Sandy stratum tunnel water gushing risk prediction method

Abstract

The invention provides a sandy stratum tunnel water burst risk prediction method. The method comprises the steps that a radar instantaneous energy spectrum in front of tunneling is obtained through a geological radar, a micro-seismic spectrum is obtained through a micro-seismic array, and a residual distance is obtained through laser ranging; aligning and coupling the two spectrums to obtain a water inrush risk index, inputting the water inrush risk index and the residual distance into a self-adaptive threshold comparator, and if the water inrush risk index and the residual distance exceed the limit, dynamically adjusting the rotating speed of a cutter head, propulsive force, the rotating speed of a screw conveyor and grouting pressure, and generating a risk avoidance path; and calling a water inrush risk model library constructed by LSTM-Attention, inputting a risk index sequence, a residual distance, cutterhead torque, a soil pressure gradient, a current change rate and other auxiliary variables, and outputting a future multi-step risk level. According to the method, accurate prediction and real-time avoidance of the water gushing risk of the sandy stratum tunnel can be realized, and the safety and efficiency of tunnel construction are improved.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction safety monitoring and intelligent control technology, and more specifically, to a method for predicting the risk of sudden water inrush in tunnels in sandy strata. Background Technology

[0002] In tunnel construction, sudden water inrush in sandy strata is a common geological hazard, posing a serious threat to construction safety and project progress. Existing methods for predicting sudden water inrush risks mainly rely on geological exploration data, empirical formulas, and simple analysis of field monitoring data. While these methods can provide risk warnings to some extent, they have many limitations. For example, geological exploration data has limited spatiotemporal resolution, making it difficult to accurately reflect the dynamic changes in sandy strata tunnels; empirical formulas are often based on specific engineering experience and lack universality; and the analysis of field monitoring data is usually lagging, making it impossible to predict sudden water inrush risks in real time and accurately.

[0003] With the development of technology, some new technical means have been introduced into the prediction of sudden water inrush risks, such as ground-penetrating radar (GPR) and microseismic monitoring. GPR can detect the electromagnetic wave reflection characteristics of the geological body ahead, while microseismic monitoring can capture microseismic waveforms induced during tunneling. However, these technologies are mostly used independently, lacking fusion analysis of multi-source data, resulting in insufficient accuracy and reliability of the prediction results. In addition, existing methods are not timely and accurate enough in dynamically adjusting construction parameters after risk prediction, making it difficult to effectively avoid risks.

[0004] In implementing the embodiments of the present invention, the prior art has at least the following problems or defects: On the one hand, the existing prediction methods mostly rely on a single data source and cannot make full use of multi-source heterogeneous data for comprehensive analysis, resulting in low prediction accuracy; on the other hand, the existing technology is not intelligent and dynamic enough in adjusting construction parameters after risk prediction, and cannot quickly generate the optimal construction instructions according to the real-time risk level, thus failing to effectively reduce the risk of sudden water inrush. Summary of the Invention

[0005] This invention provides a method for predicting the risk of sudden water inrush in tunnels in sandy strata, comprising: The instantaneous energy spectrum of the radar is obtained by transmitting step-frequency electromagnetic waves forward through ground-penetrating radar and receiving the echoes, and then performing Hilbert-Huang transform. The micro-vibration waveforms induced by tunneling vibration are captured by a micro-seismic sensor array, and the instantaneous energy spectrum of the micro-seismic waves is obtained after wavelet packet decomposition. The remaining distance from the cutterhead end face to the sandy interface is measured in real time using a laser rangefinder. The instantaneous energy spectrum of the radar and the instantaneous energy spectrum of the micro-seismic waves are aligned within the same time window and energy coupling calculation is performed to obtain the risk index of sudden water inrush at the sandy interface. The risk index of sudden water inrush and the remaining distance are input into an adaptive threshold comparator. When the risk index is higher than the threshold or the remaining distance is lower than the warning line, the excavation parameters are immediately dynamically adjusted. The adjustment objects include the cutterhead speed, propulsion force, screw conveyor speed and synchronous grouting pressure. After the adjustment is completed, the risk avoidance path of the tunneling machine is generated based on the adjusted parameters. The target model corresponding to the current sandy stratum tunnel type, burial depth and geological age is retrieved from the sudden water inrush risk model library. The target model is implemented with LSTM-Attention architecture. The input is the sudden water inrush risk index sequence, the remaining distance sequence and the auxiliary variable sequence. The output is the sudden water inrush risk level for multiple future time periods. The auxiliary variable sequence includes the cutterhead torque fluctuation index, the soil pressure gradient, and the screw conveyor current change rate. The sudden water inrush risk model library is constructed as follows: A three-field coupled simulation environment of sandy strata tunnel-tunneling machine-groundwater is established in GTSNX. Training sample sets are generated according to different sandy strata tunnel types, different burial depths, and different geological ages. The sample sets include radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, residual distance, auxiliary variables, and corresponding real labels of sudden water inrush. Bayesian optimization is used to optimize the hyperparameters of the LSTM-Attention architecture. The optimization objective is to minimize the cross-entropy loss between the prediction level and the real label. After optimization, the optimal hyperparameters are solidified into a sudden water inrush risk model for the combination of strata-burial depth-age and are added to the library.

[0006] Furthermore, using the tunneling cycle as a time window, the instantaneous energy spectrum of the radar and the instantaneous energy spectrum of the microseismic event are discretized according to time periods; The dynamic weighted allocation values ​​of radar instantaneous energy and microseismic instantaneous energy within the same time period are calculated to obtain the coupled energy value; wherein, the dynamic weighted allocation is adjusted according to the stratigraphic heterogeneity characteristics, specifically as follows: Feature extraction was performed on the instantaneous energy spectrum of radar and the instantaneous energy spectrum of microseismic activity. The extracted features included the peak value and frequency distribution of the energy spectrum. The weighting coefficients of radar instantaneous energy and microseismic instantaneous energy in the current time period are calculated based on the extracted features. The weighting coefficients are related to the formation heterogeneity. The higher the formation heterogeneity, the greater the difference in the weighting coefficients. The instantaneous radar energy and the instantaneous microseismic energy are weighted and summed according to the weighting coefficients to obtain the coupled energy value; The risk index of sudden water inrush at the sandy interface is obtained by summing all the coupled energy values ​​within the time window and dividing by the time window length.

[0007] Furthermore, after inputting the sudden water inrush risk index and the remaining distance into the adaptive threshold comparator, the dynamic adjustment of the threshold by the adaptive threshold comparator includes: The current tunneling speed value is read in real time in the tunneling machine PLC, and the speed change is obtained by subtracting the tunneling speed value from the previous cycle. If the change in speed is positive and exceeds the set rate, the threshold increases linearly. If the change in speed is negative and its absolute value is higher than the set rate, the threshold decreases exponentially. If the absolute value of the speed change does not exceed the set rate, the threshold remains unchanged; The adjusted threshold is immediately written to the adaptive threshold comparator and takes effect.

[0008] Furthermore, the risk avoidance path for the tunneling machine generated based on the adjusted parameters includes: The tunnel axis is discretized into a set of nodes at fixed intervals; For each node, call the inrush water risk model to calculate the maximum inrush water risk level of that node over a certain number of future time periods. If the maximum risk level of sudden water inrush exceeds the preset level, the node will be marked as a risk node. Construct a cost function, which is a weighted sum of the node risk level and the Euclidean distance from the node to the current cutterhead position; The A* algorithm is used to search for the minimum cost path from the current cutterhead position to the end of the tunnel design axis in the node set. The cost of risk nodes is multiplied by a penalty coefficient, which increases monotonically with the risk level. The minimum cost paths are connected in sequence to generate a continuous risk avoidance path, which is then sent to the tunneling machine navigation system for execution.

[0009] Furthermore, the sudden water inrush risk model library is constructed in the following ways, including: In the GTSNX platform, for each type of sandy stratum tunnel—burial depth—geological age ternary set, a three-dimensional simulation scene is established that fully couples the physical model of the sandy stratum tunnel, the physical model of the tunnel boring machine, and the groundwater seepage field. The set of working conditions covering the parameter space is generated by Latin hypercube sampling. The input of each working condition is the radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, remaining range and auxiliary variables, and the output is the true label of the sudden water inrush. The simulation was run to extract the time of the sudden water inrush, the inrush volume, and the peak pore water pressure to form training samples. Bayesian optimization is used to globally optimize the hyperparameters of the LSTM-Attention architecture, with cross-entropy loss as the optimization objective; The optimal hyperparameters are solidified into the ternary water inrush risk model and stored in the water inrush risk model library with the stratigraphic type-burial depth-geological age as the key. At the same time, the model weights and normalized parameters are saved.

