Tunnel engineering structure deformation monitoring method based on deep learning

Abstract

The invention discloses a tunnel engineering structure deformation monitoring method based on deep learning. The method comprises the following steps: S1, constructing a data set; s2, data preprocessing; s3, extracting spatio-temporal features by adopting a graph attention spatio-temporal converter; s4, constructing a prediction model by adopting a Poincare sphere embedding and control differential equation; s5, adopting a hyperbolic element learning and variational Bayesian optimization model; s6, monitoring in real time and adopting a self-adaptive early warning mechanism; and S7, establishing an intelligent feedback mechanism, and optimizing monitoring parameters. According to the method, accurate prediction and efficient monitoring of tunnel engineering structure deformation are realized, the prediction precision, the calculation efficiency and the early warning capability are remarkably improved, the generalization capability and the anomaly detection performance of the model are enhanced, the method effectively adapts to deformation characteristics of different tunnel structures, and the method is suitable for large-scale popularization and application. The method can be widely applied to the field of tunnel engineering structure deformation monitoring in a complex geological environment.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and particularly to a method for monitoring the deformation of tunnel engineering structures based on deep learning. Background Art

[0002] The long-term stability of tunnel engineering structures has a crucial impact on transportation, underground facilities, and infrastructure construction. With the continuous expansion of global infrastructure, more and more tunnels are being built in complex geological environments, such as soft soil layers, high groundwater level areas, and active fault zones. These complex environments can cause the deformation of tunnel structures during long-term operation, and even pose safety hazards such as cracks, settlements, and distortions. Therefore, the deformation monitoring of tunnel structures has important engineering value and significance for safety assurance. Currently, deformation monitoring mainly relies on high-precision sensors, geological exploration techniques, and computer numerical analysis methods to evaluate the health status of tunnel structures through data collection, processing, and prediction. However, existing monitoring methods still have many technical bottlenecks in complex environments and are difficult to meet the requirements of accurate prediction and real-time control.

[0003] Traditional tunnel deformation monitoring techniques mainly include optical measurement methods, radar interferometry, terrestrial laser scanning, and numerical simulation methods based on finite element analysis (FEM). Optical measurement methods usually use total stations or laser rangefinders to periodically scan the tunnel surface to obtain displacement information. Although this method has high measurement accuracy, it has disadvantages such as long measurement periods, poor environmental adaptability, and inability to monitor in real time. Radar interferometry (InSAR) can use satellite remote sensing technology to monitor large-scale deformation data, but its accuracy is greatly affected by factors such as weather, vegetation, and signal interference, and it is difficult to apply in internal tunnel monitoring. Terrestrial laser scanning technology can generate high-precision three-dimensional point cloud models of tunnels, but the data processing is complex, the calculation cost is high, and it is difficult to monitor hidden deformations inside the tunnel in real time. Numerical simulation methods based on finite element analysis can predict tunnel deformation by combining geological data, but their results depend on initial modeling parameters, are difficult to adapt to dynamic changes in the actual environment, and have high computational complexity and cannot update prediction results in real time.

[0004] In recent years, with the development of the Internet of Things (IoT) and deep learning technologies, tunnel deformation monitoring has gradually developed towards intelligence and automation. The introduction of IoT technology enables tunnel monitoring sensor networks to collect large-scale deformation data and provide richer spatio-temporal information. Deep learning technology can build data-driven prediction models based on large-scale monitoring data to improve the accuracy and timeliness of deformation prediction. However, existing deep learning deformation monitoring methods still have multiple technical difficulties.

[0005] First, existing deep learning methods have difficulty effectively integrating spatial correlations and temporal dynamic characteristics during modeling. Many studies use convolutional neural networks for spatial feature extraction of deformation data. However, convolutional neural networks are mainly used for regular grid data and have difficulty handling the topological relationships of complex sensor networks. Although recurrent neural networks and long short-term memory networks can model time series, they have weak long-term dependence modeling capabilities and are difficult to effectively fuse with spatial information, resulting in information loss when the model extracts spatio-temporal features. In addition, tunnel sensor data is usually irregular and dynamically changing topological data, which is difficult for traditional time series models to directly process.

[0006] Second, existing deformation prediction models have insufficient modeling capabilities in non-Euclidean spaces. The deformation data of tunnel structures often evolves in high-dimensional non-linear spaces, while existing prediction methods are mostly based on Euclidean spaces for calculation and cannot accurately describe complex deformation patterns. Recently, non-Euclidean geometry methods such as Poincaré ball embedding have made some progress in deep learning, but their application in deformation monitoring is still in the exploratory stage. Existing studies usually only use non-Euclidean embedding for static feature extraction and lack a complete temporal modeling framework.

[0007] In addition, existing deformation prediction models are difficult to adapt to the deformation characteristics of different tunnel structures during the optimization process. The geological environments, construction methods, and support structures of different tunnels vary, resulting in strong scenario dependence of deformation patterns. Existing deep learning models often adopt a unified training strategy, train on different tunnel data sets and then conduct tests, resulting in weak generalization ability. The model may have large prediction deviations in new tunnel environments. How to use meta-learning to enable the model to quickly adapt to new deformation patterns with a small amount of data and a limited number of trainings is an important challenge in current research.

