Real-time dynamic modeling method for tunnel safety based on coupling of rock stress and deformation

Abstract

The application belongs to the technical field of tunnel safety monitoring, more specifically relates to a tunnel safety real-time dynamic modeling method of rock stress-deformation coupling, which embeds rock mass mechanical relationship as a hard constraint into a neural network model, establishes an explicit mapping of monitoring data and physical field, adopts set Kalman filtering to dynamically optimize model parameter field in real time based on field detection data, designs a data-driven and physical mechanism collaborative deduction mechanism, effectively suppresses long-term prediction error accumulation, directly calculates disaster risk indexes with clear mechanical significance by using stress field, deformation field and other physical quantities deduced by the model, and realizes physically interpretable early warning decision through fuzzy reasoning fusion of multi-source risks; The technical problems of physical misalignment of data-driven model, weak long-term generalization and poor early warning interpretability are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tunnel safety monitoring, and more particularly to a rock stress-deformation coupling tunnel safety real-time dynamic modeling method. BACKGROUND

[0002] With the complex and changeable geological environment faced by engineering construction and operation, real-time perception and accurate prediction of the stress-deformation state of surrounding rock are core challenges to ensure construction and structural safety. Traditional safety monitoring mainly relies on real-time data-driven methods based on field sensors. Such methods can quickly respond to changes in monitoring data, but lack embedding of rock mechanics constitutive laws. Essentially, they are "black box" or "gray box" models, whose prediction results are easily disturbed by noise and have weak extrapolation ability. In particular, the reliability is insufficient in predicting future stress / deformation evolution and inferring the state of unmeasured points, and it is difficult to ensure that the prediction results comply with basic physical laws. Therefore, using such data-driven models leads to technical problems such as prediction results that violate basic mechanics principles, poor generalization, weak interpretability, and unreliable long-term prediction. SUMMARY

[0003] The present application provides a rock stress-deformation coupling tunnel safety real-time dynamic modeling method, which aims to solve the technical problems of current data-driven model prediction methods, such as prediction results that violate basic mechanics principles, poor generalization, weak interpretability, and unreliable long-term prediction.

[0004] The rock stress-deformation coupling tunnel safety real-time dynamic modeling method comprises the following steps:

[0005] S1. Collecting tunnel surrounding rock stress, deformation, and environmental data, binding the space-time label through Beidou timing and tunnel mileage coordinates, and establishing a sensor-rock mass unit mapping matrix to associate the data to the discretized rock mass grid unit, obtaining a multi-dimensional sensor data stream aligned in space and time;

[0006] S2. Based on the multi-dimensional sensor data stream aligned in space and time, a physical mechanism guided space-time convolutional neural network is constructed. In network training, the rock mass constitutive equation is embedded as a constraint term, while extracting the space-time coupling features. The real-time sensor data is converted to the prediction field using the set Kalman filter, and the rock mass parameter space distribution is dynamically corrected to obtain the dynamically updated stress tensor field and deformation tensor field of the whole tunnel;

[0007] S3. Based on the stress tensor field and the deformation tensor field, a double-engine deduction based on the combination of mechanism engine and data engine is adopted. The double-engine deduction results are weighted and fused according to the confidence, and the deduction time window is adaptively adjusted, and the future stress, deformation evolution field and uncertainty quantification results are output;

[0008] S4. Based on the future gravity and deformation evolution field, the triple risk index of each rock mass unit is calculated in real time, multi-index fusion is carried out based on the fuzzy inference system, the risk level is dynamically divided combined with the change rate, and the risk confidence is generated by associating the uncertainty quantization result, to obtain the real-time risk thermal map of the whole space of the tunnel and the positioning warning information;

[0009] S5. Real-time acquisition of actual response data, when the actual response data deviates from the warning result, triggering the closed-loop optimization.

[0010] The application embeds the rock mass mechanical relationship as a hard constraint into the neural network model, establishes an explicit mapping between the monitoring data and the physical field, ensures that the model structure strictly follows the physical principle, and then uses the ensemble Kalman filter to dynamically optimize the model parameter field in real time. Field detection data, and design a data-driven and physical mechanism co-evolution mechanism, effectively suppresses the long-term prediction error accumulation, significantly improves the adaptability and prediction stability of the model under complex geological conditions; On this basis, the stress field, deformation field and other physical quantities calculated by the model evolution output are directly used to calculate the disaster risk index with clear mechanical significance, and the physical interpretable warning decision is realized through the fuzzy inference of multi-source risk; Finally, through the dynamic feedback loop of the actual disaster response data to the warning result, the key parameters and warning thresholds of the model are continuously calibrated, forming a self-iterative optimization warning system, thereby systematically solving the technical problems of physical misalignment, weak long-term generalization and poor warning interpretability of data-driven models.