[0010] Furthermore, the acquisition of the auxiliary variable sequence, including the cutterhead torque fluctuation index, includes: A torque sensor is installed at the spindle of the cutter head drive motor to continuously collect the raw torque signal of the cutter head at a fixed frequency; The original signal is subjected to moving average filtering to obtain the filtered torque signal; The filtered torque signal is segmented into segments of fixed duration and the variance of each segment is calculated to obtain a short-time variance sequence. The short-time variance sequence is subjected to minimum-maximum normalization to obtain the tool head torque fluctuation index; When the cutterhead torque fluctuation index exceeds the set limit, an early warning signal is triggered and the tunneling cycle number is recorded.

[0011] Furthermore, the establishment of a three-field coupled simulation environment for sandy strata tunnel-tunneling machine-groundwater in GTSNX includes: A three-dimensional geological model of a tunnel in sandy strata was established in GTSNX and nonlinear elastoplastic constitutive and permeability parameters were assigned. Establish a model of the tunneling machine casing and set the boundary conditions for the cutterhead, screw conveyor, and grouting holes; Establish a groundwater seepage field and set the hydraulic head boundary and permeability coefficient; The three-field model was meshed and a coupled solver was set up. The coupling variables included displacement, pore water pressure and seepage velocity.

[0012] Furthermore, retrieving target models corresponding to the current type, depth, and geological age of tunnels in sandy strata from the sudden water inrush risk model database also includes: if no perfectly matching model exists, performing similarity retrieval and transfer learning. Calculate the cosine similarity between the current stratigraphic feature vector and the corresponding stratigraphic feature vectors of all models in the library; The model with the highest similarity is selected as the source model, its attention layer weights are frozen, and only the LSTM layer weights are fine-tuned. The LSTM layer weights are updated online using real-time data collected from the current tunneling section. After the update is completed, the model is fixed and stored in the database.

[0013] Furthermore, updating the LSTM layer weights online using real-time data collected from the current tunneling section includes: Construct a sliding time window of fixed length, with a step size of one tunneling cycle; Within the time window, outlier removal and missing value filling are performed on the real-time acquired radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, remaining distance, and auxiliary variables; The preprocessed data is input into the source model of the frozen attention layer to obtain the predicted risk level of sudden water inrush, and the cross-entropy loss is calculated by comparing it with the measured level on site. An adaptive optimizer is used to update the gradient of the LSTM layer weights, and the learning rate adaptively decreases with the sample variance within the time window. Add a resilient weighting term to the loss function to prevent catastrophic forgetting; When the cross-entropy loss of several consecutive time windows is lower than the set threshold and the decrease in loss is lower than the set proportion, the update stops, the current weights are solidified to form a new inrush water risk model, and the original reservoir record is overwritten with the current stratum type-burial depth-geological age as the key.

[0014] Furthermore, after outputting the risk levels of sudden water inrush for multiple future time periods, the system also includes generating construction instructions, including: Map the risk level of sudden water inrush to a risk color code; Risk color codes are overlaid on the tunnel BIM model to form a visual risk cloud map; Based on the risk cloud map, the shield machine control commands are automatically generated. The commands include quantitative values ​​of cutterhead speed, thrust, synchronous grouting pressure and screw conveyor speed. Control commands are sent to the tunnel boring machine's PLC via industrial Ethernet for execution.

[0015] The embodiments of the present invention have at least the following beneficial effects: 1. By fusing multi-source data from ground-penetrating radar, microseismic sensor arrays, and laser rangefinders, comprehensive real-time monitoring and accurate assessment of the risk of sudden water inrush in tunnels in sandy strata were achieved. This solved the problem of insufficient prediction accuracy caused by the reliance on a single data source in traditional methods, and improved the reliability of risk prediction.

[0016] 2. By adopting an adaptive threshold comparator and a strategy of dynamically adjusting tunneling parameters, the tunneling parameters of the tunnel boring machine can be quickly adjusted based on the real-time monitored risk index and remaining distance, and a risk avoidance path can be generated. This solves the problem of untimely and inaccurate adjustment of construction parameters in existing technologies, reduces the risk of sudden water inrush, and ensures construction safety.

[0017] 3. The sudden water inrush risk model library built on the LSTM-Attention architecture, combined with multi-step risk level prediction and transfer learning technology, can generate accurate risk prediction models for different types of sandy strata tunnels, burial depths and geological ages, and can quickly adapt and update the models under new geological conditions. This solves the problems of insufficient prediction ability and poor model adaptability of existing technologies under complex geological conditions, and improves the universality and practicality of the prediction system. Attached Figure Description

[0018] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein: Figure 1 This is a flowchart illustrating a method for predicting the risk of sudden water inrush in tunnels in sandy strata according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a tunnel intrusion water risk prediction system in sandy strata provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] like Figure 1 As shown, this application proposes a method for predicting the risk of sudden water inrush in tunnels in sandy strata, including: S1. A step-frequency electromagnetic wave is emitted towards the tunneling front by a ground-penetrating radar and the echo is received. The instantaneous energy spectrum of the radar is obtained after Hilbert-Huang transform. The micro-vibration waveform induced by tunneling vibration is captured by a micro-seismic sensor array and the instantaneous energy spectrum of the micro-seismic wave is obtained after wavelet packet decomposition. The remaining distance from the cutterhead end face to the sandy interface is measured in real time by a laser rangefinder. The instantaneous energy spectrum of the radar and the instantaneous energy spectrum of the micro-seismic wave are aligned within the same time window and energy coupling calculation is performed to obtain the risk index of sudden water inrush at the sandy interface. S2. Input the sudden water inrush risk index and the remaining distance into the adaptive threshold comparator. When the risk index is higher than the threshold or the remaining distance is lower than the warning line, the tunneling parameters are immediately dynamically adjusted. The adjustment objects include the cutterhead speed, propulsion force, screw conveyor speed and synchronous grouting pressure. After the adjustment is completed, the risk avoidance path of the tunneling machine is generated based on the adjusted parameters. S3. Retrieve the target model corresponding to the current sandy stratum tunnel type, burial depth and geological age from the sudden water inrush risk model library. The target model is implemented with LSTM-Attention architecture. The input is the sudden water inrush risk index sequence, the remaining distance sequence and the auxiliary variable sequence. The output is the sudden water inrush risk level for multiple future time periods. The auxiliary variable sequence includes the cutterhead torque fluctuation index, the soil pressure gradient, and the screw conveyor current change rate.

[0021] This application achieves accurate prediction of water inrush risk in tunnels in sandy strata and real-time closed-loop control of construction parameters by simultaneously acquiring electromagnetic reflection, mechanical vibration, and spatial distance data, combined with depth time-series models and physical simulation technology. This system overcomes the limitations of traditional single-data source analysis, utilizing multi-physics coupled simulation to generate training data, significantly improving the generalization ability of the risk model. Simultaneously, through an adaptive threshold mechanism and online parameter adjustment, it ensures the timeliness of risk response and the accuracy of control.

[0022] Specifically, a ground-penetrating radar emits stepped-frequency electromagnetic waves and receives the echoes, obtaining the radar's instantaneous energy spectrum through Hilbert-Huang transform. Simultaneously, a microseismic sensor array captures microseismic waveforms induced by tunneling vibrations, obtaining the microseismic instantaneous energy spectrum through wavelet packet decomposition. A laser rangefinder measures the remaining distance from the cutterhead end face to the sandy interface in real time. The radar's instantaneous energy spectrum and the microseismic instantaneous energy spectrum are aligned within the same time window and energy coupling calculations are performed to obtain the risk index of sudden water inrush at the sandy interface.

[0023] The risk index of sudden water inrush and the remaining distance are input into an adaptive threshold comparator. When the risk index is higher than the threshold or the remaining distance is lower than the warning line, the tunneling parameters are dynamically adjusted, including the cutterhead speed, thrust, screw conveyor speed, and synchronous grouting pressure. After the adjustment is completed, a risk avoidance path for the tunneling machine is generated based on the adjusted parameters.

[0024] The target model corresponding to the current tunnel type, burial depth, and geological age in sandy strata is retrieved from the sudden water inrush risk model library. The target model adopts an LSTM-Attention architecture, with the input being a sudden water inrush risk index sequence, a residual distance sequence, and an auxiliary variable sequence, and the output being the sudden water inrush risk level for multiple future time periods. The auxiliary variable sequence includes the cutterhead torque fluctuation index, the soil pressure gradient, and the screw conveyor current change rate.

