Underground comprehensive pipe gallery geological risk identification method and system

Abstract

The invention relates to the field of underground comprehensive pipe gallery engineering geological risk assessment, and provides an underground comprehensive pipe gallery geological risk identification method and system, and the method comprises the steps: obtaining geological survey data along a pipe gallery; constructing a layered decoupling variational auto-encoder, and extracting geological heterogeneity risk factors; establishing a soft soil flow field coupling time-space diagram attention network model, and performing spatial modeling to output spatial distribution characteristics; carrying out weight optimization by adopting a soft soil parameter sensitive particle swarm-reinforcement learning hybrid optimization algorithm to output an adaptive weight; constructing a memory physical information neural network model, and modeling pipe gallery-stratum interaction characteristics; establishing a stress transfer relation between pipe gallery sections, and converting the stress transfer relation into space coupling characteristics; constructing a risk prediction model and outputting a prediction result; and calculating a geological risk level based on the adaptive weight and the prediction result. According to the method, the accuracy and reliability of underground comprehensive pipe gallery geological risk identification are improved, and the requirement of high-precision risk identification in a soft soil stratum complex engineering environment is met.

Description

Technical Field

[0001] This invention relates to the field of geological risk assessment for underground utility tunnel projects, and in particular to a method and system for identifying geological risks in underground utility tunnels. Background Technology

[0002] Underground utility tunnels are a type of integrated underground infrastructure that, through unified planning, design, and construction, centrally lays various municipal utility pipelines such as electricity, communications, water supply, and gas within underground tunnel spaces. They are widely used in key engineering fields such as modern urban construction, underground space development, and municipal infrastructure. During long-term operation, underground utility tunnels are susceptible to various geological risks, including structural deformation, leakage, settlement, and cracking, due to factors such as changes in geological conditions, soft soil creep, groundwater level fluctuations, and soil settlement. Geological surveys and risk identification technologies serve as the primary means of safety assessment for underground utility tunnels, identifying and assessing the degree of geological risks by detecting changes in geological parameters and risk status along the tunnel route.

[0003] In existing technologies, geological risk identification for underground utility tunnels mainly employs traditional geological exploration methods and basic data analysis techniques to achieve basic geological risk assessment functions. However, existing methods do not adequately consider the inherent physical correlation mechanism between the complex thixotropic characteristics of soft soil and the geological risk state. They struggle to organically integrate objective geomechanical constraints with the actual creep evolution characteristics of soft soil, resulting in insufficient modeling capabilities for the time-dependent characteristics of soft soil thixotropy. Consequently, they cannot accurately describe the dynamic changes in soft soil strength over time, and thus cannot achieve high-precision automatic identification of the geological risk state of underground utility tunnels. Summary of the Invention

[0004] In view of this, the present invention proposes a geological risk identification method and system for underground integrated utility tunnels. This method solves the problems of existing methods failing to adequately consider the inherent physical correlation mechanism between the complex thixotropic characteristics of soft soil and the geological risk status, making it difficult to organically integrate objective geomechanical physical constraints with the actual creep evolution characteristics of soft soil. Consequently, the methods lack the ability to model the time-dependent characteristics of soft soil thixotropy, cannot accurately describe the dynamic change law of soft soil strength over time, and cannot achieve high-precision automatic identification of the geological risk status of underground integrated utility tunnels.

[0005] The technical solution of this invention is implemented as follows: On the one hand, this invention provides a method for identifying geological risks in underground integrated utility tunnels, including the following steps: Obtain geological survey data along the utility tunnel; A hierarchical decoupled variational autoencoder sensitive to the thixotropy of soft soil is constructed to extract features from the geological exploration data and output geological heterogeneity risk factors. A depth-adaptive spatiotemporal graph attention network model for soft soil flow field coupling is established to spatially model the geological heterogeneity risk factors and output their spatial distribution characteristics. A hybrid optimization algorithm based on memory enhancement for soft soil parameter sensitivity particle swarm optimization and reinforcement learning is used to optimize the weights of the spatial distribution features and output adaptive weights. A memory physical information neural network model based on soft soil creep history data is constructed to model the interaction of the geological heterogeneity risk factors and output the interaction characteristics of the pipe gallery-stratum. Establish the stress transfer relationship between the segments of the utility tunnel, and convert the interaction characteristics of the utility tunnel and the stratum into spatial coupling characteristics; Construct a risk prediction model and output risk prediction results based on the spatial coupling characteristics; The geological risk level is calculated based on the adaptive weights and the risk prediction results, and the geological risk identification results are output.

[0006] Based on the above technical solutions, preferably, the hierarchical decoupled variational autoencoder sensitive to the thixotropy of soft soil includes a thixotropy time-dependent modeling unit, a hierarchical decoupled coding unit, and an anisotropic feature extraction unit. The geological survey data is preprocessed in layers according to the soft soil burial depth and thixotropic strength. A thixotropic time-effect modeling unit is constructed to encode the strength recovery process of soft soil after disturbance. A layered decoupling coding unit is constructed to independently encode the consolidation parameters of soft soil at different depths. An anisotropic feature extraction unit is constructed to separately encode the horizontal and vertical parameters of soft soil. The thixotropic time-effect coding results, layered decoupling coding results, and anisotropic feature coding results are fused through a variational inference mechanism to output the geological heterogeneity risk factor.

[0007] Based on the above technical solutions, preferably, the thixotropic aging modeling unit includes: Data on the strength variation of soft soil samples under different disturbance levels over time are obtained. Thixotropic decay curves and strength recovery curves of soft soil are established. A time decay weight function is constructed to quantify the influence of thixotropy. A recurrent neural network structure is used to model the time dependence of soft soil strength. The time decay weight and the output of the recurrent neural network are weighted and combined to obtain the thixotropic time-effect encoding result. The thixotropic time-effect encoding result is input into the variational inference mechanism of the hierarchical decoupled variational autoencoder for fusion processing. The formula for calculating the thixotropic time-dependent coding result is as follows: ; in, This is the result of thixotropic time-dependent encoding; For time steps; for The degree of disturbance at any given time is Time decay weights; For recurrent neural networks in the first... Step output, For recurrent neural networks in the first... The hidden state at all times For the first The current input data at any given moment; It is the intensity recovery regulating factor; For the first Intensity recovery value at any given time; This is the initial strength value; This represents the maximum strength value.

[0008] Based on the above technical solutions, preferably, the depth-adaptive soft soil flow field coupled spatiotemporal graph attention network model includes a graph structure construction unit, a flow field coupling modeling unit, a depth-adaptive adjustment unit, and a spatiotemporal attention calculation unit; The geological heterogeneity risk factors are converted into a three-dimensional spatial node graph structure through the graph structure construction unit to obtain spatial node features; The flow field coupling modeling unit is used to couple the seepage field and stress field of soft soil to obtain the flow field coupling characteristics. The depth adaptive adjustment unit is used to adaptively adjust the network depth according to the change in soft soil layer thickness to obtain the depth weight parameters. The spatiotemporal attention calculation unit calculates the spatiotemporal correlation weights between nodes to obtain the attention weight matrix. The spatial distribution features are output by weighted fusion based on the spatial node features, flow field coupling features, depth weight parameters, and attention weight matrix.

[0009] Based on the above technical solutions, preferably, the depth adaptive adjustment unit includes: Data on the distribution of soft soil layer thickness and groundwater level changes along the utility tunnel are obtained. A mapping relationship between soft soil layer thickness and network depth is established. A depth sensitivity evaluation function is constructed to calculate the influence weight coefficients of soft soil parameters at different depths on the accuracy of network modeling. A dynamic depth selection mechanism is used to solve for the optimal number of network layers based on the influence weight coefficients. The optimal number of network layers is converted into depth weight parameters. The depth weight parameters are input into the soft soil flow field coupled spatiotemporal graph attention network model for weighted fusion calculation.