[0011] Preferably, the binding space-time label comprises the following steps:

[0012] Time reference synchronization: all data acquisition sensors are equipped with clock chips, which are synchronized at the hardware level through the Beidou second pulse signal, and the original time stamp is corrected in two steps: first, compensate for the fixed transmission delay, and then eliminate the clock crystal oscillator jitter error through the Kalman filter to obtain the sensor data after time stamp alignment;

[0013] Spatial coordinate conversion: taking the tunnel entrance control point as the origin and the tunnel design centerline as the reference axis to establish the tunnel mileage coordinate system;

[0014] Convert the original sensor coordinates to the tunnel mileage coordinate system: eliminate the control point offset by coordinate translation, and correct the tunnel orientation deviation by rotation matrix;

[0015] Rock mass grid construction: discretize the tunnel surrounding rock space into tetrahedral element grids, assign a unique identifier to each grid element, and record the spatial position of each node.

[0016] Preferably, the data processing steps of the physical mechanism guided space-time convolutional neural network are as follows:

[0017] Spatial feature extraction: sliding in the spatial dimension with a pre-defined size of convolution kernel, extracting local spatial features;

[0018] Temporal feature extraction: unfolding in the temporal dimension with a pre-defined size of convolution kernel, extracting dynamic features in time series;

[0019] Feature fusion: fusing spatial and temporal features, generating high-level feature representation, and outputting high-level feature tensor based on this.

[0020] Preferably, a physical mechanism is introduced into the loss function of the spatio-temporal convolutional neural network:

[0021] The loss function includes data fitting loss and physical constraint loss; the data fitting loss measures the difference between the predicted field and the true field using mean square error; the physical loss constraint calculates the residual between the predicted maximum principal stress and the equivalent strain based on the Hoek-Brown criterion; and the weights of the data fitting loss and the physical constraint loss are adjusted to ensure balance between the two during training.

[0022] Preferably, the mechanism engine is based on the explicit finite difference method, and uses the updated rock mass parameter field to calculate the stress-deformation evolution in the future time, including the following steps:

[0023] Time step advancement: advancing the calculation step by step according to the set time step, in each time step, using the current gravity and displacement data to calculate the stress and displacement change of the next time step; at the end of each time step, updating the stress and displacement data of the rock mass to provide initial conditions for the calculation of the next time step;

[0024] Deformation and stress statistical index calculation: in each time step, calculate the gradient of the displacement field, and use the displacement gradient to calculate the deformation and stress statistical index.

[0025] Preferably, the data engine is based on real-time stress, deformation, displacement and address parameter data, and uses an LSTM model for prediction to generate deformation and stress statistical indicators in the future T time.

[0026] Preferably, the weight in the weighted fusion in step S3 is adjusted based on the following steps:

[0027] The fusion weight of the mechanism engine is greater than that of the data engine in stable regions, and the fusion weight of the data engine is greater than that of the mechanism engine in unstable regions;

[0028] Wherein the regional stability is evaluated based on the geological conditions and monitoring data of the rock mass; the nonlinearity degree is evaluated based on the deformation characteristics of the rock mass.

[0029] Preferably, the step S4 comprises the following steps:

[0030] Obtaining the current stress value of the tunnel surrounding rock and the compressive strength of the rock, dividing the current stress value by the compressive strength of the rock to obtain a strength stress ratio;

[0031] Based on the three principal stresses of the current stress tensor, the corresponding yield stress is calculated according to the Hoek-Brown criterion, and then the ratio of the square difference between the current stress tensor and the yield stress tensor is calculated to obtain the plastic yield proximity;

[0032] Based on the stress tensor and the strain tensor, the dot product of the gravity tensor and the strain tensor is calculated, and the dot product result is integrated in volume to obtain the energy release rate;

[0033] Based on the time series data of the deformation rate, the average value of the deformation rate is calculated, and then the square of the difference between the deformation rate at each time and the average value is calculated, and the sum is averaged, and then the square root is taken to obtain the standard deviation of the deformation rate;

[0034] Based on the strength stress ratio, the plastic yield proximity, the energy release rate and the standard deviation of the deformation rate, three levels of fuzzy sets are divided as input variables, and the respective membership functions are defined; according to the expert experience and time data, the fuzzy rules are established;

[0035] Through fuzzy logic reasoning, the fuzzy values of the input variables are comprehensively obtained to obtain the fuzzy output of the risk level, and the fuzzy output is converted into a clear risk level;

[0036] Based on the current predicted uncertainty and the maximum uncertainty value, the confidence is calculated, and based on this, the risk level and the confidence are spatially displayed in the form of a heat map.