[0025] The LSTM-Attention model operates in a parallel coupled manner: firstly, convolutional spatial features are extracted from the instantaneous energy spectra of radar and microseismic events; secondly, LSTM temporal features are extracted from the inrush water risk index sequence, the remaining distance sequence, and the auxiliary variable sequences (cutterhead torque fluctuation index, soil pressure gradient, and screw conveyor current change rate). The two types of features are concatenated after dimensionality unification and input into the LSTM-Attention network under a Bayesian optimization framework. Combining prior probabilities with real-time data, probabilistic inference is performed, outputting the future multi-step inrush water risk level with the maximum posterior probability, thus achieving collaborative decision-making based on multi-source information.

[0026] The sudden water inrush risk model library is constructed by establishing a three-field coupled simulation environment of sandy strata tunnels, tunnel boring machines, and groundwater in GTS NX. Training sample sets are generated based on different types of sandy strata tunnels, different burial depths, and different geological ages, including radar instantaneous energy spectra, microseismic instantaneous energy spectra, residual distances, auxiliary variables, and corresponding real labels for sudden water inrush. Bayesian optimization is used to optimize the hyperparameters of the LSTM-Attention architecture, with the optimization objective being to minimize the cross-entropy loss between the predicted level and the real label. After optimization, the optimal hyperparameters are solidified into a sudden water inrush risk model for that combination of strata-depth-age and added to the library.

[0027] Specifically, the sudden water inrush risk model library is constructed as follows: A three-field coupled simulation environment of sandy strata tunnel-tunneling machine-groundwater is established in GTSNX. Training sample sets are generated according to different sandy strata tunnel types, different burial depths, and different geological ages. The sample sets include radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, residual distance, auxiliary variables, and corresponding real labels of sudden water inrush. Bayesian optimization is used to optimize the hyperparameters of the LSTM-Attention architecture. The optimization objective is to minimize the cross-entropy loss between the predicted level and the real label. After optimization, the optimal hyperparameters are solidified into a sudden water inrush risk model for the combination of strata-burial depth-age and added to the library.

[0028] It should be noted that the process of ground-penetrating radar transmitting step-frequency electromagnetic waves forward and receiving the echoes refers to using the reflection characteristics of electromagnetic waves in different media to detect the geological structure. Specifically, step-frequency continuous wave radar can be used. By transmitting broadband electromagnetic waves and receiving the reflected signals, the non-uniformity characteristics of the sandy strata tunnel ahead can be obtained.

[0029] The Hilbert-Huang transform for obtaining the instantaneous energy spectrum of radar refers to the empirical mode decomposition and Hilbert spectrum analysis of radar echo signals. Specifically, it adaptively decomposes non-stationary signals into intrinsic mode function components and calculates instantaneous frequency and energy distribution, which can accurately characterize the time-frequency characteristics of reflected energy at the tunnel interface in sandy strata.

[0030] The micro-seismic sensor array captures the micro-seismic waveforms induced by tunneling vibration. This refers to the use of multi-axis accelerometers to monitor the elastic waves generated by mechanical vibration. Specifically, it employs an array of piezoelectric accelerometers. Through synchronous acquisition at multiple points in space, it can effectively identify micro-seismic events induced by the expansion of cracks in sandy strata tunnels.

[0031] Wavelet packet decomposition to obtain the instantaneous energy spectrum of microseismic events refers to multi-scale frequency band decomposition of microseismic signals and calculation of the energy of each sub-band. Specifically, a three-level decomposition and reconstruction is performed using the db4 wavelet basis function. By extracting the energy distribution characteristics of different frequency bands, the dynamic process of damage evolution inside tunnels in sandy strata can be reflected. Real-time measurement of the remaining distance from the cutterhead end face to the sandy interface using a laser rangefinder refers to measuring spatial distance using the laser time-of-flight principle. This is specifically achieved using a phase-type laser ranging module. By continuously measuring the distance between the excavation face and the geological interface, spatial location parameters are provided for dynamic risk assessment.

[0032] Energy coupling calculation of radar instantaneous energy spectrum and microseismic instantaneous energy spectrum refers to spatiotemporal alignment and joint analysis of the two types of energy data. Specifically, it uses time window synchronous discretization to calculate the product coupling energy. By fusing the dual characteristics of electromagnetic reflection and mechanical vibration, the reliability of seepage channel identification at sandy interfaces is enhanced. Adaptive threshold comparator dynamically adjusts trigger conditions by automatically correcting the risk judgment threshold based on changes in tunneling speed. Specifically, it dynamically optimizes the sensitivity of risk warning by calculating the speed difference in real time and setting a linear proportional adjustment rule, avoiding misjudgments or missed judgments caused by fixed thresholds.

[0033] The sudden water inrush risk model library, implemented with an LSTM-Attention architecture, refers to constructing a deep learning model with the ability to extract temporal features and focus on key information. Specifically, it uses a Long Short-Term Memory network to capture the temporal dependence of the risk index sequence and combines an attention mechanism to strengthen the weights of key features, enabling accurate prediction of risk levels over multiple time periods. The three-field coupled simulation environment established in GTS NX involves constructing a multi-physics interactive model of geomechanics, mechanical motion, and seepage dynamics. Specifically, by defining the nonlinear constitutive relationship of tunnels in sandy strata, the boundary conditions of the tunnel boring machine, and seepage field parameters, simulation data that conforms to actual working conditions is generated, providing multi-dimensional sample support for model training. Bayesian optimization for hyperparameter optimization refers to using a probabilistic model to guide the parameter search process. Specifically, by constructing a Gaussian process surrogate model and iteratively optimizing the acquisition function, the optimal number of layers, units, and attention heads of the LSTM-Attention network can be efficiently determined, improving the model's prediction accuracy.

[0034] Specifically, the hyperparameters in the LSTM-Attention architecture include the number of LSTM layers, the number of LSTM units per layer, the number of attention heads, the dimension of each attention head, the learning rate, the batch size, and the regularization parameter.

[0035] The number of LSTM layers and the number of units per layer determine the model's learning ability and complexity on time-series data. Increasing the number of LSTM layers and units enhances the model's ability to learn complex patterns. The number of attention heads and the dimension of each attention head determine the flexibility of the attention mechanism and its sensitivity to features. Through multi-head attention mechanisms, the model can learn the importance of input features from multiple perspectives, thereby improving its sensitivity to different features. In this invention, the number of LSTM layers can be selected between 1 and 3. For example, for complex sandy strata tunnel conditions, 2 to 3 LSTM layers are needed to capture the complex patterns of dynamic changes in the strata. The number of units per LSTM layer is selected between 64 and 256 units to ensure that the model can learn sufficiently rich features while avoiding overfitting. For simpler strata conditions, 64 units can be selected; while for complex strata conditions, 128 or 256 units can be selected.

[0036] The number of attention heads is typically between 4 and 8, with each attention head having a dimension between 32 and 64. For simpler formation conditions, 4 attention heads can be selected, each with a dimension of 32; while for complex formation conditions, 8 attention heads can be selected, each with a dimension of 64.

[0037] The learning rate determines the step size for parameter updates. An excessively high learning rate may prevent the model from converging, while an excessively low learning rate will make the training process too slow. Batch size affects the model's training efficiency and memory usage; a larger batch size can improve training efficiency.

[0038] Regularization parameters (such as L2 regularization coefficient and Dropout rate) are used to prevent model overfitting and improve the model's generalization ability by limiting its complexity. The L2 regularization coefficient is between 0.001 and 0.01, and the Dropout rate is between 0.2 and 0.5. For simpler formation conditions, a smaller L2 regularization coefficient (e.g., 0.001) and a lower Dropout rate (e.g., 0.2) can be chosen; for complex formation conditions, a larger L2 regularization coefficient (e.g., 0.01) and a higher Dropout rate (e.g., 0.5) can be chosen.

[0039] Preferably, during the tunnel boring machine's (TBM) excavation process, the ground-penetrating radar emits step-frequency electromagnetic waves forward, with a frequency range of 100MHz to 1GHz. After receiving the echo signal, Hilbert-Huang transform processing is performed to obtain the radar's instantaneous energy spectrum. A microseismic sensor array is arranged on the TBM shell, with a sampling frequency of 1kHz, to capture the microseismic waveforms induced by tunneling vibrations. Wavelet packet decomposition is performed on the microseismic waveforms to obtain the instantaneous energy spectrum of the microseismic waves. A laser rangefinder is installed at the center of the cutterhead, with a measurement accuracy of ±1mm, to measure the remaining distance from the cutterhead end face to the sand-formed interface in real time.

[0040] The instantaneous energy spectrum of the radar and the instantaneous energy spectrum of the microseismic event are aligned within a time window of one tunneling cycle. The product of the instantaneous radar energy and the instantaneous microseismic energy within the same time period is calculated to obtain the coupled energy value. All coupled energy values ​​within the time window are summed and divided by the time window length to obtain the risk index of sudden water inrush at the sandy interface.