[0008] Furthermore, existing deformation prediction methods lack modeling of uncertainty. In practical applications, tunnel monitoring data may be affected by factors such as noise, sensor failures, and external environmental disturbances, resulting in large uncertainties in the prediction model. However, most traditional deep learning methods use point estimation for prediction and cannot quantify the uncertainty of the model. This may lead to incorrect decisions by the monitoring system in the face of abnormal situations and affect engineering safety. Therefore, combining probability methods such as variational Bayesian inference to model the uncertainty of the prediction model is an important direction to improve the reliability of deformation prediction.

[0009] Finally, existing methods still have deficiencies in the intelligent feedback mechanism of the deformation monitoring system. In traditional monitoring systems, the prediction model and the sensor data processing system are often independent, lacking an automated feedback adjustment mechanism. Even if some research has introduced a feedback mechanism, most of them use fixed rules to adjust monitoring parameters and cannot dynamically optimize key parameters such as sensor layout and sampling frequency according to the actual situation. How to combine evolutionary optimization algorithms to continuously optimize the model and monitoring parameters during the feedback process to enable the system to have an adaptive adjustment ability is the key issue in improving the intelligence level of monitoring.

[0010] In summary, although deep learning and Internet of Things technologies have made remarkable progress in the field of tunnel deformation monitoring, there are still obvious deficiencies in aspects such as spatio-temporal feature extraction, non-Euclidean space modeling, model generalization ability, uncertainty modeling, and intelligent feedback mechanism. Therefore, proposing a deformation feature extraction method based on graph attention spatio-temporal transformers, a time-series deformation prediction model combining Poincaré ball embedding and controlled differential equations, and optimizing it using hyperbolic meta-learning and variational Bayesian optimization, while constructing an intelligent feedback mechanism, has important theoretical significance and engineering application value for improving the accuracy, adaptability, and stability of tunnel deformation monitoring systems.

[0011] Therefore, how to provide a deep learning-based tunnel engineering structure deformation monitoring method is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0012] An object of the present invention is to propose a deep learning-based tunnel engineering structure deformation monitoring method. The present invention uses a graph attention spatio-temporal transformer to extract deformation features, uses Poincaré ball embedding combined with controlled differential equations to model the time-series evolution, and adjusts the latent variable distribution through Sinkhorn-Knopp Wasserstein optimization. Finally, the variational Bayesian optimization strategy based on hyperbolic meta-learning is used to improve the prediction accuracy and generalization ability. The present invention details deformation feature extraction, non-Euclidean space modeling, time-series deformation prediction, optimization strategy, and intelligent feedback mechanism, and has the advantages of high prediction accuracy, strong adaptability, strong uncertainty modeling ability, high computational stability, and intelligent feedback optimization ability.

[0013] A deep learning-based tunnel engineering structure deformation monitoring method according to an embodiment of the present invention includes the following steps:

[0014] S1. Connect tunnel monitoring sensors using Internet of Things technology, collect deformation data, and construct a data set;

[0015] S2. Preprocess the data set, including data synchronization, noise filtering, outlier removal, time-series alignment, and data normalization;

[0016] S3. Extract deformation features using the graph attention spatio-temporal transformer, construct a dynamic adjacency matrix to learn spatial correlations, use the spatio-temporal transformer to extract temporal patterns, and adaptively fuse spatial and temporal features to generate the final spatio-temporal features;

[0017] S4. Represent deformation features using Poincaré ball embedding, model temporal evolution based on the control differential equation, combine Sinkhorn-Knopp Wasserstein optimization to adjust the latent variable distribution, introduce non-conservative Hamiltonian dynamics and fractional dissipation terms to optimize the prediction model, and use the adaptive step Runge-Kutta method to improve the calculation accuracy to generate the prediction output;

[0018] S5. Adopt the variational Bayesian optimization strategy of hyperbolic meta-learning, construct an optimization framework based on meta-learning, calculate the hyper-curvature gradient for hyperbolic optimization, then adjust the meta-parameters using information geometry, combine the variational Bayesian method for probabilistic inference optimization, and finally adjust the deformation prediction model through contrastive learning to calculate the final optimization result;

[0019] S6. Based on the prediction model, conduct real-time monitoring of the tunnel structure and adopt an adaptive early warning mechanism for hierarchical early warning;

[0020] S7. Establish an intelligent feedback mechanism, collect monitoring data to optimize the prediction model, and dynamically adjust monitoring parameters in combination with the evolutionary optimization strategy.

[0021] Optionally, the specific steps of S3 include:

[0022] S31. Construct a dynamic adjacency matrix of the monitoring sensors, use the graph attention mechanism to calculate the spatial correlations between monitoring points, extract weighted spatial features, and perform multi-layer updates to enhance the topological relationship modeling ability;

[0023] S32. Segment the monitoring time series, use the spatio-temporal transformer to calculate the feature correlations at different time steps, extract short-term mutation patterns and long-term deformation trends, and maintain the integrity of time information through residual connections;

[0024] S33. Perform adaptive weighted fusion on the spatial and temporal features, normalize the spatio-temporal features at different scales, perform noise suppression, and output the fused spatio-temporal features as the input for deformation prediction.