[0010] Based on the above technical solutions, preferably, the soft soil parameter-sensitive particle swarm optimization algorithm based on memory enhancement includes: The memory bank features are obtained by storing and updating historical optimization data through memory enhancement units; The sensitivity coefficients of soft soil parameters to weight optimization are calculated using the parameter sensitivity analysis unit, and the sensitivity weights are obtained. The particle swarm optimization results are obtained by performing a swarm search using particle swarm optimization units. Reinforcement learning strategies are obtained by learning strategies based on environmental feedback through reinforcement learning guidance units; The memory features, sensitivity weights, particle swarm optimization results, and reinforcement learning strategies are fused by a hybrid optimization coordination unit, and adaptive weights are obtained by coordinating and optimizing the calculations.

[0011] Based on the above technical solutions, preferably, the memory enhancement unit includes: A historical optimization data storage structure is established to store the correlation data between changes in soft soil parameters and the effects of weighted optimization. A memory decay function is constructed to assign time weights to historical data from different time periods. A similarity matching mechanism is used to retrieve historical optimization parameter configurations that are similar to the current soft soil conditions from the historical optimization data. The current optimization results are integrated into the historical database for dynamic updates through a memory update strategy. The historical optimization parameter configurations and time weights are weighted and combined to generate memory bank features. The memory bank features are then input into the hybrid optimization coordination unit for coordinated optimization calculations.

[0012] Based on the above technical solutions, preferably, the memory physical information neural network model based on soft soil creep history data includes a creep history data processing unit, a physical constraint encoding unit, a memory mechanism unit, and an interaction modeling unit; The creep history data processing unit extracts the temporal features of the creep history data of soft soil to obtain the creep temporal features. The physical laws of soft soil mechanics are encoded into constraint conditions through physical constraint coding units to obtain physical constraint characteristics. Memory-enhancing features are obtained by storing and retrieving historical interaction patterns through memory mechanism units; Based on the geological heterogeneity risk factors, the mechanical response relationship between the utility tunnel and the strata is calculated using the interaction modeling unit to obtain the interaction response characteristics; The creep time-series characteristics, physical constraint characteristics, memory enhancement characteristics, and interaction response characteristics are fused and calculated to output the pipe gallery-formation interaction characteristics.

[0013] Based on the above technical solutions, preferably, the physical constraint encoding unit includes: Data on the stress-strain relationship and parameters of the creep constitutive equation for soft soil are obtained. A physical law constraint matrix is ​​established to characterize the physical boundary conditions of the soft soil's mechanical behavior. A constraint weight allocation function is constructed to quantify the importance of different physical laws. The physical law constraint matrix and constraint weights are embedded into a neural network structure using a constraint embedding mechanism. The consistency deviation between the neural network output and the physical laws is calculated using a constraint loss calculation function. The network parameters are adjusted based on the consistency deviation feedback to generate physical constraint features. These physical constraint features are then input into the memory physical information neural network model for fusion calculation.

[0014] On the other hand, the present invention also provides a geological risk identification system for underground integrated utility tunnels, the system comprising: The geological survey data acquisition module is used to acquire geological survey data along the utility tunnel. The geological heterogeneity risk factor extraction module is used to construct a hierarchical decoupled variational autoencoder that is sensitive to the thixotropy of soft soil, extract features from the geological exploration data, and output geological heterogeneity risk factors. The spatial distribution feature modeling module is used to establish a depth-adaptive soft soil flow field coupled spatiotemporal map attention network model, to spatially model the geological heterogeneity risk factors, and output spatial distribution features. The adaptive weight optimization module is used to optimize the weights of the spatial distribution features using a memory-enhanced soft soil parameter-sensitive particle swarm optimization-reinforcement learning hybrid optimization algorithm, and output adaptive weights. The utility tunnel-soil interaction modeling module is used to construct a memory physical information neural network model based on soft soil creep history data, to model the interaction of the geological heterogeneity risk factors, and to output the utility tunnel-soil interaction characteristics. The spatial coupling feature conversion module is used to establish the stress transfer relationship between pipe gallery segments and convert the pipe gallery-soil interaction characteristics into spatial coupling characteristics. The risk prediction module is used to construct a risk prediction model and output risk prediction results based on the spatial coupling features. The geological risk identification module is used to calculate the geological risk level based on the adaptive weights and the risk prediction results, and output the geological risk identification results.

[0015] The geological risk identification method and system for underground integrated utility tunnels of the present invention have the following advantages over the prior art: (1) By using a layered decoupled variational autoencoder sensitive to the thixotropy of soft soil and a spatiotemporal graph attention network model based on depth adaptation for soft soil flow field coupling, weight adaptive optimization is performed using a memory-enhanced soft soil parameter-sensitive particle swarm-reinforcement learning hybrid optimization algorithm. Combined with a memory physical information neural network model based on soft soil creep history data, the interaction mechanism between the pipe gallery and the stratum is established. Spatial coupling features are constructed through stress transfer relationship, which improves the accuracy and reliability of geological risk identification of underground integrated pipe gallery. At the same time, memory enhancement and physical constraint guidance meet the needs of high-precision risk identification in complex engineering environments of soft soil strata. (2) By combining the thixotropic time-dependent modeling unit, the hierarchical decoupling coding unit and the anisotropic feature extraction unit, the recurrent neural network structure is used to model the time dependence of soft soil strength, the time decay weight function is constructed to quantify the thixotropic influence, the time-dependent feature characterization mechanism is established through the soft soil thixotropic decay curve and the strength recovery curve, and the multi-unit coding results are fused through the variational reasoning mechanism, which improves the pertinence of the extraction of geological heterogeneity risk factors. (3) By integrating the graph structure construction unit, the flow field coupling modeling unit, the depth adaptive adjustment unit and the spatiotemporal attention calculation unit, the network layer number adaptive solution and the influence weight coefficient of the depth sensitivity evaluation function are calculated using the dynamic depth selection mechanism. The network depth parameters are dynamically adjusted by combining the soft soil layer thickness distribution data and the groundwater level change data. The spatiotemporal correlation weight is calculated and optimized according to the coupling characteristics of the soft soil seepage field and stress field, which improves the adaptability of the spatiotemporal modeling of the soft soil flow field coupling. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a geological risk identification method for underground integrated utility tunnels according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a method for identifying geological risks in underground utility tunnels, comprising the following steps: Obtain geological survey data along the utility tunnel; A hierarchical decoupled variational autoencoder sensitive to the thixotropy of soft soil is constructed to extract features from the geological exploration data and output geological heterogeneity risk factors. A depth-adaptive spatiotemporal graph attention network model for soft soil flow field coupling is established to spatially model the geological heterogeneity risk factors and output their spatial distribution characteristics. A hybrid optimization algorithm based on memory enhancement for soft soil parameter sensitivity particle swarm optimization and reinforcement learning is used to optimize the weights of the spatial distribution features and output adaptive weights. A memory physical information neural network model based on soft soil creep history data is constructed to model the interaction of the geological heterogeneity risk factors and output the interaction characteristics of the pipe gallery-stratum. Establish the stress transfer relationship between the segments of the utility tunnel, and convert the interaction characteristics of the utility tunnel and the stratum into spatial coupling characteristics; Construct a risk prediction model and output risk prediction results based on the spatial coupling characteristics; The geological risk level is calculated based on the adaptive weights and the risk prediction results, and the geological risk identification results are output.