[0041] An adaptive threshold comparator reads the current tunneling speed in real time and calculates the speed change by subtracting it from the speed of the previous cycle. The threshold is then dynamically adjusted based on this speed change. When the risk index of sudden water inrush exceeds the threshold or the remaining distance falls below the warning line, dynamic adjustments to the tunneling parameters are triggered. These adjustments include cutterhead speed, propulsion force, screw conveyor speed, and synchronous grouting pressure.

[0042] The target model corresponding to the current tunnel type, burial depth, and geological age in sandy strata is retrieved from the sudden water inrush risk model library. The target model adopts an LSTM-Attention architecture, with the input being a sudden water inrush risk index sequence, a remaining distance sequence, and an auxiliary variable sequence, and the output being the sudden water inrush risk level for the next 10 time periods. The auxiliary variable sequence includes the cutterhead torque fluctuation index, the soil pressure gradient, and the screw conveyor current change rate.

[0043] A three-field coupled simulation environment of tunnel-tunneling machine-groundwater in sandy strata was established in GTS NX. Training sample sets were generated based on different tunnel types, burial depths, and geological ages in sandy strata. Bayesian optimization was used to optimize the hyperparameters of the LSTM-Attention architecture, with the optimization objective being to minimize the cross-entropy loss between the predicted level and the true label. After optimization, the optimal hyperparameters were solidified into a sudden water inrush risk model for that combination of strata-burial depth-age and was stored in the database.

[0044] This application further proposes to discretize the radar instantaneous energy spectrum and the microseismic instantaneous energy spectrum by time period, using the tunneling cycle as the time window; The dynamic weighted allocation values ​​of radar instantaneous energy and microseismic instantaneous energy within the same time period are calculated to obtain the coupled energy value; wherein, the dynamic weighted allocation is adjusted according to the stratigraphic heterogeneity characteristics, specifically as follows: Feature extraction was performed on the instantaneous energy spectrum of radar and the instantaneous energy spectrum of microseismic activity. The extracted features included the peak value and frequency distribution of the energy spectrum. The weighting coefficients of radar instantaneous energy and microseismic instantaneous energy in the current time period are calculated based on the extracted features. The weighting coefficients are related to the formation heterogeneity. The higher the formation heterogeneity, the greater the difference in the weighting coefficients. The instantaneous radar energy and the instantaneous microseismic energy are weighted and summed according to the weighting coefficients to obtain the coupled energy value; The risk index of sudden water inrush at the sandy interface is obtained by summing all the coupled energy values ​​within the time window and dividing by the time window length.

[0045] The time window is divided based on the tunneling cycle to ensure that data acquisition is synchronized with the construction progress. Discretization uses a fixed time period to keep the sampling points aligned for the two energy spectrum time series. Dynamic weight allocation is adjusted according to the formation heterogeneity characteristics, which can improve the sensitivity of the risk index to formation heterogeneity and make risk prediction more accurate.

[0046] Specifically, the tunneling cycle is set as the length of a time window, and the radar and microseismic data acquisition systems are synchronously triggered via a unified clock signal. Within each time period, weighted allocation coefficients for the radar instantaneous energy and microseismic instantaneous energy are calculated based on the formation heterogeneity characteristics. These weighted allocation coefficients are related to formation heterogeneity; the higher the formation heterogeneity, the greater the difference in weighted allocation coefficients. The radar instantaneous energy and microseismic instantaneous energy are weighted and summed according to the weighted allocation coefficients to obtain the coupled energy value. The coupled energy values ​​of all time periods within the time window are accumulated and divided by the total number of time periods to obtain a normalized risk index. For example, when the time window contains 100 time periods, the coupled energy values ​​of each time period are accumulated and divided by 100 to form a standardized risk index between 0 and 1. This calculation method can effectively identify the coordinated change patterns of the two energy spectra, avoiding misjudgments caused by fluctuations in data from a single sensor.

[0047] Preferably, aligning the radar instantaneous energy spectrum and the microseismic instantaneous energy spectrum within the same time window and performing energy coupling calculations includes the following steps: First, using the tunneling cycle as a time window, the instantaneous radar energy spectrum and the instantaneous microseismic energy spectrum are discretized by time interval. Specifically, a complete tunneling cycle is selected as the time window, typically ranging from 30 minutes to 2 hours, depending on the specific tunneling conditions. Within this time window, the instantaneous radar energy spectrum and the instantaneous microseismic energy spectrum are sampled at the same time interval, for example, once every 5 seconds, thus obtaining two sets of discretized energy spectrum data sequences.

[0048] The dynamic weighted allocation values ​​of radar instantaneous energy and microseismic instantaneous energy within the same time period are calculated to obtain the coupled energy value. Specifically, for each sampling time point, features are extracted from both the radar instantaneous energy spectrum and the microseismic instantaneous energy spectrum. Extracted features include peak values ​​and frequency distributions of the energy spectra. Based on the extracted features, the weighted allocation coefficients of the radar instantaneous energy and microseismic instantaneous energy in the current time period are calculated. These weighted allocation coefficients are related to stratigraphic heterogeneity; the higher the stratigraphic heterogeneity, the greater the difference in weighted allocation coefficients. The radar instantaneous energy and microseismic instantaneous energy are then weighted and summed according to the weighted allocation coefficients to obtain the coupled energy value. This step achieves energy coupling between two signals from different sources, enabling a more comprehensive reflection of the characteristics of tunnels in sandy strata.

[0049] Finally, the coupling energy values ​​within the time window are summed and divided by the time window length to obtain the risk index of sudden water inrush at the sandy interface. For example, if the time window is 1 hour and the sampling interval is 5 seconds, there are a total of 720 coupling energy values. Adding these 720 values ​​and then dividing by 3600 seconds (1 hour) gives the average coupling energy within that time window, which is the risk index of sudden water inrush at the sandy interface.

[0050] This application further proposes to read the current tunneling speed value in the tunneling machine PLC in real time, and obtain the speed change by differentiating the tunneling speed value of the previous cycle; if the speed change is positive and higher than the set rate, the threshold increases linearly; if the speed change is negative and the absolute value is higher than the set rate, the threshold decreases exponentially; if the absolute value of the speed change does not exceed the set rate, the threshold remains unchanged; the adjusted threshold is immediately written into the adaptive threshold comparator and takes effect.

[0051] The current tunneling speed is acquired via a PLC bus at a fixed sampling frequency, and differential calculation is performed using the average speed difference between two adjacent tunneling cycles. The set rate is pre-defined based on the tunnel type in sandy strata; for example, a rate of 0.2 m / min is set in loose sand. Linear increases use a function relating the threshold increment to the speed change, while exponential decreases use an exponential function where the threshold multiplier increases with the absolute value of the speed change. The adjusted threshold is updated in real-time to the comparator kernel via shared memory, ensuring immediate effect in the next tunneling cycle.

[0052] Specifically, when the tunneling speed increases significantly, a linear increase in the threshold can prevent false alarms caused by excessively rapid advancement; when the speed drops sharply, an exponential decrease in the threshold can quickly respond to potential risks. The threshold adjustment process forms a closed-loop feedback with the tunneling speed, ensuring the real-time nature and accuracy of risk warnings by dynamically balancing risk sensitivity and false alarm rate. For example, when the speed change reaches 1.5 times the set rate, the threshold is adjusted by increasing by 0.5% per unit speed change, while when the speed decreases beyond the set rate, the threshold is adjusted exponentially by decreasing by 1.2% per unit speed change. This mechanism allows the risk assessment threshold to adaptively match the current construction status, effectively improving the accuracy of risk warnings.

[0053] Preferably, the current tunneling speed value is read in real time in the tunneling machine PLC, and the speed change is obtained by subtracting the tunneling speed value from the previous cycle. Specifically, the current signal from the tunneling speed sensor can be acquired through the analog input module of the tunneling machine PLC, and the 4-20mA current signal can be converted into a speed value of 0-100mm / min. At the end of each tunneling cycle, the current speed value is subtracted from the speed value recorded at the end of the previous cycle to obtain the speed change.

[0054] If the speed change is positive and exceeds the set rate, the threshold increases linearly. For example, if the set rate is 5 mm / min, and the speed change is 8 mm / min, the threshold increases by (8-5) / 5*100%=60%.

[0055] If the velocity change is negative and its absolute value is higher than the set rate, the threshold decreases exponentially. For example, if the set rate is 5 mm / min, and the velocity change is -7 mm / min, the threshold decreases by [percentage missing]. .

[0056] If the absolute value of the speed change does not exceed the set rate, the threshold remains unchanged.

[0057] The adjusted threshold is immediately written to the adaptive threshold comparator and takes effect. In practice, the adjusted threshold value can be transmitted to the memory of the adaptive threshold comparator via industrial Ethernet, and the comparator reads the latest threshold before each comparison operation.