[0025] Optionally, the specific steps of S4 include:

[0026] S41. Construct a Poincaré ball embedding space to obtain the Poincaré ball embedding feature representation;

[0027] S42. Based on the control variational autoencoder, perform variational encoding on the Poincaré ball embedding features to model the temporal latent variables;

[0028] S43. Introduce non-conservative Hamiltonian dynamics, constrain the temporal latent variables, and establish the control differential equation;

[0029] S44. Based on the modeling results of the control differential equation, obtain the predicted output Y out 。

[0030] Optionally, the S41 specifically includes:

[0031] S411. Project the spatio-temporal features into the Poincaré sphere embedding space Perform non-Euclidean transformation to obtain the deformed feature representation Z0 embedded in the hyperbolic space;

[0032] S412. In the embedding space, calculate the hyperbolic geodesic distance of the deformed features at different times to measure the non-Euclidean relationship between the deformed features:

[0033]

[0034] where, ‖·‖ represents the modulus of the vector, represents the geodesic distance between the deformed features Z1 and Z2 in the Poincaré sphere space, Z1 and Z2 are the deformed features at two different time steps respectively, and arccosh(·) is the inverse hyperbolic cosine function;

[0035] S413. Calculate the manifold embedding gradient based on the geodesic distance and optimize the embedding features:

[0036]

[0037] where, is the gradient of the geodesic distance with respect to the deformed feature;

[0038] S414. The finally optimized Poincaré sphere embedding feature representation is:

[0039]

[0040] where, η is the learning rate.

[0041] Optionally, the S42 specifically includes:

[0042] S421. Construct a variational autoencoder to perform variational encoding on the optimized Poincaré sphere embedding feature to obtain the temporal latent variable Z t ;

[0043] S422. Use the control differential equation to describe the time evolution of the latent variable, so that the temporal features change in the continuous time domain:

[0044]

[0045] where, Zt The sequential latent variable representing time t, f θ (·) is the control dynamic equation, C t is the control path signal, defined as follows:

[0046]

[0047] where, g φ (·) is the sequential control mapping function, H τ is the spatio-temporal feature at time τ;

[0048] S423. Based on ensuring the consistency between the latent variable and the real data during the time evolution, introduce the Wasserstein optimal transport optimization. In the latent variable Z t space, calculate the Wasserstein-2 optimal transport distance based on the Sinkhorn-Knopp algorithm:

[0049]

[0050] where, is the optimal transport loss, γ i,j is the transport weight matrix, representing the optimal transport weight between the deformation features Z i and Z j ; is the geodesic distance between Z i and Z j in the Poincaré ball space;

[0051] The iterative update formula of the Sinkhorn-Knopp transport weight matrix is as follows:

[0052]

[0053] where, is the transport matrix at the k-th iteration, a i and b j are the balance parameters, controlling the stability of the Sinkhorn algorithm during the iteration.

[0054] Optionally, the S43 specifically includes:

[0055] S431. In the Poincaré ball space construct a control differential equation based on non-conservative Hamiltonian dynamics, and make the time evolution of the deformation feature conform to the physical law:

[0056]

[0057] where, is the rate of change of the latent variable with time, To calculate the gradient of the Hamiltonian H(Z t , C t ) with respect to Z t , λ is the dissipation coefficient, ‖C t ‖ 2 is the energy of the control signal, 〈Z t , C t 〉 is the interaction term between the deformation state and the control signal, and D(Z t ) is the dissipation term, which is modeled using fractional derivatives:

[0058]

[0059] where Γ(·) is the gamma function and α is the fractional order, which is used to control the dissipation degree of the deformation characteristics over time;

[0060] S432. Use the non - adaptive step - size Runge - Kutta method to numerically solve the established control differential equation. Calculate the change rate of the deformation state at the current time step, and use the fourth - order Runge - Kutta formula for iterative update. Calculate multiple intermediate predicted values and then take the weighted average to obtain the final result. Subsequently, adjust the step size according to the local truncation error, and repeat the calculation until the entire time interval is covered to obtain the time - series evolution trajectory of the deformation state.

[0061] Optionally, the specific content of S5 includes:

[0062] S51. Based on the meta - learning optimization strategy, construct a high - order gradient optimization model to optimize the predicted output Y out . Define the model parameters as θ, and the high - order gradient update formula is:

[0063]

[0064] where is the gradient of the parameter, α is the first - order gradient learning rate, β is the high - order gradient adjustment coefficient, the loss function the first - order gradient of the loss function with respect to the parameter θ, the second - derivative of the loss function with respect to θ;

[0065] S52. Construct a hyperbolic optimization space, introduce the Finsler metric in the hyperbolic space, and perform hyper - curvature optimization on θ′:

[0066]

[0067] Calculate the hyperbolic Finsler gradient:

[0068]

[0069] where G θ′is the hyperbolic Finsler metric tensor, is the inverse matrix of the Finsler metric tensor, λ is the super-curvature scaling factor, is the Finsler metric norm;