[0020] Specifically, this embodiment employs a layered decoupled variational autoencoder sensitive to the thixotropy of soft soil and a depth-adaptive spatiotemporal graph attention network model coupled with the soft soil flow field. It utilizes a memory-enhanced soft soil parameter-sensitive particle swarm optimization-reinforcement learning hybrid optimization algorithm for weight adaptive optimization, and combines a memory-based physical information neural network model based on soft soil creep history data to establish the interaction mechanism between the utility tunnel and the geological strata. Furthermore, it constructs spatial coupling features through stress transfer relationships. This multi-level risk prediction and weight fusion mechanism addresses the insufficient modeling of the physical correlation mechanism between the complex thixotropic characteristics of soft soil and the geological risk state, improving the accuracy and reliability of geological risk identification for underground integrated utility tunnels. Simultaneously, memory enhancement and physical constraint guidance meet the requirements for high-precision risk identification in complex engineering environments with soft soil strata. The acquisition of geological survey data along the utility tunnel includes: Based on the distribution characteristics of soft soil strata and the pipeline route, the location of geological exploration points is determined and the coordinates of the exploration points are output.

[0021] In one specific embodiment, topographic data and soft soil distribution range along the utility tunnel are acquired. Based on the gradient of soft soil layer thickness variation and the area of ​​groundwater level variation, a dense exploration interval is determined. Within the dense exploration interval, exploration points are set up at unequal intervals according to the magnitude of soft soil layer thickness variation. Smaller intervals are used in areas with large soft soil layer thickness variation, and larger intervals are used in areas with relatively stable soft soil layer thickness. The coordinates of the exploration points are then output.

[0022] Geological parameters are obtained at the coordinates of the exploration point using diverse geological exploration methods, and geological exploration data is output, including borehole data, geophysical data, and in-situ test data.

[0023] In one specific embodiment, boreholes are drilled at the coordinates of the exploration point to obtain the physical and mechanical parameters of the soft soil. Static and dynamic cone penetration tests are used to obtain the bearing capacity data of the soft soil. Ground-penetrating radar and high-density resistivity method are used to obtain the distribution data of the soft soil layer. Indoor geotechnical tests are conducted on soft soil samples at different depths to obtain thixotropic and rheological parameters. The physical and mechanical parameters of the soft soil, the bearing capacity data of the soft soil, the distribution data of the soft soil layer, the thixotropic parameters, and the rheological parameters are combined to obtain geological exploration data.

[0024] Specifically, this embodiment analyzes the characteristics of soft soil strata distribution and plans the pipeline route. It utilizes the gradient of soft soil thickness variation and the area of ​​groundwater level change to determine the density of exploration zones and employs a strategy of unequal-spacing exploration points. This is combined with diverse geological exploration methods such as borehole sampling, static cone penetration testing, dynamic cone penetration testing, ground-penetrating radar, and high-density resistivity methods to obtain the physical and mechanical parameters and thixotropic rheological properties of the soft soil. The spacing density of exploration points is dynamically adjusted according to the variation in soft soil thickness. This unequal-spacing adaptive layout mechanism solves the problem that traditional equal-spacing geological exploration cannot adapt to variations in soft soil thickness and the complexity of geological conditions, improving the relevance and comprehensiveness of geological exploration data acquisition. Furthermore, the combination of diverse exploration methods and specialized thixotropic parameter testing meets the data acquisition requirements for accurate identification of complex soft soil strata features.

[0025] The hierarchical decoupled variational autoencoder sensitive to the thixotropy of soft soil includes a thixotropy time-dependent modeling unit, a hierarchical decoupled coding unit, and an anisotropic feature extraction unit. The geological survey data is preprocessed in layers according to the soft soil burial depth and thixotropic strength. A thixotropic time-effect modeling unit is constructed to encode the strength recovery process of soft soil after disturbance. A layered decoupling coding unit is constructed to independently encode the consolidation parameters of soft soil at different depths. An anisotropic feature extraction unit is constructed to separately encode the horizontal and vertical parameters of soft soil. The thixotropic time-effect coding results, layered decoupling coding results, and anisotropic feature coding results are fused through a variational inference mechanism to output the geological heterogeneity risk factor.

[0026] The thixotropic time-dependent modeling unit includes: Data on the strength variation of soft soil samples under different disturbance levels over time are obtained. Thixotropic decay curves and strength recovery curves of soft soil are established. A time decay weight function is constructed to quantify the influence of thixotropy. A recurrent neural network structure is used to model the time dependence of soft soil strength. The time decay weight and the output of the recurrent neural network are weighted and combined to obtain the thixotropic time-dependent encoding result. The thixotropic time-dependent encoding result is input into the variational inference mechanism of the hierarchical decoupled variational autoencoder for fusion processing.

[0027] In one specific embodiment, the time decay weighting function is calculated as follows: ; in, The degree of disturbance is Recovery time is Time decay weights; The initial coefficient of thixotropic strength; The attenuation coefficient function is related to the disturbance. ; The basic attenuation coefficient; This is the disturbance impact coefficient; For perturbation nonlinearity; It is a time-decaying nonlinear exponent; This refers to the Sigmoid steepness parameter; This is the critical perturbation threshold. This is the amplitude coefficient of the periodic oscillation; The maximum disturbance level; This refers to the phase offset angle; It is an exponential function.

[0028] Specifically, the time decay weight function in this embodiment differs from traditional thixotropic modeling, which uses a single linear decay model and ignores the nonlinear influence of perturbation on the decay process. It employs a perturbation-adaptive decay mechanism, nonlinear time evolution, critical perturbation identification, and periodic perturbation effects, including: The aforementioned adaptive disturbance attenuation mechanism utilizes a basic attenuation function. and This term allows the attenuation coefficient to be dynamically adjusted according to the degree of disturbance; The nonlinear time evolution, through The term captures the non-uniform decay characteristics at different recovery stages; The critical perturbation identification is performed using the Sigmoid activation function. Achieve a smooth transition near the critical threshold; The periodic disturbance effect, through the sine term Simulate the cumulative effect of repeated loading.

[0029] The formula for calculating the thixotropic time-dependent coding result is as follows: ; in, This is the result of thixotropic time-dependent encoding; For time steps; for The degree of disturbance at any given time is Time decay weights; For recurrent neural networks in the first... Step output, For recurrent neural networks in the first... The hidden state at all times For the first The current input data at any given moment; It is the intensity recovery regulating factor; For the first Intensity recovery value at any given time; This is the initial strength value; This represents the maximum strength value.

[0030] Specifically, the thixotropic time-dependent encoding result in this embodiment differs from traditional recurrent neural networks, which cannot effectively characterize the intensity recovery characteristics of thixotropy. It employs time-varying weight modulation, intensity recovery coupling, and temporal information aggregation, including: The time-varying weight modulation will As a dynamic weight, it enables precise quantification of time decay; The intensity recovery coupling is achieved by introducing a normalized intensity recovery factor. Integrating material mechanical behavior into the coding process; The time-series information aggregation, through Accumulated thixotropic evolutionary information over multiple time steps.

[0031] This embodiment achieves refined modeling of the time-dependent thixotropic properties of soft soil by constructing a recurrent neural network encoding mechanism that couples perturbation-adaptive time decay weight function with intensity recovery. Compared with traditional methods, this embodiment can accurately capture multiple nonlinear coupling relationships between perturbation degree, recovery time, and intensity evolution, thereby improving the accuracy of thixotropic time-dependent encoding results, and significantly improving prediction reliability, especially in scenarios with repeated perturbations and long-term recovery.