[0058] This application further proposes to discretize the tunnel axis into a set of nodes at fixed intervals; for each node, call the inrush water risk model to calculate the maximum inrush water risk level of the node in the future for several time periods; if the maximum inrush water risk level exceeds the preset level, the node is marked as a risk node; construct a cost function, which consists of the weighted sum of the node risk level and the Euclidean distance from the node to the current cutterhead position; use the A* algorithm to search for the minimum cost path from the current cutterhead position to the end point of the tunnel design axis in the node set, multiply the cost of the risk node by a penalty coefficient, which monotonically increases with the risk level; connect the minimum cost paths in sequence to generate a continuous risk avoidance path and send it to the tunneling machine navigation system for execution.

[0059] The discretized spacing of the tunnel axis is set to an integer multiple of the tunnel boring machine's advance step length to ensure that the node spacing matches the mechanical motion characteristics. Risk node marking is based on multi-time-period prediction results output by the sudden water inrush risk model, and future risk peaks are extracted through a sliding time window. The cost function weight coefficients are dynamically configured according to the project's safety level, with the risk level weight accounting for over 70% in high-risk conditions. The A* algorithm uses Manhattan distance as a heuristic function, and the penalty coefficient for risk nodes during path search is divided into three levels of gradients according to risk level, with each gradient corresponding to a 50% increase in penalty coefficient.

[0060] Specifically, the tunnel axis is discretized into a set of nodes spaced 1 meter apart, with each node corresponding to the risk prediction results for the next three tunneling cycles. The inrush water risk model outputs the risk level of each node for the future time period; if the level exceeds a threshold at any time, the node is marked as a risk node. The risk level weight in the cost function is set to 0.6, and the Euclidean distance weight is set to 0.4, ensuring that path selection prioritizes avoiding high-risk areas. During the search process, the A* algorithm multiplies the base cost of risk nodes by a three-level penalty coefficient of 1.5, 2.25, or 3.375, corresponding to low, medium, and high risk levels. The final minimum cost path is converted into tunnel boring machine (TBM) navigation instructions through the BIM system, controlling the cutterhead to automatically avoid high-risk node areas while ensuring the shortest total path length. This method achieves an optimal balance between safety and efficiency in path planning through dynamic risk node marking and gradient penalty mechanisms.

[0061] Preferably, the tunnel axis is discretized into a set of nodes at 5-meter intervals. For each node, a sudden water inrush risk model is called to calculate the maximum sudden water inrush risk level for that node over the next 10 time periods. If the maximum sudden water inrush risk level exceeds level 3, the node is marked as a risk node. A cost function is constructed, consisting of the weighted sum of the node's risk level and the Euclidean distance from the node to the current cutterhead position. The A* algorithm is used to search for the minimum cost path from the current cutterhead position to the end point of the tunnel's designed axis within the node set. The cost of the risk node is multiplied by a penalty coefficient, which monotonically increases with the risk level. The minimum cost paths are connected sequentially to generate a continuous risk avoidance path, which is then sent to the tunnel boring machine's navigation system for execution.

[0062] Specifically, the tunnel axis is first discretized into a set of nodes at 5-meter intervals. Then, a sudden water inrush risk model is applied to each node to calculate its maximum sudden water inrush risk level over the next 10 time periods. If the maximum sudden water inrush risk level exceeds level 3, the node is marked as a risk node. Next, a cost function is constructed, which consists of the weighted sum of the node's risk level and the Euclidean distance from the node to the current cutterhead position. The cost function can be expressed as: ; Where w1 and w2 are weighting coefficients, RiskLevel is the node risk level, and Distance is the Euclidean distance from the node to the current cutterhead position. Specifically, in high-risk conditions, to ensure construction safety, the value of w1 is usually higher (e.g., 0.7 to 0.9). In low-risk conditions, the value of w1 can be appropriately reduced (e.g., 0.5 to 0.7) to balance path length and risk avoidance. In high-risk conditions, the value of w2 is usually lower (e.g., 0.1 to 0.3). In low-risk conditions, the value of w2 can be appropriately increased (e.g., 0.3 to 0.5) to optimize path length.

[0063] Furthermore, the A* algorithm is used to search the node set for the minimum-cost path from the current cutterhead position to the end point of the tunnel design axis. For risky nodes, their cost is multiplied by a penalty coefficient, which monotonically increases with the risk level. For example, the penalty coefficient can be set as follows: ; Finally, the minimum cost paths obtained from the search are connected in sequence to generate a continuous risk avoidance path, and the path is sent to the tunneling machine navigation system for execution via industrial Ethernet.

[0064] This application further proposes to establish a three-dimensional simulation scene on the GTS NX platform, for each sandy stratum tunnel type-depth-geological age triplet, fully coupled with the physical model of the sandy stratum tunnel, the physical model of the tunnel boring machine, and the groundwater seepage field; to generate a set of working conditions covering the parameter space using Latin hypercube sampling, the input of each working condition is the radar instantaneous energy spectrum, the microseismic instantaneous energy spectrum, the remaining distance and auxiliary variables, and the output is the true label of the inrush water; to run the simulation and extract the inrush water occurrence time, inrush volume, and peak pore water pressure to form training samples; to use Bayesian optimization to globally optimize the hyperparameters of the LSTM-Attention architecture, with cross-entropy loss as the optimization objective; to solidify the optimal hyperparameters into the inrush water risk model of the triplet, and store it in the inrush water risk model library with stratum type-depth-geological age as the key, while saving the model weights and normalized parameters.

[0065] Among them, the solid model of the sandy stratum tunnel is given nonlinear elastoplastic constitutive and permeability parameters, the solid model of the tunnel boring machine is set with boundary conditions for the cutterhead, screw conveyor and grouting hole, and the groundwater seepage field is set with hydraulic head boundary and permeability coefficient; Latin hypercube sampling uniformly distributes sampling points in the parameter space to ensure that the training samples cover all possible working condition combinations; Bayesian optimization models the relationship between hyperparameters and cross-entropy loss through Gaussian process, selects the next set of candidate hyperparameters through the acquisition function, and iteratively optimizes until convergence.

[0066] Specifically, the solid model of the sandy strata tunnel simulates the deformation and seepage characteristics of the strata through mesh generation and nonlinear constitutive equations. The solid model of the tunnel boring machine reflects the cutting process of the cutterhead and the soil removal process of the screw conveyor through boundary condition settings. The groundwater seepage field simulates the pore water pressure distribution through Darcy's law and the continuity equation. Latin hypercube sampling divides each parameter dimension into equal probability intervals, selecting only one sample point in each interval to eliminate sample clustering. During simulation, the three-field coupled solver calculates the interaction of the displacement field, seepage field, and stress field in real time, extracting the inflow rate and peak pore water pressure at the moment of sudden water inrush as true labels. In the Bayesian optimization process, the Gaussian process surrogate model constructs a probability distribution based on historical hyperparameter combinations and cross-entropy loss, and selects the next set of hyperparameters through the expectation improvement function to achieve global optimal search. When saving model weights and normalized parameters, the connection weights of the LSTM layer and attention layer are stored in binary serialization format, and the mean and standard deviation of the input data are recorded for data standardization during online inference.

[0067] Preferably, in the GTS NX platform, for each sandy stratum tunnel type-depth-geological age triplet, a three-dimensional simulation scene is established that fully couples the sandy stratum tunnel entity model, the tunnel boring machine entity model, and the groundwater seepage field. A set of working conditions covering the parameter space is generated using Latin hypercube sampling. The input for each working condition is the radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, remaining distance, and auxiliary variables; the output is the true label of the inrush water. The simulation is run to extract the inrush water occurrence time, inrush volume, and peak pore water pressure to form training samples. Bayesian optimization is used to globally optimize the hyperparameters of the LSTM-Attention architecture, with cross-entropy loss as the optimization objective. The optimal hyperparameters are solidified into the inrush water risk model for that triplet and stored in the inrush water risk model library using stratum type-depth-geological age as the key, while also saving the model weights and normalized parameters.

[0068] Specifically, the first step is to establish a solid model of the sandy strata tunnel in the GTS NX platform. This model includes the tunnel's geometry, physical parameters, and mechanical parameters. For example, parameters such as porosity, permeability coefficient, elastic modulus, and Poisson's ratio are set. Next, a solid model of the tunnel boring machine (TBM) is established, including major components such as the cutterhead, shield, and screw conveyor. Finally, a groundwater seepage field model is established, setting the initial hydraulic head and boundary conditions.