[0070] S53. Optimize θ′ using information geometric metric and calculate the Fisher-Rao metric:

[0071]

[0072] where D KL (·‖·) is the KL divergence, and adjust the output distribution p(Y out |θ′) of the current deformation prediction model to approximate the output distribution p(Y out |θ ″ ) of the optimized deformation prediction model;

[0073] Perform information geometric optimization based on the Fisher-Rao metric:

[0074]

[0075] where is the learning rate of information geometric optimization;

[0076] S54. Perform variational Bayesian optimization on the optimized θ′ and define the optimization objective:

[0077]

[0078] Calculate the finally optimized parameters:

[0079]

[0080] where represents the expectation of the loss function under the variational distribution q(θ′), p(θ′) is the prior distribution of the model parameters, q(θ′) is the variational posterior distribution, and ρ is the learning rate of Bayesian optimization;

[0081] S55. Optimize the final deformation prediction result using contrastive learning and finally calculate the optimized deformation prediction output:

[0082]

[0083] where ε is the learning rate of variational Bayesian optimization, μ is the weight parameter of contrastive learning optimization, is the contrastive learning loss function.

[0084] The beneficial effects of the present invention are:

[0085] (1) The present invention uses a graph attention spatio-temporal transformer to deeply model sensor data. Compared with traditional CNN and RNN structures, this method can simultaneously learn the dynamic topological relationships between monitoring points and extract short-term mutation patterns and long-term trend changes in the deformation sequence through the spatio-temporal transformer, ensuring the integrity of spatio-temporal information.

[0086] (2) Traditional deformation prediction methods rely on Euclidean space modeling and are difficult to describe complex high-dimensional deformation features. The present invention adopts the Poincaré ball embedding method to project deformation features into a non-Euclidean space, enabling them to be modeled in a hyperbolic space that more conforms to actual physical characteristics, improving the discrimination ability and stability of deformation features.

[0087] (3) Based on the Poincaré ball embedding, the present invention uses a controlled variational autoencoder for time-series latent variable modeling and combines non-conservative Hamiltonian dynamics and fractional derivatives to model dissipation terms, ensuring that the prediction model can more accurately depict the deformation evolution process, while avoiding the problem of error accumulation in long-term prediction and improving the long-term stability of the model.

[0088] (4) During the time-series modeling process, the present invention introduces Sinkhorn-Knopp Wasserstein optimization, using the optimal transport method to adjust the distribution of deformation latent variables, making the latent variables more conform to the distribution characteristics of real tunnel deformation data. Compared with traditional variational inference, this method can more effectively align the data distribution, improve the generalization ability of the model, and make it applicable to different tunnel structure environments.

[0089] (5) Existing deep learning models need to be retrained with a large amount of data in a new environment. The present invention adopts a variational Bayesian optimization strategy of hyperbolic meta-learning, combining methods such as Finsler metric optimization, information geometry optimization, and variational Bayesian inference, enabling the model to quickly adapt to different tunnel deformation patterns, reducing the dependence on a large amount of labeled data, and improving the applicability in a new environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0091] Figure 1 is the overall flowchart of a tunnel engineering structure deformation monitoring method based on deep learning proposed by the present invention;

[0092] Figure 2 is the flowchart of optimizing the deformation prediction model by the variational Bayesian optimization strategy of hyperbolic meta-learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0093] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0094] Reference Figure 1 and Figure 2 , a deformation monitoring method for tunnel engineering structures based on deep learning, comprising the following steps:

[0095] S1. Connect tunnel monitoring sensors using Internet of Things technology, collect deformation data, and construct a data set;

[0096] S2. Preprocess the data set, including data synchronization, noise filtering, outlier removal, time series alignment, and data normalization;

[0097] S3. Use a graph attention spatio-temporal transformer to extract deformation features, construct a dynamic adjacency matrix to learn spatial correlations, use the spatio-temporal transformer to extract time series patterns, and adaptively fuse spatial and time features to generate final spatio-temporal features;

[0098] S4. Use Poincaré ball embedding to represent deformation features, model time series evolution based on control differential equations, combine Sinkhorn-Knopp Wasserstein optimization to adjust the latent variable distribution, introduce non-conservative Hamiltonian dynamics and fractional dissipation terms to optimize the prediction model, and use the adaptive step Runge-Kutta method to improve calculation accuracy to generate a prediction output;

[0099] S5. Adopt a variational Bayesian optimization strategy based on hyperbolic element learning, construct an optimization framework based on meta-learning, calculate the hyper-curvature gradient for hyperbolic optimization, then use information geometry to adjust meta-parameters, combine the variational Bayesian method for probabilistic inference optimization, and finally adjust the deformation prediction model through contrastive learning to calculate the final optimization result;

[0100] S6. Based on the prediction model, conduct real-time monitoring of the tunnel structure, and adopt an adaptive early warning mechanism for hierarchical early warning;

[0101] S7. Establish an intelligent feedback mechanism, collect monitoring data to optimize the prediction model, and combine an evolutionary optimization strategy to dynamically adjust monitoring parameters.