[0037] The beneficial effects of the present application include:

[0038] The application embeds the rock mass mechanical relationship as a hard constraint into the neural network model, establishes an explicit mapping of monitoring data and physical fields, ensures that the model structure strictly follows the physical principle, and then adopts the ensemble Kalman filter to dynamically optimize the model parameter field in real time by using the field detection data, and designs a data-driven and physical mechanism cooperative deduction mechanism, which effectively suppresses the long-term prediction error accumulation, significantly improves the adaptability and prediction stability of the model under complex geological conditions; on this basis, the stress field, deformation field and other physical quantities deduced by the model are directly used to calculate the disaster risk index with clear mechanical significance, and the physical interpretable early warning decision is realized by fusing multi-source risk through fuzzy reasoning; finally, through the dynamic feedback loop of the actual disaster response data to the early warning result, the key parameters and early warning thresholds of the model are continuously calibrated, forming a self-iterative optimization early warning system, thereby systematically solving the technical problems of physical misalignment, weak long-term generalization and poor early warning interpretability of data-driven models. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 The overall step block diagram provided by the embodiment of the present application.

[0041] Figure 2 The step block diagram of step S2 provided by the embodiment of the present application.

[0042] Figure 3 The specific step block diagram of the double-engine deduction provided by the embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the technical problems, technical solutions and beneficial effects of the present application more clearly understood, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0044] Referring to Figure 1 As shown in the figure, the tunnel safety real-time dynamic modeling method of rock stress-deformation coupling includes the following steps:

[0045] S1. Collecting tunnel surrounding rock stress, deformation and environmental data, binding space-time tags through Beidou timing and tunnel mileage coordinates, and establishing a sensor-rock element mapping matrix, correlating the data to the discretized rock element grid, obtaining a multi-dimensional sensor data stream aligned in space and time;

[0046] In the embodiment, the stress data is mainly obtained by the fiber grating sensor and the piezoelectric micro stress sensor; the fiber grating sensor is used for continuous stress monitoring along the tunnel axis, and the piezoelectric sensor is used for point stress monitoring in the key area;

[0047] The deformation data is monitored by the distributed fiber strain sensor, the microseismic monitoring system is used for monitoring the point source position and energy, and the three-dimensional laser scanning system provides surface displacement data.

[0048] The environmental data includes underground water level meter, high-precision inclinometer, ground temperature sensor data, etc., and is used for monitoring the environmental changes inside and outside the tunnel.

[0049] For the above collected data, the collected data is preprocessed, specifically including the following steps:

[0050] Time reference synchronization: the built-in clock chip of all data collection sensors is synchronized by the Beidou second pulse signal at the hardware level, the original time stamp is corrected in two steps: first, the fixed transmission delay is compensated, and then the clock crystal oscillator jitter error is eliminated by the Kalman filter to obtain the sensor data after time stamp alignment; In the embodiment, the time synchronization by the Beidou pulse signal is only one embodiment of the present application, and GPS or the like can also be used; the specific time reference synchronization is a conventional technical means in the art, and therefore will not be described in detail in the embodiment;

[0051] Spatial coordinate conversion: taking the tunnel entrance control point as the origin and the tunnel design centerline as the reference axis to establish the tunnel mileage coordinate system, the sensor original coordinates are converted into the tunnel mileage coordinate system: the control point offset is eliminated by coordinate translation, and the tunnel azimuth deviation is corrected by the rotation matrix.

[0052] ;

[0053] In the formula: R represents the rotation matrix calculated from the tunnel design azimuth angle; C represents the coordinates of the tunnel entrance control point; S represents the coordinates of the sensor in the northeast coordinate system (engineering control network); TCCS represents the converted TCCS coordinates (tunnel mileage coordinate system), wherein M represents the mileage, D represents the lateral offset, H represents the elevation;

[0054] Rock mass grid construction: the tunnel surrounding rock space is discretized into tetrahedral element grid, each grid element is assigned a unique identifier, and each node records the spatial position; the purpose of this step is to map the point, line and surface data to the unified rock mass grid, which is as follows:

[0055] The surrounding rock of the tunnel is divided into a tetrahedral element grid, where the nodes are , the number of elements is , and the node coordinates are ;

[0056] For point data (such as inclinometers, water level gauges, etc.), inverse distance weighted mapping is used to the element nodes: ;

[0057] In the formula: , the physical quantity estimation value (such as stress, temperature) of node i; K represents the number of adjacent sensors; , the weight coefficient, is the weight of each sensor's measurement value in the estimated value; , the position of the kth adjacent sensor; , the measurement value of the kth adjacent sensor; , the attenuation coefficient; , the distance between node i and the kth sensor.

[0058] For linear data (such as fiber Bragg grating distributed stress sensors and distributed fiber strain sensors), integration is performed along the fiber path to map to the grid edge: ;

[0059] In the formula: , the strain value on the grid edge; , the strain measurement value at the fiber path s; L represents the length of the grid edge; , and represent the start and end points of the fiber and the grid edge, and the strain is integrated in this interval.