[0069] Furthermore, the Latin hypercube sampling method is used to generate a set of working cases. For example, for a specific combination of tunnel depth and geological age in sandy strata, 100 different working cases are generated. Each working case includes different tunneling parameters, such as cutterhead rotation speed and advance speed, and strata parameters, such as the range of variation of permeability coefficient and strength parameters.

[0070] Therefore, simulations are run and key data is extracted. For each working condition, the tunneling process is simulated and the time of water inrush, water inflow rate, and changes in pore water pressure are recorded. For example, when the pore water pressure exceeds a set threshold, it is recorded as a water inrush, and the parameter values ​​at that moment are extracted.

[0071] In this study, Bayesian optimization was employed to optimize the hyperparameters of the LSTM-Attention model. These hyperparameters included the number of LSTM layers, the number of hidden units, and the type of attention mechanism. During optimization, cross-entropy loss was used as the objective function, and the optimal combination of hyperparameters was found through multiple iterations.

[0072] Finally, the optimized model parameters are saved to the sudden water inrush risk model library. Each model is uniquely identified by its sandy stratum tunnel type, burial depth, and geological age, facilitating rapid retrieval and application later.

[0073] This application further proposes installing a torque sensor at the main shaft of the cutterhead drive motor to continuously acquire the raw torque signal of the cutterhead at a fixed frequency; performing a moving average filter on the raw signal to obtain a filtered torque signal; dividing the filtered torque signal into segments of fixed duration and calculating the variance of each segment to obtain a short-time variance sequence; performing max-min normalization on the short-time variance sequence to obtain the cutterhead torque fluctuation index; and triggering an early warning signal and recording the tunneling cycle number when the cutterhead torque fluctuation index exceeds a set limit.

[0074] The torque sensor employs a ring structure that fits tightly against the motor spindle surface, ensuring synchronous measurement of axial and radial forces. The window length of the moving average filter is set to an integer multiple of the torque signal's dominant frequency period, effectively suppressing high-frequency noise interference. The interval between fixed-duration segments is dynamically adjusted according to the tunneling speed; as the tunneling speed increases, the segment duration is shortened to improve sampling density. Maximum-minimum normalization uses extreme values ​​within the sliding window for dynamic range compression, eliminating interference from dimensional differences on subsequent model inputs. Upon triggering a warning signal, the system automatically associates the spatiotemporal coordinates of the current tunneling cycle, providing data support for risk tracing.

[0075] Specifically, the torque sensor captures the raw signal of the main shaft torque in real time at a sampling frequency of 2000Hz. Mean filtering is performed through a sliding window with 50 sampling points to eliminate instantaneous impact noise. The filtered signal is divided into 10-second time intervals, and the variance of the torque value within each interval is calculated to form a short-time variance sequence. This sequence is then normalized using the maximum and minimum values ​​of the most recent 100 time intervals as a window, mapping the variance values ​​to the 0-1 range to generate a torque fluctuation index. When the index exceeds the 0.85 threshold for three consecutive time intervals, the system triggers a level-three warning and marks the current cycle number in the tunneling log. This method, through high-frequency sampling and dynamic processing, effectively identifies abnormal friction states at the contact surface between the cutterhead and the sandy strata tunnel, providing real-time and reliable auxiliary variable input for predicting the risk of sudden water inrush.

[0076] Preferably, a high-precision torque sensor is installed at the spindle of the cutterhead drive motor, with a sampling frequency set to 1000Hz, to continuously acquire the raw cutterhead torque signal. A 20-point moving average filter is applied to the raw signal to obtain the filtered torque signal. The filtered torque signal is then segmented into 10-second intervals, and the variance of each segment is calculated to form a short-time variance sequence. The short-time variance sequence is then subjected to max-min normalization to obtain the cutterhead torque fluctuation index. When the cutterhead torque fluctuation index exceeds 0.8, an early warning signal is triggered, and the current tunneling cycle number is recorded.

[0077] Specifically, the torque sensor is a strain gauge torque sensor with a range of 0-100 kN·m and an accuracy class of 0.1. The moving average filtering uses a 20-point equally weighted average. The short-time variance is calculated using an unbiased estimation method. The maximum-minimum normalization formula is: , Where x is the current value, and min and max are the minimum and maximum values ​​of the short-time variance sequence, respectively. Early warning signals are transmitted to the control room and on-site LED displays via industrial Ethernet. Tunneling cycle numbers are automatically generated and recorded by the PLC.

[0078] This application further proposes to establish a three-dimensional geological model of a sandy stratum tunnel in GTS NX and assign nonlinear elastoplastic constitutive and permeability parameters; establish a tunnel boring machine shell model and set boundary conditions for the cutterhead, screw conveyor and grouting holes; establish a groundwater seepage field and set hydraulic head boundary and permeability coefficient; mesh the three-field model and set a coupled solver, with coupling variables including displacement, pore water pressure and seepage velocity.

[0079] Among them, the three-dimensional geological model of the sandy stratum tunnel adopts a nonlinear elastoplastic constitutive equation to describe the stratum mechanical behavior, and the permeability parameter is related to porosity and permeability coefficient through Darcy's law; in the tunnel boring machine shell model, the cutterhead boundary condition is set as the combined action of rotation speed and propulsion force, the screw conveyor boundary condition is the friction coefficient of the contact surface between the screw blade and the soil, and the grouting hole boundary condition is pressure-driven fluid injection; the hydraulic head boundary of the groundwater seepage field is set with an initial hydraulic gradient based on geological exploration data, and the permeability coefficient is calibrated through laboratory tests; the meshing of the three-field model adopts a mixture of tetrahedral and hexahedral elements, and the coupled solver calculates the displacement field, seepage field and pore water pressure field synchronously and iteratively based on the finite element method.

[0080] Specifically, the three-dimensional geological model of the sandy strata tunnel is constructed by importing geological exploration data to establish its geometry. Nonlinear elastoplastic constitutive parameters include the internal friction angle, cohesion, and dilatation angle. Permeability parameters are calculated using porosity and the Kozeny-Carman equation. In the tunnel boring machine (TBM) shell model, the cutterhead rotation speed and propulsion force are coupled with the soil interaction model through kinematic equations. The friction coefficient between the screw conveyor blades and the soil contact surface is set based on the material surface roughness. The grouting hole pressure is correlated with the stratum permeability through fluid dynamics equations. The hydraulic head boundary of the groundwater seepage field is initialized based on borehole water level data, and the permeability coefficient is obtained through triaxial permeability tests. During mesh generation, the contact area between the strata and the TBM uses locally refined meshes, with element sizes controlled between 5-10 mm. The coupled solver synchronously updates displacement, pore water pressure, and seepage velocity using the Newton-Raphson iterative method, with each iteration converging under the condition that the residual norm is less than 1e-4. This scheme ensures that the simulation data accurately reflects the actual working conditions by precisely coupling the three physical fields, providing high-fidelity samples for model training.

[0081] Preferably, a three-dimensional geological model of the tunnel in sandy strata is established in GTS NX software. This model includes multiple strata, each assigned nonlinear elastoplastic constitutive parameters and permeability parameters. The nonlinear elastoplastic constitutive parameters include elastic modulus, Poisson's ratio, cohesion, and internal friction angle, etc. The permeability parameters include permeability coefficient and porosity, etc.

[0082] A model of the tunnel boring machine (TBM) casing is established. This model includes the cutterhead, screw conveyor, and grouting holes. Rotational speed and feed rate conditions are set at the cutterhead boundary. Soil discharge rate conditions are set at the screw conveyor boundary. Grouting pressure and flow rate conditions are set at the grouting hole boundary.

[0083] Establish a groundwater seepage field model. Set a fixed hydraulic head boundary condition at the model boundary. Set an initial pore water pressure distribution within the formation. Assign a permeability coefficient to each formation unit.

[0084] Mesh the three models. The sandy strata tunnel model uses tetrahedral elements. The tunnel boring machine shell model uses hexahedral elements. The groundwater seepage field model uses triangular elements.

[0085] A coupled solver is configured. The coupled variables include the displacement field, pore water pressure field, and seepage velocity field. An iterative solution method is used, alternately updating the solutions for the three fields in each iteration. The convergence criterion is set as the relative error of displacement, pore water pressure, and seepage velocity being less than 1e-6.

[0086] This application further proposes that when searching for the target model corresponding to the current sandy stratum tunnel type, burial depth and geological age in the sudden water inrush risk model library, if there is no perfectly matching model, then similarity retrieval and transfer learning are performed.