[0102] In this embodiment, the specific content of S3 includes:

[0103] S31. Construct a dynamic adjacency matrix of the monitoring sensors, use the graph attention mechanism to calculate the spatial correlations between monitoring points, extract weighted spatial features, and perform multi-layer updates to enhance the topological relationship modeling ability;

[0104] S32. Segment the monitoring time series, calculate the feature correlation at different time steps using a spatio-temporal transformer, extract short-term mutation patterns and long-term deformation trends, and maintain the integrity of time information through residual connections;

[0105] S33. Perform adaptive weighted fusion on spatial features and time features, normalize spatio-temporal features at different scales, suppress noise, and output the fused spatio-temporal features as the input for deformation prediction.

[0106] In this embodiment, the specific steps of S4 include:

[0107] S41. Construct a Poincaré ball embedding space to obtain a Poincaré ball embedding feature representation;

[0108] S42. Based on a controlled variational autoencoder, perform variational encoding on the Poincaré ball embedding features to model temporal latent variables;

[0109] S43. Introduce non-conservative Hamiltonian dynamics to constrain the temporal latent variables and establish a control differential equation;

[0110] S44. Based on the modeling result of the control differential equation, obtain the prediction output Y out .

[0111] In this embodiment, the specific steps of S41 include:

[0112] S411. Project the spatio-temporal features onto the Poincaré ball embedding space Perform non-Euclidean transformation to obtain a deformation feature representation Z0 embedded in the hyperbolic space;

[0113] S412. In the embedding space, calculate the hyperbolic geodesic distance between deformation features at different times to measure the non-Euclidean relationship between deformation features:

[0114]

[0115] where, ‖·‖ represents the modulus of a vector, represents the geodesic distance between deformation features Z1 and Z2 in the Poincaré ball space, Z1 and Z2 are deformation features at two different time steps respectively, and arccosh(·) is the inverse hyperbolic cosine function;

[0116] S413. Calculate the manifold embedding gradient based on the geodesic distance to optimize the embedding features:

[0117]

[0118] where, is the gradient of the geodesic distance with respect to the deformation feature;

[0119] S414. The finally optimized Poincaré sphere embedding feature representation is as follows:

[0120]

[0121] where η is the learning rate.

[0122] In this embodiment, the S42 specifically includes:

[0123] S421. Construct a variational encoder to perform variational encoding on the optimized Poincaré sphere embedding feature to obtain the temporal latent variable Z t ;

[0124] S422. Use a controlled differential equation to describe the time evolution of the latent variable, so that the temporal features change in the continuous time domain:

[0125]

[0126] where Z t represents the temporal latent variable at time t, and f θ (·) is the control dynamic equation, and C t is the control path signal, which is defined as follows:

[0127]

[0128] where g φ (·) is the temporal control mapping function, and H τ is the spatio-temporal feature at time τ;

[0129] S423. Based on ensuring the consistency between the latent variable and the real data during the time evolution, introduce Wasserstein optimal transport optimization, and calculate the Wasserstein-2 optimal transport distance in the latent variable Z t space based on the Sinkhorn-Knopp algorithm:

[0130]

[0131] where is the optimal transport loss, γ i,j is the transport weight matrix, representing the optimal transport weight between the deformation features Z i and Z j , is the geodesic distance between Z i and Z j in the Poincaré sphere space;

[0132] The iterative update formula of the Sinkhorn-Knopp transport weight matrix is as follows:

[0133]

[0134] Among them, is the transmission matrix of the k-th iteration, a i and b j are balance parameters that control the stability of the Sinkhorn algorithm during the iteration process.

[0135] In this embodiment, the S43 specifically includes:

[0136] S431. In the Poincaré sphere space , construct a control differential equation based on non-conservative Hamiltonian dynamics, and make the time evolution of the deformation characteristics conform to physical laws:

[0137]

[0138] Among them, is the rate of change of the latent variable with time, is the gradient of the calculation of the Hamiltonian H(Z t , C t ) with respect to Z t , λ is the dissipation coefficient, ‖C t ‖ 2 is the energy of the control signal, <Z t , C t > is the interaction term between the deformation state and the control signal, D(Z t ) is the dissipation term, which is modeled using fractional derivatives:

[0139]

[0140] Among them, Γ(·) is the gamma function, and α is the fractional order, which is used to control the dissipation degree of the deformation characteristics with time;

[0141] S432. Use the non-adaptive step Runge-Kutta method to numerically solve the established control differential equation, calculate the rate of change of the deformation state at the current time step, perform iterative updates using the fourth-order Runge-Kutta formula, calculate multiple intermediate predicted values and then take the weighted average to obtain the final result, and then adjust the step size according to the local truncation error, and repeat the calculation until the entire time interval is covered to obtain the time series evolution trajectory of the deformation state.