[0060] Planar data (three-dimensional point cloud): point cloud data is fitted to the grid surface by the least squares method, where is the coordinate of the jth point cloud data, is the displacement of the jth point cloud data, and the specific fitting expression is as follows: ; ; In the formula: , the position of the grid node; , the Gaussian weight function, where , the distance between the point cloud coordinate and the grid node ; , the fitted displacement field, is a displacement function fitted by the least squares method, i.e., a linear basis function; , the coefficient vector to be solved; , the coordinate of the point cloud data;

[0061] In this embodiment, the construction vibration interference is eliminated by adaptive band-pass filtering, and a specific expression is as follows: ;

[0062] The AR model filtering is used, wherein represents an original vibration signal; represents an autoregressive coefficient; represents an order of the model; represents a filtered signal; represents a filtered signal value at a time point ;

[0063] Based on this, a time-space aligned multi-dimensional data stream is generated, and each data frame contains a millimeter-level unified timestamp, a grid cell identifier, a stress, a strain, a displacement and other physical quantity values.

[0064] In this embodiment, by establishing a unified spatial coordinate system, the sensor data is converted to the coordinate system, the tunnel surrounding rock space is discretized through grid division, a unified spatial carrier is provided for data fusion, and according to the characteristics of different data types, a corresponding data fusion method is used to convert the original measurement value to a grid physical quantity. At the same time, through an adaptive filter, environmental noise is suppressed, and data quality is improved. The data format of the fused data is converted into a structured multi-dimensional data stream, so that the subsequent modules can use the data to reconstruct the physical field.

[0065] S2. Based on the time-space aligned multi-dimensional sensor data stream, a physical mechanism guided space-time convolutional neural network is constructed, the rock mass constitutive equation is embedded as a constraint term in network training, and the space-time coupled features are extracted. The real-time sensor data is converted to a predicted field by using ensemble Kalman filtering, the spatial distribution of rock mass parameters is dynamically corrected, and the dynamically updated stress tensor field and deformation tensor field of the whole tunnel are obtained.

[0066] Referring to Figure 2 In this embodiment, a physical mechanism guided space-time convolutional neural network is used to reconstruct the stress-deformation coupled field, and the network architecture includes an input layer, a space-time convolutional layer, a physical constraint embedding layer and an output layer.

[0067] The input is a time-space aligned multi-dimensional sensor data stream, including stress, strain, displacement, environmental parameters and the like, and the data is represented as a three-dimensional tensor with a shape of (H, W, C), wherein H and W represent the spatial resolution (the number of rows and columns of the sensor grid), and C represents the number of data channels (stress, strain, displacement, etc.) of different sensors.

[0068] First, the input data is standardized, and then the standardized data is input to the physical mechanism guided space-time convolutional neural network through the input layer.

[0069] The spatio-temporal convolution layer is used to extract spatial and temporal features, wherein the spatial feature extraction uses a 3x3 convolution kernel to slide in the spatial dimension to extract local spatial features, and the temporal feature extraction uses a 1x1 convolution kernel to expand in the time dimension to extract dynamic features in the time sequence;

[0070] Then the spatial and temporal features are fused to generate high-level feature representation, and a high-level feature tensor is output based on this, and the expression is as follows: ;

[0071] In the formula: represents the local spatial features extracted by the spatio-temporal convolution; represents the dynamic features in the time sequence extracted; is a nonlinear activation function; represents the fused features;

[0072] The output layer adopts a linear transformation layer to output the stress field and the deformation field based on the fused features, and the specific expression is as follows: ;

[0073] In the formula: represents the predicted stress field, represents the predicted deformation field, wherein represents the spatial coordinates, and t represents the time; represents the linear transformation layer, for example, a fully connected layer.

[0074] The physical constraint embedding layer is introduced into the loss function to introduce a physical constraint term, to ensure that the predicted maximum principal stress and the equivalent strain satisfy the Hoek-Brown criterion, so that the loss function includes a data fitting loss and a physical constraint loss, wherein the data fitting loss measures the difference between the predicted field and the real field by using the mean square error; the physical loss constraint calculates the residual between the predicted maximum principal stress and the equivalent strain based on the Hoek-Brown criterion; and then by adjusting the weights of the data fitting loss and the physical constraint loss, the balance between the two is ensured in the training process, and the specific expression of the loss function is as follows: ;

[0075] In the formula: represents the data fitting loss, which uses the mean square error; represents the physical constraint loss; and represent weight coefficients for balancing the data fitting and the physical constraint; wherein the expression of the physical constraint loss is as follows: ;

[0076] In the formula: represents the predicted maximum principal stress; represents the equivalent strain; denotes the number of constraints; and denotes the material parameter; m denotes the index parameter.