[0087] The process involves calculating the cosine similarity between the current formation feature vector and the corresponding formation feature vectors of all models in the database. The model with the highest similarity is selected as the source model, and its attention layer weights are frozen. Only the LSTM layer weights are fine-tuned. The LSTM layer weights are updated online using real-time data collected from the current tunneling section. After the update, the model is fixed and added to the database. Specifically, the cosine similarity is calculated using the ratio of the vector dot product to the product of the magnitudes. Freezing the attention layer weights preserves the source model's ability to extract key features. Fine-tuning the LSTM layer weights adapts to the time-series characteristics of new formations. The online update uses a sliding time window mechanism, with each time window step being one tunneling cycle. The preprocessing steps include outlier removal and missing value imputation. The adaptive optimizer adjusts the learning rate based on the sample variance, and the elastic weight consolidation term prevents the forgetting of existing knowledge during the fine-tuning process.

[0088] Specifically, when a model is not found in the database, the current stratigraphic type, depth, and geological age are first converted into a multi-dimensional feature vector. The closest source model is then matched using cosine similarity. The attention layer weights of the source model are fixed, and only the LSTM layer parameters are updated. Real-time acquired tunneling data is processed through a sliding window and input into the model. The cross-entropy loss between the predicted and measured levels is calculated, and the LSTM layer weights are adjusted using gradient descent. The learning rate dynamically decays based on the variance of the data within the window; when the variance is large, the learning rate is reduced to stabilize the training process. To prevent damage to the original parameters of the source model during fine-tuning, an elastic weight consolidation term is introduced into the loss function to constrain the magnitude of weight updates. When the loss value remains consistently below a threshold, updates are stopped, and the adjusted model is re-stored in the model database, overwriting the original record. This process enables the model to quickly adapt to new stratigraphic conditions while maintaining efficient reuse of historical knowledge.

[0089] As a preferred embodiment, the solution of this application is implemented as follows: During shield tunneling, if the type, depth, and geological age combination of the current sandy stratum tunnel do not match a corresponding model in the risk model library, a model similarity retrieval is performed. Specifically, the current stratum type is encoded as a discrete categorical variable, the depth is converted into a normalized value, and the geological age is mapped to a geological age index; these three constitute a stratum feature vector. By calculating the cosine similarity between this vector and the feature vectors of all models in the model library, candidate models with similarity higher than a preset threshold are selected.

[0090] The preset threshold is determined through statistical analysis of historical tunneling data combined with Bayesian optimization methods. By collecting radar instantaneous energy spectra, microseismic instantaneous energy spectra, and remaining distance data just moments before a sudden water inrush occurs under different sandy strata tunnel conditions, cluster analysis is used to determine the typical distribution intervals of risk characteristics. Then, a Bayesian optimization algorithm is employed to optimize the threshold parameter with the objective of minimizing cross-entropy loss, thereby obtaining the optimal discrimination value that balances the false alarm rate and the missed alarm rate. This threshold is also dynamically adjusted according to changes in tunneling speed and takes effect in real time through an adaptive threshold comparator.

[0091] The candidate model with the highest similarity was selected as the source model, retaining the weight parameters of its attention layer, and only unfreezing the weights of the hidden units in the LSTM layer. The instantaneous radar energy spectrum, microseismic energy spectrum, remaining distance, and auxiliary variable sequences acquired in real-time for the current tunneling section were input into the source model. A sliding time window mechanism was used to extract continuous data blocks according to the tunneling cycle step size. Outlier detection was performed on the data blocks, and outliers were removed using the interquartile range method. Missing values ​​were filled in using linear interpolation.

[0092] The preprocessed data is input into the model with the frozen attention layer to generate a predicted risk level sequence and calculate the cross-entropy loss with the measured level sequence. An Adam optimizer with elastic weight consolidation terms is used to update the gradient of the LSTM layer weights, with the learning rate dynamically adjusted based on the sample variance within the time window. The weight update process terminates when the cross-entropy loss for three consecutive tunneling cycles is below a preset threshold and the loss decreases by less than 5%. Finally, the adjusted model parameters are stored in the risk model library and associated with the current formation feature vector as a new key.

[0093] This application further proposes to calculate the cosine similarity between the current formation feature vector and the corresponding formation feature vectors of all models in the library, select the model with the highest similarity as the source model, freeze its attention layer weights, fine-tune only the LSTM layer weights, use the real-time data collected in the current tunneling section to update the LSTM layer weights online, and solidify the model and add it to the library after the update is completed.

[0094] The process involves constructing a fixed-length sliding time window with a step size equal to one tunneling cycle. Within this time window, outliers are removed and missing values ​​are filled in for real-time acquired radar instantaneous energy spectra, microseismic instantaneous energy spectra, remaining distances, and auxiliary variables. The preprocessed data is then input into the source model of the frozen attention layer to obtain the predicted inrush water risk level. The cross-entropy loss is calculated by comparing the predicted level with the measured level. An adaptive optimizer is used to update the LSTM layer weights using gradients. The learning rate adaptively decreases with the sample variance within the time window. An elastic weight consolidation term is added to the loss function to prevent catastrophic forgetting. The update stops when the cross-entropy loss of several consecutive time windows is lower than a set threshold and the decrease in loss is lower than a set percentage. The current weights are then solidified to form a new inrush water risk model, and the original database record is overwritten using the current stratum type-depth-geological age as the key.

[0095] Specifically, the sliding time window is set to ten tunneling cycles. Data within each time window undergoes a three-standard-deviation method to remove outliers, and missing values ​​are filled using linear interpolation. After preprocessing, the data is input into the source model of the frozen attention layer. Cross-entropy loss is calculated between the predicted risk level and the actual monitored water inflow level. The adaptive optimizer dynamically adjusts the learning rate based on the sample variance within the time window; the larger the variance, the smaller the learning rate decay coefficient, ensuring stable model updates even with significant data fluctuations. The elastic weight consolidation term constrains the parameter update direction by calculating the Fisher information matrix of the LSTM layer parameters between the source and target tasks, preserving key features of the source model. When the cross-entropy loss is below 0.1 for three consecutive time windows and the loss decreases by less than 5%, the model is considered converged, updates cease, and the updated parameters overwrite the corresponding formation type model in the original database. Through the frozen attention layer weights and elastic weight consolidation mechanism, both the source model's ability to capture macroscopic formation features and the adaptive adjustment of the LSTM layer to the current dynamic characteristics of the formation are preserved, effectively solving the catastrophic forgetting problem during model migration.

[0096] As a preferred embodiment, the specific implementation of this application's solution is as follows: A fixed-length sliding time window is constructed, with the sliding step size set to one tunneling cycle. The time window contains real-time data from multiple consecutive tunneling cycles. Within the time window, outliers are removed from the real-time acquired radar instantaneous energy spectrum. Data points exceeding the upper and lower limits are identified and deleted using the interquartile range method. Missing values ​​in the microseismic instantaneous energy spectrum are filled using a linear imputation method. The preprocessed radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, remaining distance, and auxiliary variables are input into the source model of the frozen attention layer, outputting the predicted inrush water risk level. The predicted level is compared with the actual measured level on-site, and the cross-entropy is calculated. Loss value; The adaptive moment estimation algorithm is used to update the gradient of the LSTM layer weights. The initial learning rate is set to 0.001 and dynamically adjusted according to the sample variance within the time window. The larger the variance, the smaller the learning rate decay coefficient. An elastic weight consolidation term based on the Fisher information matrix is ​​added to the loss function to retain the key weight parameters of the source model. When the cross-entropy loss value of five consecutive time windows is lower than 0.2 and the loss decrease is less than 5%, the online update process is terminated. The current optimized LSTM layer weights are combined with the frozen attention layer to form a new inrush water risk model. The records in the model library are updated with the current stratum type, burial depth and geological age as indexes.

[0097] This application further proposes mapping the risk level of sudden water inrush to a risk color mark, overlaying the risk color mark on the tunnel BIM model to form a visualized risk cloud map, and automatically generating shield machine control commands based on the risk cloud map. The commands include quantitative values ​​of cutterhead speed, thrust, synchronous grouting pressure and screw conveyor speed. The control commands are then sent to the shield machine PLC for execution via industrial Ethernet.

[0098] The risk color scheme uses four colors: red, orange, yellow, and green. Red represents the highest risk level, corresponding to a reduction in cutterhead speed to a preset minimum threshold, a reduction in thrust to 50% of the original set value, and an increase in synchronous grouting pressure to 120% of the design limit. The visualized risk cloud map binds the color scheme to the tunnel axis coordinates through a 3D rendering engine, updating the risk distribution area in real time. The automatically generated shield machine control commands are implemented through a preset fuzzy control rule base, which stores parameter adjustment coefficient matrices corresponding to different risk levels. The industrial Ethernet uses the Profinet protocol, with a command transmission latency of less than 50 milliseconds.