[0142] In this embodiment, the S5 specifically includes:

[0143] S51. Based on the meta-learning optimization strategy, construct a high-order gradient optimization model to optimize the predicted output Y out , define the model parameter as θ, and the high-order gradient update formula is:

[0144]

[0145] Among them, To calculate the gradient of the parameter, α is the first-order gradient learning rate, and β is the high-order gradient adjustment coefficient. Loss function The first-order gradient of the loss function with respect to the parameter θ The second derivative of the loss function with respect to θ;

[0146] S52. Construct a hyperbolic optimization space, introduce the Finsler metric in the hyperbolic space, and perform hyper-curvature optimization on θ′:

[0147]

[0148] Calculate the hyperbolic Finsler gradient:

[0149]

[0150] Among them, G θ′ Is the hyperbolic Finsler metric tensor, Is the inverse matrix of the Finsler metric tensor, λ is the hyper-curvature scaling factor, Is the Finsler metric norm;

[0151] S53. Optimize θ′ using the information geometric metric and calculate the Fisher-Rao metric:

[0152]

[0153] Among them, D KL (·‖·) is the KL divergence, which adjusts the output distribution p(Y out |θ′) of the current deformation prediction model to approximate the output distribution p(Y out |θ ″ ) of the optimized deformation prediction model;

[0154] Perform information geometric optimization based on the Fisher-Rao metric:

[0155]

[0156] Among them, Is the information geometric optimization learning rate;

[0157] S54. Perform variational Bayesian optimization on the optimized θ′ and define the optimization objective:

[0158]

[0159] Calculate the finally optimized parameter:

[0160]

[0161] in, Represents the loss function under the variational distribution q(θ′) , p(θ′) is the prior distribution of model parameters, q(θ′) is the variational posterior distribution, and ρ is the Bayesian optimization learning rate;

[0162] S55. Use contrastive learning to optimize the final deformation prediction result, and finally calculate the optimized deformation prediction output:

[0163]

[0164] Among them, ε is the learning rate of variational Bayesian optimization, μ is the weight parameter of contrastive learning optimization, is the contrastive learning loss function.

[0165] Embodiment 1:

[0166] In order to verify the feasibility and superiority of the present invention in deformation monitoring of tunnel engineering, the present invention is applied to the deformation monitoring of a highway tunnel in a mountainous area in Sichuan. The tunnel is 5.3 kilometers long and passes through multiple areas with complex geological structures. The main strata include soft soil layers, fault fracture zones and swelling rock layers, and are affected by high groundwater infiltration and periodic geological movements. Since the tunnel was built, problems such as lining cracking, local settlement, and lining deformation exceeding the limit have occurred many times, affecting operational safety. Due to the complex geological environment of the tunnel, the existing monitoring methods (such as total station measurement and lidar scanning) have unstable measurement accuracy in high humidity and high vibration environments, and lack real-time prediction capabilities, and cannot effectively warn of sudden deformation disasters. Therefore, the tunnel was selected as the test scene of the present invention to verify its effectiveness in a complex engineering environment.

[0167] The present invention deploys 120 sets of high-precision displacement sensors, 60 sets of accelerometers, 50 sets of ground stress monitors and 30 sets of temperature and humidity sensors in the tunnel, and collects data through the Internet of Things system. The sensor network collects tunnel structure data at intervals of 5 seconds to form a monitoring data set with high time resolution.

[0168] In the data processing stage, the present invention first performs data preprocessing on the collected raw data, including time synchronization, noise filtering, outlier removal, time series alignment and normalization. After processing, the graph attention spatiotemporal transformer is used to extract deformation features, construct a dynamic adjacency matrix, learn the spatial correlation between monitoring points, and extract short-term mutation patterns and long-term deformation trends through the spatiotemporal transformer. Compared with traditional RNN and CNN prediction models, this method can more accurately extract deformation features of different time scales and avoid feature information loss.

[0169] In the prediction modeling stage, the present invention adopts Poincaré sphere embedding to project spatio-temporal features into the hyperbolic space, so as to improve the distinguishability of complex deformation patterns, and models the temporal latent variables based on the controlled variational autoencoder. In addition, the present invention introduces non-conservative Hamiltonian dynamics and fractional derivative dissipation terms to optimize the physical consistency of the deformation prediction model, and adopts an adaptive step-size method for numerical solution to improve the prediction accuracy and computational stability.

[0170] The present invention has been continuously operating in this tunnel for 6 months, accumulating 120 million pieces of deformation data, and comparing with traditional monitoring methods. The experimental results show that the present invention is superior to traditional methods in terms of deformation prediction accuracy, early warning ability, computational efficiency, etc.

[0171] To further verify the prediction accuracy of the present invention, we select the deformation data of the past 3 months, compare the prediction results of the present invention with the actual measured values, and conduct a comparative analysis with the traditional LSTM model and the finite element numerical simulation method. The comparison data are shown in Table 1 below:

[0172] Table 1: Performance comparison between the present invention and traditional methods in tunnel deformation monitoring

[0173] Evaluation index The present invention LSTM prediction model Finite element numerical simulation Average prediction error (mm) 0.82 2.65 3.18 Maximum prediction error (mm) 1.45 4.82 5.67 Deformation trend matching degree (%) 94.6％ 79.3％ 82.1％ Prediction lead time (hours) 48 12 24 Computation time (seconds / time) 0.68 3.52 10.74

[0174] It can be seen from the data in Table 1 that the prediction error of the present invention is significantly lower than that of traditional methods. In terms of the average prediction error, the error of the present invention is only 0.82 mm, which is reduced by 69.0% compared with 2.65 mm of the LSTM prediction model and 74.2% compared with 3.18 mm of the finite element numerical simulation method, indicating that the present invention can more accurately predict the tunnel structure deformation and reduce the prediction error. In terms of the maximum prediction error, the error of the present invention is 1.45 mm, which is much lower than 4.82 mm of the LSTM method and 5.67 mm of the finite element method, indicating that the present invention can better control the prediction deviation and avoid the problem of error amplification in extreme cases.