[0032] The depth-adaptive soft soil flow field coupled spatiotemporal graph attention network model includes a graph structure construction unit, a flow field coupling modeling unit, a depth-adaptive adjustment unit, and a spatiotemporal attention calculation unit. The geological heterogeneity risk factors are converted into a three-dimensional spatial node graph structure through the graph structure construction unit to obtain spatial node features; The flow field coupling modeling unit is used to couple the seepage field and stress field of soft soil to obtain the flow field coupling characteristics. The depth adaptive adjustment unit is used to adaptively adjust the network depth according to the change in soft soil layer thickness to obtain the depth weight parameters. The spatiotemporal attention calculation unit calculates the spatiotemporal correlation weights between nodes to obtain the attention weight matrix. The spatial distribution features are output by weighted fusion based on the spatial node features, flow field coupling features, depth weight parameters, and attention weight matrix.

[0033] The depth adaptive adjustment unit includes: Data on the distribution of soft soil layer thickness and groundwater level changes along the utility tunnel are obtained. A mapping relationship between soft soil layer thickness and network depth is established. A depth sensitivity evaluation function is constructed to calculate the influence weight coefficients of soft soil parameters at different depths on the accuracy of network modeling. A dynamic depth selection mechanism is used to solve for the optimal number of network layers based on the influence weight coefficients. The optimal number of network layers is converted into depth weight parameters. The depth weight parameters are input into the soft soil flow field coupled spatiotemporal graph attention network model for weighted fusion calculation.

[0034] In one specific embodiment, the depth sensitivity evaluation function is calculated as follows: ; in, The thickness of the soft soil layer Time Depth sensitivity evaluation value of layer network; For the number of parameters of soft soil; For the first Normalized weights of each parameter, ; For the network in Modeling accuracy at different layers; For the first Layer thickness; For reference soft soil thickness; The standard deviation of thickness sensitivity; This is the groundwater level influence coefficient; This refers to the volume of groundwater. Total volume; It is an exponential function.

[0035] Specifically, the depth sensitivity evaluation function in this embodiment differs from traditional network depth selection, which relies on experience or grid search and lacks quantitative evaluation of soft soil parameter sensitivity. It employs multi-parameter gradient fusion, a thickness-sensitive Gaussian kernel, a groundwater influence factor, and normalized weights, including: The multi-parameter gradient fusion, through This section comprehensively considers the impact of various soft soil parameters on modeling accuracy; The thickness-sensitive Gaussian kernel, through The term is given a higher weight near the reference thickness to reflect local characteristics; The groundwater influencing factors, through Explicit modeling of the moderating effect of groundwater on depth selection; The normalized weights are obtained through Ensure a reasonable balance in the importance of each parameter.

[0036] The formula for calculating the depth weight parameter is: ; in, For depth Depth weight parameters at the location; The optimal number of network layers; The thickness of the soft soil layer Time Depth sensitivity evaluation value of layer network, The thickness of the soft soil layer Time Depth sensitivity evaluation value of layer network, For depth The thickness of the soft soil at that location; For depth The optimal layer number at that location; The standard deviation of the layer number distribution; This is the thickness adjustment coefficient; This represents the average thickness of the soft soil. This represents the standard deviation of soft soil thickness. It is the hyperbolic tangent activation function; It is an exponential function.

[0037] Specifically, the depth weight parameters in this embodiment differ from traditional fixed depth weights, which cannot adapt to different depth and thickness conditions. Instead, they employ position-related normalization, optimal Gaussian depth distribution, and nonlinear thickness adjustment, including: The location-related normalization is achieved through... The project implements dynamic normalization of sensitivity at each layer; The optimal depth Gaussian distribution, through The terms form a smooth weight distribution near the optimal layer number; The thickness nonlinear adjustment is achieved through... Weight adjustment when the thickness of the item deviates from the average value.

[0038] This embodiment achieves intelligent dynamic adjustment of network depth by establishing a depth evaluation mechanism based on multi-parameter sensitive gradients and thickness-groundwater coupling, as well as a location-adaptive depth weight allocation strategy. This embodiment overcomes the limitations of traditional fixed depths, enabling the network structure to adaptively optimize based on the soft soil thickness and groundwater conditions at different depth locations. This improves modeling accuracy under different geological conditions, while reducing invalid network layers and significantly enhancing computational efficiency and model generalization ability.

[0039] The soft soil parameter sensitivity particle swarm optimization algorithm based on memory enhancement includes: The memory bank features are obtained by storing and updating historical optimization data through memory enhancement units; The sensitivity coefficients of soft soil parameters to weight optimization are calculated using the parameter sensitivity analysis unit, and the sensitivity weights are obtained. The particle swarm optimization results are obtained by performing a swarm search using particle swarm optimization units. Reinforcement learning strategies are obtained by learning strategies based on environmental feedback through reinforcement learning guidance units; The memory features, sensitivity weights, particle swarm optimization results, and reinforcement learning strategies are fused by a hybrid optimization coordination unit, and adaptive weights are obtained by coordinating and optimizing the calculations.

[0040] The memory enhancement unit includes: A historical optimization data storage structure is established to store the correlation data between changes in soft soil parameters and the effects of weighted optimization. A memory decay function is constructed to assign time weights to historical data from different time periods. A similarity matching mechanism is used to retrieve historical optimization parameter configurations that are similar to the current soft soil conditions from the historical optimization data. The current optimization results are integrated into the historical database for dynamic updates through a memory update strategy. The historical optimization parameter configurations and time weights are weighted and combined to generate memory bank features. The memory bank features are then input into the hybrid optimization coordination unit for coordinated optimization calculations.

[0041] In one specific embodiment, the memory decay function is calculated as follows: ; in; For memory decay weights, For the current time, For historical time; This is the basic coefficient for memory strength; This is the memory decay time constant; Similarity enhancement factor; Let Gaussian similarity function be used. , Given the current soft soil parameters, Historical soft soil parameters, The standard deviation of the similarity function; It is an exponential function.

[0042] Specifically, the memory decay function in this embodiment differs from traditional memory decay, which only considers time distance and ignores the similarity between historical data and the current state. It employs exponential time decay, a similarity enhancement mechanism, dynamic weight adjustment, and a normalized similarity ratio, including: The exponential time decay, through The project establishes a basic decay mechanism based on time distance; The similarity enhancement mechanism uses The similarity of the parameter space is quantified using the Gaussian kernel function; The dynamic weight adjustment is achieved through... The coefficient controls the degree to which similarity enhances memory strength; The normalized similarity ratio is expressed in fractional form to ensure that the similarity contribution is within a reasonable range.

[0043] The formula for calculating the memory bank features is: ; in, Features of the memory bank; Total number of historical records; For the first One historical optimization parameter configuration; To normalize the optimization effect weights, For the first Optimization effect of historical records; To optimize the performance weighting factors; To optimize the variance of the results; To optimize the mean of the results; This is a small amount for numerical stability.

[0044] Specifically, the memory features in this embodiment differ from traditional memory systems that use simple averaging or nearest neighbor methods, which cannot effectively utilize the statistical characteristics of optimization history. Instead, they employ weighted memory aggregation, normalized weighted averaging, effect variance adjustment, and numerical stability, including: The weighted memory aggregation, through The overall time decay, historical configuration, and optimization effects are considered. The normalized weighted average, through the denominator The item implements automatic weight normalization; The effect variance adjustment is achieved through... The stability of the optimization was evaluated using the coefficient of variation. The numerical stability, through This item prevents division by zero errors.