[0077] Further, in this embodiment, the ensemble Kalman filter is used for data assimilation, and the parameters of the physical mechanism guided spatio-temporal convolutional neural network (PG-STCNN model) are dynamically corrected. The specific steps are as follows:

[0078] The initial values and covariance matrices of the rock mass mechanical parameters such as elastic modulus and Poisson's ratio are initialized to reflect the uncertainty of the parameters: ;

[0079] wherein, denotes the elastic modulus; denotes the Poisson's ratio; denotes the parameter of Hoek-Brown criterion;

[0080] The PG-STCNN is used to generate a prediction field as the predicted state of the model: ;

[0081] wherein, denotes the model equation (prediction process of the PG-STCNN model); denotes the process noise; denotes the update of the state at the previous time calculated by the PG-STCNN model; denotes the state vector at time k;

[0082] The actual observation data (such as stress and strain) are compared with the prediction field to calculate the observation residual: ;

[0083] wherein, denotes the observation model, which is used to map the state vector to the observation space; denotes the observation noise; denotes the observation vector at time k;

[0084] The observation model converts the state vector into the predicted observation value by combining the physical equations of rock mass mechanics, and the specific process is as follows:

[0085] wherein the physical equation is the relationship between the maximum principal stress and the equivalent strain according to the Hoek-Brown criterion:

[0086] ;

[0087] wherein, Maximum principal stress, predicted by Hoek-Brown criterion Equivalent strain

[0088] The relationship between stress and strain is calculated using the generalized Hooke's law, taking into account the elastic properties of the material:

[0089]

[0090] Where: The stress is determined by the elastic modulus and the strain ;

[0091] Based on the above, the observation model is constructed , which is expressed as follows: ;

[0092] Based on the state covariance matrix and the observation covariance matrix, the Kalman gain is calculated, which is used to adjust the state vector: ;

[0093] In the formula: The observation matrix is denoted by The observation noise covariance matrix is denoted by The state covariance matrix is denoted by The Kalman gain is denoted by T represents the transpose

[0094] State update: using the Kalman gain and the observation residual, the state vector is updated to obtain more accurate model parameters: ;

[0095] In the formula: The predicted state vector at time k is denoted by The updated state vector at time k is denoted by The predicted observation value at time k is denoted by

[0096] The updated model parameters are used for subsequent prediction field generation, based on which high-confidence full-space stress tensor field and deformation tensor field are generated; then through inverse normalization processing, the inverse normalized predicted value is obtained.

[0097] ​In this embodiment, by deeply coupling physical mechanism and data driving, Hoek-Brown criterion is embedded into loss function, and network output is forced to comply with rock mass constitutive relation, and dynamic updating stress and deformation tensor field is obtained by combining ensemble Kalman filter to dynamically correct rock mass parameters; the introduction of physical mechanism makes the model more consistent with the actual mechanical behavior of rock mass, avoiding the situation that pure data model driving is contrary to physical law; secondly, by combining Kalman filter to effectively utilize real-time sensor data to correct prediction, the parameter space distribution is more consistent with the actual situation, thereby improving the accuracy and dynamics of stress and deformation tensor field.

[0098] S3. Based on the stress tensor field and the deformation tensor field, a double-engine deduction based on the combination of mechanism engine and data engine is adopted, the double-engine deduction results are weighted and fused according to confidence, the deduction time window is adaptively adjusted, and future stress and deformation evolution field and uncertainty quantization results are output;

[0099] Referring to Figure 3 The mechanism engine is based on explicit finite difference method, and the updated rock mass parameter field is used to calculate the stress-deformation evolution in future time, including the following steps:

[0100] According to the set time step, the calculation is gradually advanced, and in each time step, the stress and displacement change of the next time step is calculated by using the current gravity and displacement data; at the end of each time step, the stress and displacement data of the rock mass are updated to provide initial conditions for the calculation of the next time step;

[0101] In each time step, the gradient of the displacement field is calculated, and the displacement gradient is used to calculate the deformation and stress statistical indicators, and the expression is as follows:

[0102] ;

[0103] ;

[0104] In the formula: ρ represents the density of the rock mass; F represents the external force; g represents the acceleration of gravity; σ represents the stress tensor the partial derivative of time t; ε represents the divergence of the deformation tensor ; σ represents the divergence of the stress tensor; ε represents the partial derivative of time t of the deformation tensor ;

[0105] Based on the explicit finite difference method, the stress and deformation evolution in future time step is calculated.

[0106] The data engine generates future deformation and stress statistical indicators within T time based on real-time stress, deformation, displacement and address parameter data, and uses an LSTM model for prediction. In this embodiment, the LSTM model includes an input layer, multiple LSTM layers and an output layer, and outputs future deformation and stress statistical indicators through the output layer.

