weight function and the deformation continuity parameter into the feature mapping function, solve the robust matching objective function containing a regularization term, and obtain the microstructure deformation data.

6. The method according to claim 1, wherein Input the observed data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts data feature associations through dynamic topology reconstruction and contrast learning. The obtained point position change prediction results include: Construct a feature association graph, map the point stress distribution data, three-dimensional displacement data, and environmental monitoring data in the observed data set to the nodes in the graph, and establish dynamic connection weights based on the mutual information measure between the data; Perform an adaptive pruning operation on the feature association graph, calculate the conditional mutual information between node pairs, and delete the connection relationships with conditional mutual information lower than the adaptive threshold, where the adaptive threshold is dynamically adjusted according to historical prediction errors; Adopt a grouped contrast learning method, randomly sample multiple groups of subgraphs from the feature association graph, apply different intensities of data augmentation operations to each group of subgraphs to generate contrast samples, and train the feature extractor by maximizing the mutual information between the subgraphs; Apply the trained feature extractor to the complete feature association graph, obtain the latent representations of the nodes in the graph, and calculate the dynamic influence coefficients between the nodes based on the attention mechanism; Perform adaptive aggregation on the latent representations of the nodes according to the dynamic influence coefficients, and use a non-linear transformer to map the aggregated features to the prediction space to obtain the point position change prediction results.

7. An automatic tunnel point position change observation system combined with an AI algorithm, which is used to implement the method described in any one of the foregoing claims 1-6, characterized in that, Including: The first unit is used to obtain the visual data of the points to be observed in the tunnel, including the irregular speckle patterns formed by the speckle projector on the tunnel wall; The second unit is used to collect the irregular speckle patterns from multiple angles by using a polarized light camera array, obtain the speckle intensity data in different polarization directions, establish a speckle-stress correlation analysis model based on the speckle intensity data, and obtain the point stress distribution data through polarization transmission analysis and complex domain optimization iteration; The third unit is used to pre-embed the encoded microstructures generated based on quantum state superposition in the irregular speckle patterns, obtain the deformation data of the encoded microstructures, and calculate the three-dimensional displacement data according to the deformation data and the stress distribution data; The fourth unit is used to collect environmental monitoring data, and form an observed data set with the point stress distribution data and the three-dimensional displacement data; The fifth unit is used to input the observed data set into a preset self-evolving heterogeneous feature network, and the self-evolving heterogeneous feature network extracts data feature associations through dynamic topology reconstruction and contrast learning to obtain the point position change prediction results; The sixth unit is used to generate a warning message if the point position change prediction result exceeds a preset change threshold, and send the warning message to the monitoring terminal.

8. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images