[0175] In terms of the deformation trend matching degree, the present invention reaches 94.6%, which is at least 12.5% higher than 79.3% of the LSTM and 82.1% of the finite element method. This shows that the present invention can more accurately learn the long-term trend of tunnel deformation and reduce the influence of the model on noise data, making the prediction curve closer to the actual monitoring value. In addition, the prediction lead time of the present invention reaches 48 hours, which is 4 times higher than 12 hours of the LSTM and twice as high as 24 hours of the finite element numerical simulation. This result indicates that the present invention can not only more accurately predict the deformation amount, but also provide an early warning earlier, helping tunnel managers take protective measures earlier, thereby reducing potential safety hazards.

[0176] In terms of computational efficiency, the computational time of the present invention is only 0.68 seconds per time, which is 5.2 times faster than that of LSTM (3.52 seconds per time) and nearly 16 times faster than that of finite element simulation (10.74 seconds per time). This shows that the present invention has optimized the computational complexity, enabling the prediction model to complete the calculation in a shorter time and meet the real-time requirements of tunnel deformation monitoring. In contrast, although the finite element numerical simulation method has strong physical consistency, its computational cost is high and it is difficult to be used for real-time prediction. Due to the need to rely on loop calculations, LSTM has a slow computational speed and is insufficient in long-term dependence modeling.

[0177] Generally speaking, the present invention is superior to traditional methods in terms of tunnel deformation prediction accuracy, early warning ability, computational efficiency, etc. It can not only provide more accurate deformation prediction results, but also meet the real-time requirements of tunnel monitoring, and provide sufficient early warning time before abnormal deformation occurs, providing more reliable technical support for tunnel structure safety management.

[0178] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered by the protection scope of the present invention.

Claims

1. A tunnel engineering structure deformation monitoring method based on deep learning, characterized in that: The steps include: S1. Use IoT technology to connect tunnel monitoring sensors, collect deformation data, and build a data set; S2, preprocessing the data set, including data synchronization, noise filtering, outlier removal, time series alignment and data normalization; S3, using graph attention spatiotemporal transformer to extract deformation features, constructing dynamic adjacency matrix to learn spatial correlation, using spatiotemporal transformer to extract temporal patterns, and adaptively fusing spatial and temporal features to generate final spatiotemporal features; S4, using Poincare sphere embedding to characterize deformation characteristics, modeling time series evolution based on control differential equations, combining Sinkhorn-Knopp Wasserstein optimization to adjust the distribution of latent variables, introducing non-conservative Hamiltonian dynamics and fractional-order dissipative term optimization prediction model, using adaptive step size Runge-Kutta method to improve calculation accuracy, and generate prediction output; S5. Adopt the variational Bayesian optimization strategy of hyperbolic meta-learning, build an optimization framework based on meta-learning, calculate the hypercurvature gradient for hyperbolic optimization, then use information geometry to adjust the meta-parameters, combine the variational Bayesian method for probabilistic inference optimization, and finally adjust the deformation prediction model through contrastive learning to calculate the final optimization result; S6. Real-time monitoring of tunnel structure based on prediction model and hierarchical warning using adaptive early warning mechanism; S7. Establish an intelligent feedback mechanism, collect monitoring data to optimize the prediction model, and dynamically adjust the monitoring parameters in combination with evolutionary optimization strategies.

2. A tunnel engineering structure deformation monitoring method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31. Construct a dynamic adjacency matrix of monitoring sensors, use the graph attention mechanism to calculate the spatial correlation between monitoring points, extract weighted spatial features, and perform multi-layer updates to enhance the topological relationship modeling capability; S32, segmenting the monitoring time series, using a spatiotemporal transformer to calculate feature correlations at different time steps, extracting short-term mutation patterns and long-term deformation trends, and maintaining the integrity of time information through residual connections; S33, adaptively weighted fusion of spatial features and temporal features, normalization of spatiotemporal features of different scales, noise suppression, and output of fused spatiotemporal features as input for deformation prediction.

3. The method for monitoring deformation of tunnel engineering structures based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41, construct the Poincare sphere embedding space and obtain the Poincare sphere embedding feature representation; S42, based on the control variational autoencoder, the Poincare sphere embedding features are variationally encoded to model the time series latent variables; S43, introduce non-conservative Hamiltonian dynamics, constrain the time series potential variables, and establish the control differential equation; S44. Based on the modeling results of the control differential equation, the predicted output Y is obtained. out .