[0107] Based on the deformation and stress statistical indicators obtained by the data engine and the deformation and stress statistical indicators obtained by the mechanism engine, a weighted fusion method is used for fusion, wherein the weights of the weighted fusion are dynamically adjusted based on the regional stability and the nonlinearity degree, and the specific method is as follows:

[0108]

[0109] In the formula, S represents a normalized stability index; represents a normalized nonlinearity degree index; represents a mechanism engine base weight; represents a nonlinearity sensitivity coefficient; represents a mechanism engine weight; represents a data engine weight;

[0110] For example, the regional stability index S is calculated based on the intensity stress ratio of each direction:

[0111] In the formula, represents the intensity stress ratio of direction i; represents the intensity limit of direction i; represents the actual stress of direction i; respectively correspond to three principal stress directions;

[0112] The calculated intensity stress ratio is normalized by using a Sigmoid function:

[0113] In the formula, represents a critical threshold value; represents a sensitivity coefficient; e represents the base number of natural logarithm;

[0114] The comprehensive stability index is calculated based on the normalized value:

[0115]

[0116] In the formula, respectively represent the normalized stability index of x, y and z directions.

[0117] The nonlinearity degree index is calculated based on the following expression: ;​​​​​

[0118] wherein: represents the equivalent plastic strain; represents the reference strain of the material; represents the material constant.

[0119] When the prediction uncertainty increases, the step length is automatically shortened and the data communication frequency is increased, exemplarily:

[0120] Define the prediction difference index , which represents the deviation degree of the double-engine prediction results: ;

[0121] wherein: represents the stress tensor field predicted by the mechanism engine; represents the stress tensor field predicted by the data engine; represents the Frobenius norm of the tensor; represents the reference stress value;

[0122] Calculate the adaptive step length based on the difference index: ;

[0123] ;

[0124] ;

[0125] wherein: represents the basic step length; represents the sensitivity coefficient; represents the stability factor; adjusted step length; represents the minimum size of the finite difference grid; represents the maximum wave speed in the rock mass; represents the safety factor; represents the elastic model quantity, represents the density; represents the Poisson's ratio;

[0126] The step length boundary constraint is as follows: ;

[0127] wherein: represents the minimum allowed step length;

[0128] Based on the boundary constraint, it is ensured that the calculation step length is neither lower than the minimum value allowed by the hardware, nor exceeds the maximum value limited by the CFL condition, so as to find a balance between the calculation efficiency and the stability.

[0129] In the embodiment, the mechanism engine ensures the scientificity of the deduction based on physical laws, and the data engine improves the accuracy of short-term prediction with the aid of data trends, and the combination of the two meets the deduction requirements in different scenarios; the reliability of the future stress, deformation evolution field and uncertainty quantification results is further improved by confidence weighted fusion and adaptive adjustment of time window.

[0130] S4. Based on the future gravity and deformation evolution field, the triple risk indicators of each rock mass unit are calculated in real time, the multi-index fusion is carried out based on the fuzzy reasoning system, the risk level is dynamically divided combined with the change rate, the risk confidence is generated by associating the uncertainty quantification results, the real-time risk thermodynamic map of the whole space of the tunnel and the positioning warning information are obtained;

[0131] As a possible implementation manner of the embodiment, the step S4 comprises the following steps:

[0132] The current stress value of the tunnel surrounding rock and the compressive strength of the rock are obtained, and in this step, the strength stress ratio is obtained by dividing the current stress value by the compressive strength of the rock;

[0133] ;

[0134] In the formula: The strength stress ratio is represented; The current stress value is represented; The compressive strength of the rock is represented;

[0135] Based on the three principal stresses of the current stress tensor, the corresponding yield stress is calculated according to the Hoek-Brown criterion, and then the ratio is calculated after calculating the square difference of the current stress tensor and the yield stress tensor, to obtain the plastic yield proximity;

[0136] ;

[0137] In the formula: The three principal stresses of the current stress tensor are represented; The yield stress calculated according to the Hoek-Brown criterion is represented; The plastic yield proximity is represented;

[0138] Based on the stress tensor and the strain tensor, the dot product of the gravity tensor and the strain tensor is calculated, and the dot product result is integrated in volume to obtain the energy release rate;

[0139] ;

[0140] In the formula: The stress tensor is represented; The strain tensor is represented; V represents the volume; The energy release rate is represented;

[0141] Based on the time series data of strain rate, the average value of strain rate is calculated, and the standard deviation of strain rate is obtained based on the average value;

[0142] ;

[0143] In the formula: represents the strain rate at the i th time; represents the average value of the strain rate; N represents the length of the time series; represents the standard deviation of the strain rate;

[0144] Based on the intensity stress ratio, the plastic yield proximity, the energy release rate and the strain rate standard deviation as input variables, three fuzzy sets of grades are divided, and the respective membership functions are defined; according to expert experience and time data, fuzzy rules are established, for example:

[0145] The FIS input includes SSR, PYI, ERR and its strain rate standard deviation, and the output is the risk level (safe, attention, alert, danger); the fuzzy sets and their membership functions are as follows:

[0146] SSR: low (0-0.5), medium (0.3-0.7), high (0.5-1).