[0045] This embodiment achieves intelligent retrieval and reuse of historical optimization experience by constructing a dual memory mechanism coupled with time decay and similarity, and a memory bank feature generation method based on optimization effect weighting and statistical characteristic adjustment. This embodiment can automatically identify historical successful cases similar to the current geological conditions and dynamically adjust the memory strength according to time distance and optimization effect, thereby improving the quality of parameter initialization, reducing the number of optimization iterations, and significantly accelerating the model convergence process, especially in areas with similar geological conditions.

[0046] The memory-based physical information neural network model based on soft soil creep history data includes a creep history data processing unit, a physical constraint encoding unit, a memory mechanism unit, and an interaction modeling unit. The creep history data processing unit extracts the temporal features of the creep history data of soft soil to obtain the creep temporal features. The physical laws of soft soil mechanics are encoded into constraint conditions through physical constraint coding units to obtain physical constraint characteristics. Memory-enhancing features are obtained by storing and retrieving historical interaction patterns through memory mechanism units; Based on the geological heterogeneity risk factors, the mechanical response relationship between the utility tunnel and the strata is calculated using the interaction modeling unit to obtain the interaction response characteristics; The creep time-series characteristics, physical constraint characteristics, memory enhancement characteristics, and interaction response characteristics are fused and calculated to output the pipe gallery-formation interaction characteristics.

[0047] The physical constraint coding unit includes: Data on the stress-strain relationship and parameters of the creep constitutive equation for soft soil are obtained. A physical law constraint matrix is ​​established to characterize the physical boundary conditions of the soft soil's mechanical behavior. A constraint weight allocation function is constructed to quantify the importance of different physical laws. The physical law constraint matrix and constraint weights are embedded into a neural network structure using a constraint embedding mechanism. The consistency deviation between the neural network output and the physical laws is calculated using a constraint loss calculation function. The network parameters are adjusted based on the consistency deviation feedback to generate physical constraint features. These physical constraint features are then input into the memory physical information neural network model for fusion calculation.

[0048] In one specific embodiment, the formula for calculating the constraint weights is: ; in, For the first The constraint weights of a physical law; For the first The importance coefficient of each law For the first The importance coefficient of each law; For the first The gradient magnitude of a physical law, For the first The gradient magnitude of a physical law; The total number of physical laws; As an accuracy adjustment factor; Use the Sigmoid activation function; For the first The accuracy of the law's conformity; This is the minimum value that meets the accuracy requirement; To meet the maximum precision.

[0049] Specifically, the constraint weights in this embodiment differ from traditional physical constraints which use fixed weights and cannot reflect the differences in importance of each law under different conditions. Instead, they employ gradient magnitude normalization, adaptive accuracy enhancement, importance prior, and multiplicative adjustment, including: The gradient magnitude normalization is achieved through... The weights of each term are dynamically assigned based on the magnitude of the gradient of each physical law; The accuracy adaptive enhancement, through Items that conform to laws with high precision are given increased weight; The aforementioned importance prior, through Preserve expert knowledge regarding the relative importance of each law; The multiplicative adjustment achieves synergistic optimization of prior importance and real-time accuracy through the multiplication of two terms.

[0050] The formula for calculating the constraint loss is: ; in, To constrain the loss value, For neural network output; For the first The theoretical values ​​of several physical laws; These are the spatial gradient constraint weights; For spatial gradient operators; Weights are constrained by the time derivative; It is the time partial derivative operator.

[0051] Specifically, the constraint loss calculation function in this embodiment differs from traditional physical constraints, which only consider function value errors and ignore the physical consistency of gradients and time derivatives. It employs multi-level constraints, hierarchical weight adjustment, adaptive weighting, and L2 norm form, including: The multi-level constraints include: simultaneously constraining function values. Spatial gradient and time derivative ; The hierarchical weight adjustment includes: through... and Control the strength of gradient and derivative constraints separately; The adaptive weighting includes: through... To achieve differentiated constraints on various physical laws; The L2 norm form is used to ensure the convexity and differentiability of the loss function.

[0052] This embodiment achieves intelligent encoding and constraint of multiple physical laws by establishing a dynamic weight allocation mechanism based on gradient magnitude and compliance accuracy, and a physical loss function with triple constraints of function value, gradient, and derivative. This embodiment can adaptively adjust the constraint strength according to the real-time performance of each physical law during training, while ensuring the physical consistency of the prediction results through multi-level constraints. This improves the average compliance accuracy of the model with physical laws, significantly enhancing prediction reliability, especially under extreme conditions, and effectively avoiding the non-physical interpretation of neural networks.

[0053] The process of establishing stress transfer relationships between pipe gallery segments and converting the pipe gallery-soil interaction characteristics into spatial coupling characteristics includes: A stress transfer model between segments is established. Based on the geometric parameters and connection methods of the pipe gallery segments, the stress transfer between segments is calculated to obtain the stress transfer characteristics of the pipe gallery-soil interaction.

[0054] In one specific embodiment, establishing the inter-segment stress transfer model includes: The length parameters, cross-sectional parameters, and node connection stiffness data of the utility tunnel segments are obtained. A segment stiffness matrix is ​​established to characterize the structural characteristics of each segment. A stress transfer weighting function is constructed to calculate the stress transfer coefficient between adjacent segments. The interaction characteristics between the utility tunnel and the stratum are decomposed into segments using a segment coupling calculation mechanism. The stress distribution and transfer path between segments are calculated using the stress transfer coefficient and the segment stiffness matrix. The stress distribution and transfer path are combined to obtain the segment stress transfer characteristics.

[0055] A spatial coupling feature transformation model is constructed to couple and map the stress transmission characteristics of the segments with the spatial distribution information of the strata, and output the spatial coupling features.

[0056] In one specific embodiment, constructing the spatially coupled feature transformation model includes: The spatial coordinate data and stratum spatial distribution information along the utility tunnel are acquired. A spatial mapping relationship is established to associate the stress transmission characteristics of the segments with the three-dimensional spatial location. A coupling strength evaluation function is established to calculate the mutual influence intensity between the utility tunnel segments and the surrounding strata. A spatial interpolation algorithm is used to perform spatial continuity processing on the stress transmission characteristics between segments. Through the spatial mapping relationship and coupling strength, the stress transmission characteristics of the segments are converted into continuous three-dimensional spatial distribution data. The three-dimensional spatial distribution data is then standardized to obtain spatial coupling characteristics.

[0057] Specifically, this embodiment integrates the inter-segment stress transfer model with the spatial coupling feature transformation model. It utilizes a stress transfer weighting function to calculate the segment stress transfer coefficient and the segment stiffness matrix to characterize structural properties. A spatial mapping relationship is established by combining spatial coordinate data and stratum spatial distribution information. Furthermore, the interaction strength between the utility tunnel segments and the surrounding strata is dynamically evaluated based on the coupling strength evaluation function. Through spatial interpolation algorithms and segment coupling calculation mechanisms, the modeling problem of transforming the utility tunnel-stratum interaction characteristics from discrete segments to a continuous three-dimensional spatial distribution is solved, improving the accuracy and continuity of the spatial coupling feature representation. Simultaneously, the calculation of stress distribution transfer paths and the standardization of three-dimensional spatial distribution data meet the precise modeling requirements for stress transfer between utility tunnel segments and spatial coupling with the strata.