4. The method for monitoring deformation of tunnel engineering structures based on deep learning according to claim 3 is characterized in that: The S41 specifically includes: S411, projecting the spatiotemporal features into the Poincare sphere embedding space P n , perform non-Euclidean transformation to obtain the deformation feature representation Z0 embedded in the hyperbolic space; S412. In the embedding space, the hyperbolic geodesic distances of the deformation features at different times are calculated to measure the non-Euclidean relationship between the deformation features: Among them, ‖·‖ represents the modulus of the vector, represents the geodesic distance between the deformation features Z1 and Z2 in the Poincare sphere space, Z1 and Z2 are deformation features at two different time steps, respectively, and arccosh(·) is the inverse hyperbolic cosine function; S413. Calculate the manifold embedding gradient based on the geodesic distance and optimize the embedding features: in, is the gradient of the geodesic distance to the deformation feature; S414, the final optimized Poincare sphere embedding feature is expressed as: Among them, η is the learning rate.

5. The method for monitoring deformation of tunnel engineering structures based on deep learning according to claim 3, characterized in that: The S42 specifically includes: S421. Construct a variational encoder to embed the optimized Poincare sphere into features Perform variational coding to obtain the time series latent variable Z t ; S422. Use the control differential equation to describe the time evolution of the potential variables, so that the time series characteristics change in the continuous time domain: Among them, Z t represents the time series latent variable at time t, f θ (·) is the control dynamic equation, C t For the control path signal, it is defined as follows: Among them, g φ (·) is the timing control mapping function, H τ is the spatiotemporal characteristic of time τ; S423, based on ensuring that the latent variables are consistent with the real data during the time evolution, the Wasserstein optimal transmission optimization is introduced. t In space, the optimal transmission distance of Wasserstein-2 is calculated based on the Sinkhorn-Knopp algorithm: in, is the optimal transmission loss, γ i,j is the transmission weight matrix, representing the deformation feature Z i and Z j The optimal transmission weight between is Z in the Poincare sphere space i and Z j The geodesic distance between The iterative update formula of the Sinkhorn-Knopp transmission weight matrix is ​​as follows: in, is the transmission matrix of the kth iteration, a i and b j is a balance parameter that controls the stability of the Sinkhorn algorithm during the iteration process.

6. The method for monitoring deformation of tunnel engineering structures based on deep learning according to claim 3, characterized in that: The S43 specifically includes: S431, in the Poincare sphere space Based on the non-conservative Hamiltonian dynamics, the control differential equation is constructed to make the time evolution of the deformation characteristics conform to the physical laws: in, is the rate of change of the latent variable over time, To calculate the Hamiltonian H(Z t ,C t )About Z t The gradient of ,λ is the dissipation coefficient,‖C t ‖ 2 is the energy of the control signal, <Z t ,C t > is the interaction term between the deformation state and the control signal, D(Z t ) is the dissipative term, modeled using fractional derivatives: Among them, Γ(·) is the gamma function, and α is the fractional order, which is used to control the degree of dissipation of deformation characteristics over time; S432. Use the non-adaptive step-size Runge-Kutta method to numerically solve the established control differential equation, calculate the rate of change of the deformation state at the current time step, use the fourth-order Runge-Kutta formula to iteratively update, calculate multiple intermediate prediction values ​​separately and take the weighted average to get the final result, then adjust the step size according to the local truncation error, repeat the calculation until the entire time interval is covered, and obtain the time series evolution trajectory of the deformation state.

7. The method for monitoring deformation of tunnel engineering structures based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. Based on the meta-learning optimization strategy, a high-order gradient optimization model is constructed to predict the output Y out For optimization, define the model parameter as θ, and the high-order gradient update formula is: in, To find the gradient of the parameters, α is the first-order gradient learning rate, β is the high-order gradient adjustment coefficient, Loss Function The first-order gradient with respect to the parameter θ is The second derivative of the loss function with respect to θ; S52. Construct a hyperbolic optimization space, introduce the Finsler metric in the hyperbolic space, and perform hypercurvature optimization on θ′: Compute the hyperbolic Finsler gradient: Among them, G θ′ is the hyperbolic Finsler metric tensor, is the inverse matrix of the Finsler metric tensor, λ is the hypercurvature scaling factor, is the Finsler metric norm; S53. Optimize θ′ using information geometry metric and calculate Fisher-Rao metric: Among them, D KL (·‖·) is the KL divergence, which adjusts the output distribution p(Y out |θ′) approximates the output distribution p(Y out ∣θ″); Information geometry optimization based on Fisher-Rao metric: in, Optimizing learning rates for information geometry; S54. Perform variational Bayesian optimization on the optimized θ′ and define the optimization objective: Calculate the final optimized parameters: in, Represents the loss function under the variational distribution q(θ′) , p(θ′) is the prior distribution of model parameters, q(θ′) is the variational posterior distribution, and ρ is the Bayesian optimization learning rate; S55. Use contrastive learning to optimize the final deformation prediction result, and finally calculate the optimized deformation prediction output: Among them, ε is the learning rate of variational Bayesian optimization, μ is the weight parameter of contrastive learning optimization, is the contrastive learning loss function.

Images