[0147] PYI: low (0-0.3), medium (0.2-0.6), high (0.4-1).

[0148] ERR: low (0-10), medium (5-15), high (10-20).

[0149] For example, the rule base is as follows:

[0150] If SSR is high and PYI is high, the risk level is danger.

[0151] If SSR is medium and ERR is high, the risk level is alert.

[0152] According to the uncertainty index in step 3, the confidence score is calculated : ;

[0153] In the formula: represents the predicted uncertainty index; represents the maximum uncertainty value;

[0154] Based on this, a real-time risk heat map is obtained, which shows the risk level and confidence of each rock mass unit; and an advanced warning signal is output, such as "20 meters in front of the working face, the rock burst risk will rise to the alert level within 1 hour, the execution degree is 86%".

[0155] S5. Real-time acquisition of actual response data, when the actual response data deviates from the early warning result, triggering closed-loop optimization.

[0156] In the embodiment, the absolute deviation and relative error are calculated by comparing the predicted stress / strain and the measured value at the same space-time position, and the significant error area is distinguished.

[0157] Then the rock mass parameters to be corrected are extracted: elastic modulus, Poisson's ratio, cohesion, internal friction angle; the measured stress / strain is taken as "observation data" and input into the EnKF algorithm.

[0158] The parameter posteriori estimation is generated: the priori parameter estimation and observation data are fused by weighting to reduce the parameter uncertainty in the high error area.

[0159] Then the updated parameter field is fed back to step S2 in real time, and the physical field reconstruction is re-executed using the new parameters to ensure that the subsequent data processing is more accurate.

[0160] In addition, it also includes the optimization of PG-STCNN and LSTM model parameters, and how to optimize the model parameters based on the measured value and the predicted value is a conventional technical means in the art, so it is not described in detail in the present application.

[0161] The present application embeds the rock mass mechanical relationship as a hard constraint into the neural network model, establishes an explicit mapping between the monitoring data and the physical field, ensures that the model structure strictly follows the physical principle, and then uses the ensemble Kalman filter to dynamically optimize the model parameter field in real time. Field detection data, and design a data-driven and physical mechanism co-evolution mechanism, effectively suppresses the long-term prediction error accumulation, significantly improves the adaptability and prediction stability of the model under complex geological conditions; on this basis, directly use the stress field, deformation field and other physical quantities calculated by the model evolution output to calculate the disaster risk index with clear mechanical significance, and realize the physically interpretable early warning decision through the fusion of multi-source risk through fuzzy reasoning; finally, through the dynamic feedback of actual disaster response data to the early warning result, the key parameters and early warning thresholds of the model are continuously calibrated, forming a self-iterative optimization early warning system, thereby systematically solving the technical problems of physical misalignment, weak long-term generalization and poor early warning interpretability of data-driven models.

[0162] The above is only a preferred embodiment of the present application and does not limit the present application, any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A tunnel safety real-time dynamic modeling method of coupling of rock mass stress-deformation, characterized in that, The method comprises the following steps: S1. Collecting tunnel surrounding rock stress, deformation and environmental data, binding space-time tags through Beidou timing and tunnel mileage coordinates, and establishing a sensor-rock mass unit mapping matrix to correlate the data to the discretized rock mass grid unit to obtain a multi-dimensional sensor data stream aligned in time and space; S2. Based on the multi-dimensional sensor data stream aligned in time and space, a physical mechanism guided space-time convolutional neural network is constructed, the rock mass constitutive equation is embedded as a constraint term in network training, the space-time coupled features are extracted, the real-time sensor data is converted to the prediction field by using the set Kalman filter, the rock mass parameter space distribution is dynamically corrected, and the stress tensor field and the deformation tensor field of the whole tunnel are dynamically updated; S3. Based on the stress tensor field and the deformation tensor field, a double-engine deduction based on the combination of a mechanism engine and a data engine is used, the double-engine deduction results are weighted and fused according to the confidence, the deduction time window is adaptively adjusted, the future stress and deformation evolution field and the uncertainty quantification result are output; S4. Based on the future gravity and deformation evolution field, the triple risk indicators of each rock mass unit are calculated in real time, the multi-index fusion is carried out based on the fuzzy reasoning system, the risk level is dynamically divided combined with the change rate, the risk confidence is generated by correlating the uncertainty quantification result, and the real-time risk thermodynamic map of the whole tunnel space and the positioning early warning information are obtained; S5. Real-time acquisition of actual response data, when the actual response data deviates from the warning result, a closed-loop optimization is triggered.