[0058] The construction of the risk prediction model, which outputs risk prediction results based on the spatial coupling features, includes: A multi-level risk prediction model architecture is constructed. Based on the geological risk type and prediction accuracy requirements, a multi-level model structure is established, which includes a feature extraction layer, a risk classification layer, and a prediction output layer, thus obtaining the risk prediction model architecture.

[0059] In one specific embodiment, the construction of the multi-level risk prediction model architecture includes: Geological risk classification standards and prediction accuracy threshold data are obtained, a risk type mapping table is established, and classification labels for different geological risks are defined. Multi-scale feature extraction is performed on the input features through a feature extraction layer structure. A risk classification layer structure is constructed to achieve automatic identification and classification of risk types. A prediction output layer structure is established to convert the classification results into risk intensity values. A multi-level model structure is constructed by cascading the feature extraction layer, risk classification layer, and prediction output layer. The parameters of the multi-level model structure are initialized to generate a risk prediction model architecture.

[0060] Perform risk prediction calculations based on spatial coupling features, input the spatial coupling features into the risk prediction model architecture for multi-level prediction processing, and output the risk prediction results.

[0061] In one specific embodiment, the risk prediction calculation of the spatial coupling characteristics includes: The dimensional information and numerical range of the spatial coupling features are obtained, and feature standardization is performed to convert the spatial coupling features into the model input format. Multi-scale risk features are extracted through the feature extraction layer of the risk prediction model architecture to obtain the extracted feature results. The risk classification layer is used to identify the risk type of the extracted features to obtain the risk classification result. The prediction output layer converts the risk classification result into a risk intensity value to obtain a numerical prediction result. The numerical prediction results are post-processed and format-converted to generate risk prediction results.

[0062] Specifically, this embodiment integrates a multi-level risk prediction model architecture with spatially coupled feature-based risk prediction calculations. It utilizes a cascaded structure of feature extraction, risk classification, and prediction output layers to extract multi-scale risk features and automatically identify and classify risk types. A risk intensity numerical conversion mechanism is established by combining a risk type mapping table and classification label definitions. Model parameters are dynamically adjusted based on geological risk classification standards and prediction accuracy thresholds. Through multi-level prediction processing and feature standardization conversion mechanisms, the problems of accurate conversion of spatially coupled features to risk prediction results and unified modeling of risk type identification and intensity quantification are solved, improving the accuracy of geological risk prediction. Furthermore, through multi-scale feature extraction and post-processing of numerical prediction results, the need for multi-level accurate prediction of geological risks in underground integrated utility tunnels is met.

[0063] The process of calculating the geological risk level based on the adaptive weights and the risk prediction results, and outputting the geological risk identification results, includes: An adaptive weight fusion calculation is performed, and the adaptive weights are weighted and fused with the risk prediction results. A comprehensive risk index is calculated based on the weight distribution characteristics to obtain the fused risk index.

[0064] In one specific embodiment, the adaptive weight fusion calculation includes: Obtain the weight distribution data of the adaptive weights and the numerical distribution data of the risk prediction results, establish a weight matching relationship, correspond the weight values ​​to the prediction results in terms of spatial location, establish a fusion coefficient calculation function, determine the weight fusion ratio of different regions, use a weighted average algorithm to weight the risk prediction results, calculate the regional comprehensive risk intensity through the weight distribution characteristics and fusion coefficient, normalize the regional comprehensive risk intensity, and generate a fusion risk index.

[0065] Geological risk level determination calculation is performed, risk level classification standards and threshold determination rules are established, the fused risk index is compared and calculated with the risk level threshold, and the geological risk identification result is output.

[0066] In one specific embodiment, the calculation of geological risk level determination includes: Geological risk classification standard data is acquired, a multi-level risk level system is established, threshold judgment rules are constructed, numerical boundary conditions for each risk level are defined, and a hierarchical mapping algorithm is used to compare the fused risk index with the threshold of each level step by step. The risk level category of each spatial location is determined through the hierarchical mapping algorithm, a risk identification result format standard is established, the risk level is encoded and labeled, and the risk level category is combined with the spatial location information to output the geological risk identification result.

[0067] Specifically, this embodiment employs adaptive weighted fusion calculation and geological risk level determination calculation. It utilizes a weighted average algorithm and fusion coefficient calculation function for weighted calculation and regional comprehensive risk intensity calculation. A hierarchical mapping mechanism is established by combining a multi-level risk level system and threshold determination rules, and the fusion ratio is dynamically adjusted based on the weight distribution characteristics and the numerical distribution of risk prediction results. Through weight matching relationships and hierarchical mapping algorithms, the problem of effectively fusing adaptive weights with risk prediction results and accurately classifying risk levels is solved, improving the accuracy and standardization of geological risk identification results. Furthermore, through regional comprehensive risk intensity normalization processing and risk level coding labeling, the standardized output technology requirements for geological risk level identification of underground integrated utility tunnels are met.

[0068] In one specific embodiment, the application of risk identification for the soft soil strata section of an underground integrated pipe gallery in a certain city includes: The project background includes: a city planning to construct an underground utility tunnel approximately 8.5 kilometers long. This tunnel traverses the city center, with complex geological conditions along its route, including a soft soil layer 8 to 15 meters thick. This soft soil layer exhibits significant thixotropic characteristics, with a water content as high as 45% to 60%, and a groundwater level only 2 to 4 meters deep. Preliminary geological surveys indicate that the soft soil in this area has low shear strength and significant creep characteristics, making it prone to continuous settlement under long-term loads, posing a potential risk to the structural safety of the utility tunnel.

[0069] The risk identification steps are as follows: The first step was to acquire geological survey data: 86 survey points were set up in a 2.6-kilometer soft soil section. 430 sets of soft soil samples were obtained by means of drilling, static cone penetration, ground-penetrating radar and high-density resistivity method. 86 sets of thixotropic tests were completed to obtain physical parameters, mechanical parameters, thixotropic parameters and rheological properties.

[0070] The second step involves extracting geological heterogeneity risk factors: a variational autoencoder is constructed, comprising a thixotropic time-dependent modeling unit, a hierarchical decoupling coding unit, and an anisotropic feature extraction unit. The thixotropic time-dependent modeling employs a recurrent neural network, observing the intensity recovery process over 7 days, with 65% of the recovery weight allocated to the first 48 hours for strongly thixotropic soft soil.

[0071] Layered decoupled coding was used to independently encode shallow, medium, and deep soft soil layers. The consolidation coefficient for shallow soft soil was 0.8 cm² / s, and for deep soft soil it was 0.3 cm² / s. The horizontal permeability coefficient was identified through anisotropic feature extraction. Significantly larger than the vertical direction The geological heterogeneity risk factor, consisting of 86 nodes and 128 dimensions per vector, is fused through variational inference.

[0072] The third step involves spatial distribution feature modeling: converting 86 survey points into a 3D spatial node map and establishing 238 edge connections. Coupled modeling of the soft soil seepage field and stress field is performed, linking pore water pressure and effective stress through the effective stress principle. The network depth is dynamically adjusted according to the soft soil layer thickness: 5 layers are set for thin-layer areas, and 8 layers are added for thick-layer areas, improving modeling accuracy by 18%. Thin-layer areas are defined as areas with a thickness less than 10 meters, and thick-layer areas are defined as areas with a thickness greater than 12 meters. Spatiotemporal correlation weights are calculated: 0.75 for adjacent nodes, 0.1 for mid-distance nodes, and 0.15 for far-distance nodes, with a 5-year observation cycle used to analyze time-varying characteristics. The fused output provides a complete description of the spatial distribution characteristics of the soft soil strata's 3D spatial heterogeneity and spatiotemporal evolution.

[0073] The fourth step involves adaptive weight optimization: A historical database was established, collecting 120 historical records of six utility tunnel projects in the city. A memory decay function was used, with recent data within the last two years having a weight of 0.85 and earlier data from five years ago having a weight of 0.35. Similarity matching revealed that historical cases No. 37 and No. 62 were highly similar to the current project. Parameter sensitivity coefficients were calculated, with thixotropy showing the highest strength at 0.92, followed by water content at 0.78. A particle swarm size of 50 and 200 iterations were set, converging after 127 iterations. A reinforcement learning environment was constructed, and the optimization strategy was learned after 3000 training rounds. Hybrid optimization improved the modeling accuracy from 76.5% to 92.3%.

[0074] The fifth step involves modeling the interaction between the utility tunnel and the soil: collecting historical data on building settlement (15 years), road creep (10 years), and subway deformation (5 years). Based on four physical laws—the coded effective stress principle, the Mohr-Coulomb failure criterion, Terzaghi consolidation theory, and the Burgers creep model—dynamic constraint weights are assigned. The constraint loss is reduced from an initial 0.42 to 0.08. Twelve historical interaction patterns are retrieved and matched against differential settlement patterns, with a similarity of 0.81. The stress distribution of the utility tunnel is calculated, yielding a top earth pressure of 85 kPa, a bottom reaction force of 110 kPa, and sidewall stresses of 30 to 50 kPa. This predicts an additional stress of 45 kPa in the soft soil 3 meters directly beneath the utility tunnel, with a maximum settlement of 35 mm over 5 years.

[0075] Step 6: Perform spatial coupling feature transformation: Divide the utility tunnel into 52 segments, each 50 meters long, with a cross-section of 4 meters by 3 meters and a wall thickness of 0.4 meters. Establish the segment stiffness matrix and calculate the stress transfer coefficient, which is 0.55 at the joints and close to 1.0 in the middle. It was found that a certain segment experienced a maximum bending moment of 280 kNm due to a sudden change in the thickness of the soft soil layer. Kriging interpolation was used to achieve spatial continuity with a resolution of 1 meter, forming a spatial coupling feature tensor of 260×40×15×8.

[0076] Step 7: Risk prediction. A 3D convolutional neural network was constructed, consisting of four convolutional layers with 32, 64, 128, and 256 channels respectively, and a fully connected network with three hidden layers and 512, 256, and 128 neurons respectively. Using 96 training samples, 24 validation samples, and 30 test samples, after 500 training rounds, the loss was reduced from 2.35 to 0.12, and the validation accuracy was improved from 65.8% to 91.7%. Eighteen segmental settlement risks, 12 deformation risks, eight insufficient bearing capacity risks, and six leakage risks were identified. The risk intensity of the highest risk area ranged from 82 to 88, and the risk intensity of the lowest risk area ranged from 22 to 35.

[0077] Step 8: Calculate the geological risk level: Establish a five-level risk system, including extremely high, high, medium, low, and extremely low risks. Adaptive weighting is used, with a weight of 0.78 for high-risk areas and 0.42 for low-risk areas. Calculate the merged risk value, with a maximum merged risk value of 68.6 and a minimum merged risk value of 15.3.

[0078] The results showed 3 extremely high-risk segments, 15 high-risk segments, 21 medium-risk segments, 11 low-risk segments, and 2 extremely low-risk segments. The output included a complete identification result comprising a risk level distribution map, a risk type distribution table, a risk intensity curve, and a list of key risk points.

[0079] The present invention also provides a geological risk identification system for underground integrated utility tunnels, the system comprising: The geological survey data acquisition module is used to acquire geological survey data along the utility tunnel. The geological heterogeneity risk factor extraction module is used to construct a hierarchical decoupled variational autoencoder that is sensitive to the thixotropy of soft soil, extract features from the geological exploration data, and output geological heterogeneity risk factors. The spatial distribution feature modeling module is used to establish a depth-adaptive soft soil flow field coupled spatiotemporal map attention network model, to spatially model the geological heterogeneity risk factors, and output spatial distribution features. The adaptive weight optimization module is used to optimize the weights of the spatial distribution features using a memory-enhanced soft soil parameter-sensitive particle swarm optimization-reinforcement learning hybrid optimization algorithm, and output adaptive weights. The utility tunnel-soil interaction modeling module is used to construct a memory physical information neural network model based on soft soil creep history data, to model the interaction of the geological heterogeneity risk factors, and to output the utility tunnel-soil interaction characteristics. The spatial coupling feature conversion module is used to establish the stress transfer relationship between pipe gallery segments and convert the pipe gallery-soil interaction characteristics into spatial coupling characteristics. The risk prediction module is used to construct a risk prediction model and output risk prediction results based on the spatial coupling features. The geological risk identification module is used to calculate the geological risk level based on the adaptive weights and the risk prediction results, and output the geological risk identification results.

[0080] Specifically, this embodiment of an underground integrated utility tunnel geological risk identification system utilizes eight functional modules: geological survey data acquisition, extraction of geological heterogeneity risk factors, spatial distribution feature modeling, adaptive weight optimization, utility tunnel-stratum interaction modeling, spatial coupling feature transformation, risk prediction, and geological risk identification. It employs a hierarchical decoupled variational autoencoder, a deep adaptive soft soil flow field coupled spatiotemporal graph attention network, a memory-enhanced particle swarm optimization-reinforcement learning hybrid optimization algorithm, and a memory physical information neural network for multi-level intelligent modeling. Combined with the extraction of thixotropic sensitive features of soft soil and the stress transfer relationship between utility tunnel segments, a full-process risk identification mechanism is established. Furthermore, the system dynamically optimizes identification accuracy based on historical soft soil creep data and spatial coupling characteristics. This addresses the lack of systematicity, intelligence, and accuracy in underground integrated utility tunnel geological risk identification, thereby improving the systematicness, accuracy, and automation of geological risk identification.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying geological risks in underground utility tunnels, characterized in that, Includes the following steps: Obtain geological survey data along the utility tunnel; A hierarchical decoupled variational autoencoder sensitive to the thixotropy of soft soil is constructed to extract features from the geological exploration data and output geological heterogeneity risk factors. A depth-adaptive spatiotemporal graph attention network model for soft soil flow field coupling is established to spatially model the geological heterogeneity risk factors and output their spatial distribution characteristics. A hybrid optimization algorithm based on memory enhancement for soft soil parameter sensitivity particle swarm optimization and reinforcement learning is used to optimize the weights of the spatial distribution features and output adaptive weights. A memory physical information neural network model based on soft soil creep history data is constructed to model the interaction of the geological heterogeneity risk factors and output the interaction characteristics of the pipe gallery-stratum. Establish the stress transfer relationship between the segments of the utility tunnel, and convert the interaction characteristics of the utility tunnel and the stratum into spatial coupling characteristics; Construct a risk prediction model and output risk prediction results based on the spatial coupling characteristics; The geological risk level is calculated based on the adaptive weights and the risk prediction results, and the geological risk identification results are output.

2. The method for geological risk identification of underground integrated utility tunnels as described in claim 1, characterized in that, The hierarchical decoupled variational autoencoder sensitive to the thixotropy of soft soil includes a thixotropy time-dependent modeling unit, a hierarchical decoupled coding unit, and an anisotropic feature extraction unit. The geological survey data is preprocessed in layers according to the soft soil burial depth and thixotropic strength. A thixotropic time-effect modeling unit is constructed to encode the strength recovery process of soft soil after disturbance. A layered decoupling coding unit is constructed to independently encode the consolidation parameters of soft soil at different depths. An anisotropic feature extraction unit is constructed to separately encode the horizontal and vertical parameters of soft soil. The thixotropic time-effect coding results, layered decoupling coding results, and anisotropic feature coding results are fused through a variational inference mechanism to output the geological heterogeneity risk factor.