2. The method according to claim 1, wherein, The binding of space-time tags comprises the following steps: Time reference synchronization: all data acquisition sensors have built-in clock chips, which are synchronized in hardware level through Beidou second pulse signals, and the original time stamp is corrected in two steps: first, compensate for the fixed transmission delay, and then eliminate the clock crystal oscillator jitter error through the Kalman filter to obtain the sensor data after time stamp alignment; Space coordinate conversion: a tunnel mileage coordinate system is established with the tunnel entrance control point as the origin and the tunnel design centerline as the reference axis; Convert the original coordinates of the sensor to the tunnel mileage coordinate system: eliminate the control point offset by coordinate translation, and correct the tunnel azimuth deviation by rotation matrix; Rock mass grid construction: discretize the tunnel surrounding rock space into tetrahedral unit grids, assign a unique identifier to each grid unit, and record the spatial position of each node.

3. The method of claim 1, wherein, The data processing steps of the physical mechanism guided space-time convolutional neural network are as follows: Spatial feature extraction: use a predetermined size of convolution kernel to slide in the spatial dimension to extract local spatial features; Time feature extraction: use a predetermined size of convolution kernel to expand in the time dimension to extract dynamic features in the time series; Feature fusion: fuse the spatial and temporal features to generate high-level feature representation, and output high-level feature tensors based on this.

4. The method of claim 1, wherein, The physical mechanism is introduced into the loss function of the space-time convolutional neural network: The loss function includes data fitting loss and physical constraint loss; wherein the data fitting loss measures the difference between the predicted field and the true field by using mean square error; the physical loss constraint is based on the Hoek-Brown criterion to calculate the residual error between the predicted maximum principal stress and the equivalent strain; Then, by adjusting the weights of the data fitting loss and the physical constraint loss, the balance between the two in the training process is ensured.

5. The method of claim 1, wherein, The mechanism engine is based on the explicit finite difference method, and uses the updated rock mass parameter field to calculate the stress-deformation evolution in the future time, including the following steps: Time step advancing: according to the set time step, the calculation is gradually advanced, and in each time step, the stress and displacement changes of the next time step are calculated by using the current gravity and displacement data; at the end of each time step, the stress and displacement data of the rock mass are updated to provide initial conditions for the calculation of the next time step; Deformation and stress statistical index calculation: in each time step, the gradient of the displacement field is calculated, and the deformation and stress statistical index are calculated by using the displacement gradient.

6. The method of claim 1, wherein, The data engine is based on real-time stress, deformation, displacement and address parameter data, and uses an LSTM model for prediction to generate deformation and stress statistical indexes in the future T time.

7. The method of claim 1, wherein, The weights in the step S3 are adjusted based on the following steps: The fusion weights of the mechanism engine are greater than those of the data engine in stable regions, and the fusion weights of the data engine are greater than those of the mechanism engine in unstable regions; Wherein the regional stability is evaluated based on the geological conditions and monitoring data of the rock mass; The nonlinearity degree is evaluated based on the deformation characteristics of the rock mass.

8. The method of claim 1, wherein, The step S4 includes the following steps: Obtain the current stress value of the tunnel surrounding rock and the compressive strength of the rock, divide the current stress value by the compressive strength of the rock to obtain the strength stress ratio; Based on the three principal stresses of the current stress tensor, the corresponding yield stress is calculated according to the Hoek-Brown criterion, and then the square difference of the current stress tensor and the yield stress tensor is calculated to obtain the plastic yield proximity; Based on the stress tensor and the strain tensor, the dot product of the gravity tensor and the strain tensor is calculated, and the dot product result is integrated in volume to obtain the energy release rate; Based on the time series data of the deformation rate, the average value of the deformation rate is calculated, and then the square of the difference between the deformation rate at each time and the average value is calculated, and the sum is averaged and then squared to obtain the standard deviation of the deformation rate; Based on the strength stress ratio, plastic yield proximity, energy release rate and deformation rate standard deviation, three levels of fuzzy sets are divided as input variables, and the membership functions of each level are defined; according to expert experience and time data, fuzzy rules are established; Through fuzzy logic reasoning, the fuzzy values of the input variables are integrated to obtain the fuzzy output of the risk level, and the fuzzy output is converted into clear risk level; Based on the current prediction uncertainty and the maximum uncertainty value, the confidence is calculated, and based on this, the risk level and confidence are spatially displayed in the form of a heat map.

Images