3. The method for geological risk identification of underground integrated utility tunnels as described in claim 2, characterized in that, The thixotropic time-dependent modeling unit includes: Data on the strength variation of soft soil samples under different disturbance levels over time are obtained. Thixotropic decay curves and strength recovery curves of soft soil are established. A time decay weight function is constructed to quantify the influence of thixotropy. A recurrent neural network structure is used to model the time dependence of soft soil strength. The time decay weight and the output of the recurrent neural network are weighted and combined to obtain the thixotropic time-effect encoding result. The thixotropic time-effect encoding result is input into the variational inference mechanism of the hierarchical decoupled variational autoencoder for fusion processing. The formula for calculating the thixotropic time-dependent coding result is as follows: ； in, This is the result of thixotropic time-dependent encoding; For time steps; for The degree of disturbance at any given time is Time decay weight; For recurrent neural networks in the first... Step output, For recurrent neural networks in the first... The hidden state at all times For the first The current input data at any given moment; It is the intensity recovery regulating factor; For the first Intensity recovery value at any given time; This is the initial strength value; This represents the maximum strength value.

4. The method for geological risk identification of underground integrated utility tunnels as described in claim 1, characterized in that, The depth-adaptive soft soil flow field coupled spatiotemporal graph attention network model includes a graph structure construction unit, a flow field coupling modeling unit, a depth-adaptive adjustment unit, and a spatiotemporal attention calculation unit. The geological heterogeneity risk factors are converted into a three-dimensional spatial node graph structure through the graph structure construction unit to obtain spatial node features; The flow field coupling modeling unit is used to couple the seepage field and stress field of soft soil to obtain the flow field coupling characteristics. The depth adaptive adjustment unit is used to adaptively adjust the network depth according to the change in soft soil layer thickness to obtain the depth weight parameters. The spatiotemporal attention calculation unit calculates the spatiotemporal correlation weights between nodes to obtain the attention weight matrix. The spatial distribution features are output by weighted fusion based on the spatial node features, flow field coupling features, depth weight parameters, and attention weight matrix.

5. The method for geological risk identification of underground integrated utility tunnels as described in claim 4, characterized in that, The depth adaptive adjustment unit includes: Data on the distribution of soft soil layer thickness and groundwater level changes along the utility tunnel are obtained. A mapping relationship between soft soil layer thickness and network depth is established. A depth sensitivity evaluation function is constructed to calculate the influence weight coefficients of soft soil parameters at different depths on the accuracy of network modeling. A dynamic depth selection mechanism is used to solve for the optimal number of network layers based on the influence weight coefficients. The optimal number of network layers is converted into depth weight parameters. The depth weight parameters are input into the soft soil flow field coupled spatiotemporal graph attention network model for weighted fusion calculation.

6. The method for geological risk identification of underground integrated utility tunnels as described in claim 1, characterized in that, The soft soil parameter sensitivity particle swarm optimization algorithm based on memory enhancement includes: The memory bank features are obtained by storing and updating historical optimization data through memory enhancement units; The sensitivity coefficients of soft soil parameters to weight optimization are calculated using the parameter sensitivity analysis unit, and the sensitivity weights are obtained. The particle swarm optimization results are obtained by performing a swarm search using particle swarm optimization units. Reinforcement learning strategies are obtained by learning strategies based on environmental feedback through reinforcement learning guidance units; The memory features, sensitivity weights, particle swarm optimization results, and reinforcement learning strategies are fused by a hybrid optimization coordination unit, and adaptive weights are obtained by coordinating and optimizing the calculations.

7. The method for identifying geological risks in underground integrated utility tunnels as described in claim 6, characterized in that, The memory enhancement unit includes: A historical optimization data storage structure is established to store the correlation data between changes in soft soil parameters and the effects of weighted optimization. A memory decay function is constructed to assign time weights to historical data from different time periods. A similarity matching mechanism is used to retrieve historical optimization parameter configurations that are similar to the current soft soil conditions from the historical optimization data. The current optimization results are integrated into the historical database for dynamic updates through a memory update strategy. The historical optimization parameter configurations and time weights are weighted and combined to generate memory bank features. The memory bank features are then input into the hybrid optimization coordination unit for coordinated optimization calculations.

8. The method for geological risk identification of underground integrated utility tunnels as described in claim 1, characterized in that, The memory-based physical information neural network model based on soft soil creep history data includes a creep history data processing unit, a physical constraint encoding unit, a memory mechanism unit, and an interaction modeling unit. The creep history data processing unit extracts the temporal features of the creep history data of soft soil to obtain the creep temporal features. The physical laws of soft soil mechanics are encoded into constraint conditions through physical constraint coding units to obtain physical constraint characteristics. Memory-enhancing features are obtained by storing and retrieving historical interaction patterns through memory mechanism units; Based on the geological heterogeneity risk factors, the mechanical response relationship between the utility tunnel and the strata is calculated using the interaction modeling unit to obtain the interaction response characteristics; The creep time-series characteristics, physical constraint characteristics, memory enhancement characteristics, and interaction response characteristics are fused and calculated to output the pipe gallery-formation interaction characteristics.

9. The method for identifying geological risks in underground integrated utility tunnels as described in claim 8, characterized in that, The physical constraint coding unit includes: Data on the stress-strain relationship and parameters of the creep constitutive equation for soft soil are obtained. A physical law constraint matrix is ​​established to characterize the physical boundary conditions of the soft soil's mechanical behavior. A constraint weight allocation function is constructed to quantify the importance of different physical laws. The physical law constraint matrix and constraint weights are embedded into a neural network structure using a constraint embedding mechanism. The consistency deviation between the neural network output and the physical laws is calculated using a constraint loss calculation function. The network parameters are adjusted based on the consistency deviation feedback to generate physical constraint features. These physical constraint features are then input into the memory physical information neural network model for fusion calculation.

10. A geological risk identification system for underground utility tunnels, used to execute the geological risk identification method for underground utility tunnels as described in any one of claims 1-9, characterized in that, The system includes: The geological survey data acquisition module is used to acquire geological survey data along the utility tunnel. The geological heterogeneity risk factor extraction module is used to construct a hierarchical decoupled variational autoencoder that is sensitive to the thixotropy of soft soil, extract features from the geological exploration data, and output geological heterogeneity risk factors. The spatial distribution feature modeling module is used to establish a depth-adaptive soft soil flow field coupled spatiotemporal map attention network model, to spatially model the geological heterogeneity risk factors, and output spatial distribution features. The adaptive weight optimization module is used to optimize the weights of the spatial distribution features using a memory-enhanced soft soil parameter-sensitive particle swarm optimization-reinforcement learning hybrid optimization algorithm, and output adaptive weights. The utility tunnel-soil interaction modeling module is used to construct a memory physical information neural network model based on soft soil creep history data, to model the interaction of the geological heterogeneity risk factors, and to output the utility tunnel-soil interaction characteristics. The spatial coupling feature conversion module is used to establish the stress transfer relationship between pipe gallery segments and convert the pipe gallery-soil interaction characteristics into spatial coupling characteristics. The risk prediction module is used to construct a risk prediction model and output risk prediction results based on the spatial coupling features. The geological risk identification module is used to calculate the geological risk level based on the adaptive weights and the risk prediction results, and output the geological risk identification results.

Images