[0099] Specifically, after the LSTM-Attention model outputs the risk levels of sudden water inrush over multiple future time periods, the risk level data is converted into standardized color-coded labels. These color-coded labels are written into the tunnel's 3D model via the BIM interface module, generating a dynamically updated risk cloud map. The risk cloud map triggers a control command generation module, which associates the risk level with adjustments to construction parameters based on a pre-defined mapping table, generating JSON-formatted commands containing specific values. These commands are transmitted to the PLC via industrial Ethernet. The PLC parses the commands and directly adjusts the operating parameters of the cutterhead drive motor, propulsion hydraulic cylinder, grouting pump, and screw conveyor. For example, when the risk level for a certain section is orange, the cutterhead speed decreases from 1.2 rpm to 0.8 rpm, the propulsion force is adjusted from 18000 kN to 12000 kN, and the synchronous grouting pressure is increased from 0.3 MPa to 0.36 MPa. The entire process is completed within 3 seconds, achieving closed-loop control from risk prediction results to construction parameter adjustments.

[0100] As a preferred embodiment, the specific implementation of this application is as follows: During the tunnel boring machine (TBM) construction process, after the risk level is output, the risk color code adopts a four-level mapping of red, orange, yellow, and green, corresponding to high risk, medium-high risk, medium risk, and low risk, respectively. The coordinate system of the tunnel BIM model is kept synchronized with the coordinate system of the TBM navigation system. The risk color code is integrated into the three-dimensional geological structure of the BIM model through a transparency overlay algorithm to form a dynamically updated risk cloud map. The generation of TBM control commands is based on a preset mapping rule table, wherein the adjustment step of the cutterhead rotation speed is set to ±5 r / min, the adjustment amount of the thrust is calculated according to the linear relationship between the risk level and the current earth pressure chamber pressure difference, and the synchronous grouting pressure is adjusted in a gradient of 0.2 MPa for each increase in risk level. The industrial Ethernet adopts the PROFINET protocol, and the control commands are encapsulated as structured text data packets conforming to the IEC 61131-3 standard, and transmitted to the input / output module of the TBM PLC through redundant network channels.

[0101] Through the above technical solution, this application achieves real-time and visualization of construction command generation. Operators can intuitively identify the spatial distribution characteristics of high-risk areas. At the same time, the automated command generation mechanism avoids the problem of parameter lag caused by manual adjustment, ensuring that the shield machine's operating parameters are dynamically matched with the risk of sudden water inrush, and significantly reducing the probability of sudden water inrush disasters caused by untimely parameter adjustments.

[0102] like Figure 2 As shown in some embodiments, a tunnel inrush water risk prediction system in sandy strata includes: The sudden water inrush risk index calculation module 201 is used to transmit step-frequency electromagnetic waves to the tunneling front via ground-penetrating radar and receive the echoes, and obtain the radar instantaneous energy spectrum after Hilbert-Huang transform; capture the micro-vibration waveform induced by tunneling vibration through a micro-seismic sensor array, and obtain the micro-seismic instantaneous energy spectrum after wavelet packet decomposition; measure the remaining distance from the cutterhead end face to the sandy interface in real time using a laser rangefinder; align the radar instantaneous energy spectrum and the micro-seismic instantaneous energy spectrum within the same time window and perform energy coupling calculation to obtain the sudden water inrush risk index of the sandy interface; The risk avoidance path generation module 202 is used to input the sudden water inrush risk index and the remaining distance into the adaptive threshold comparator. When the risk index is higher than the threshold or the remaining distance is lower than the warning line, the excavation parameters are immediately dynamically adjusted. The adjustment objects include the cutterhead speed, propulsion force, screw conveyor speed and synchronous grouting pressure. After the adjustment is completed, the risk avoidance path of the tunneling machine is generated based on the adjusted parameters. Output module 203 is used to retrieve the target model corresponding to the current sandy stratum tunnel type, burial depth and geological age from the sudden water inrush risk model library. The target model is implemented with LSTM-Attention architecture. The input is the sudden water inrush risk index sequence, the remaining distance sequence and the auxiliary variable sequence. The output is the sudden water inrush risk level for multiple future time periods. The auxiliary variable sequence includes the cutterhead torque fluctuation index, the soil pressure gradient, and the screw conveyor current change rate.

[0103] It is understandable that the modules recorded in this sandy stratum tunnel inrush water risk prediction system are similar to those in the reference system. Figure 1 The steps in the method for predicting the risk of sudden water inrush in tunnels in sandy strata described above correspond to each other. Therefore, the operation, characteristics, and beneficial effects described above for the method for predicting the risk of sudden water inrush in tunnels in sandy strata are also applicable to the system for predicting the risk of sudden water inrush in tunnels in sandy strata and the modules contained therein, and will not be repeated here.

[0104] The following is for reference. Figure 3 The diagram illustrates a structural schematic of an electronic device 300 suitable for implementing some embodiments of the present invention. The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 3 The terminal device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0105] like Figure 3As shown, the electronic device 300 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0106] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0107] Furthermore, the storage medium in the embodiments of this application stores program instructions capable of implementing all the above methods. These program instructions can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0108] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the risk of sudden water inrush in tunnels in sandy strata, characterized in that, include: The instantaneous energy spectrum of the radar is obtained by transmitting step-frequency electromagnetic waves forward through ground-penetrating radar and receiving the echoes, and then performing Hilbert-Huang transform. The micro-vibration waveforms induced by tunneling vibration are captured by a micro-seismic sensor array, and the instantaneous energy spectrum of the micro-seismic waves is obtained after wavelet packet decomposition. The remaining distance from the cutterhead end face to the sandy interface is measured in real time using a laser rangefinder. The instantaneous energy spectrum of the radar and the instantaneous energy spectrum of the micro-seismic waves are aligned within the same time window and energy coupling calculation is performed to obtain the risk index of sudden water inrush at the sandy interface. The risk index of sudden water inrush and the remaining distance are input into an adaptive threshold comparator. When the risk index is higher than the threshold or the remaining distance is lower than the warning line, the excavation parameters are immediately dynamically adjusted. The adjustment objects include the cutterhead speed, propulsion force, screw conveyor speed and synchronous grouting pressure. After the adjustment is completed, the risk avoidance path of the tunneling machine is generated based on the adjusted parameters. The target model corresponding to the current sandy stratum tunnel type, burial depth and geological age is retrieved from the sudden water inrush risk model library. The target model is implemented with LSTM-Attention architecture. The input is the sudden water inrush risk index sequence, the remaining distance sequence and the auxiliary variable sequence. The output is the sudden water inrush risk level for multiple future time periods. The auxiliary variable sequence includes the cutterhead torque fluctuation index, the soil pressure gradient and the screw conveyor current change rate.

2. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 1, characterized in that, The sudden water inrush risk model library is constructed as follows: A three-field coupled simulation environment of sandy strata tunnel-tunneling machine-groundwater is established in GTSNX. Training sample sets are generated according to different sandy strata tunnel types, different burial depths, and different geological ages. The sample sets include radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, residual distance, auxiliary variables, and corresponding real labels of sudden water inrush. Bayesian optimization is used to optimize the hyperparameters of the LSTM-Attention architecture. The optimization objective is to minimize the cross-entropy loss between the predicted level and the real label. After optimization, the optimal hyperparameters are solidified into a sudden water inrush risk model for the combination of strata-burial depth-age and added to the library.

3. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 1, characterized in that, include: Using the tunneling cycle as the time window, the instantaneous energy spectrum of the radar and the instantaneous energy spectrum of the microseismic event are discretized according to time periods; The dynamic weighted allocation values ​​of radar instantaneous energy and microseismic instantaneous energy within the same time period are calculated to obtain the coupled energy value; wherein, the dynamic weighted allocation is adjusted according to the stratigraphic heterogeneity characteristics, specifically as follows: Feature extraction was performed on the instantaneous energy spectrum of radar and the instantaneous energy spectrum of microseismic activity. The extracted features included the peak value and frequency distribution of the energy spectrum. The weighting coefficients of radar instantaneous energy and microseismic instantaneous energy in the current time period are calculated based on the extracted features. The weighting coefficients are related to the formation heterogeneity. The higher the formation heterogeneity, the greater the difference in the weighting coefficients. The instantaneous radar energy and the instantaneous microseismic energy are weighted and summed according to the weighting coefficients to obtain the coupled energy value; The risk index of sudden water inrush at the sandy interface is obtained by summing all the coupled energy values ​​within the time window and dividing by the time window length.

4. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 1, characterized in that, After inputting the sudden water inrush risk index and the remaining distance into an adaptive threshold comparator, the dynamic adjustment of the threshold by the adaptive threshold comparator includes: The current tunneling speed value is read in real time in the tunneling machine PLC, and the speed change is obtained by subtracting the tunneling speed value from the previous cycle. If the change in speed is positive and exceeds the set rate, the threshold increases linearly. If the change in speed is negative and its absolute value is higher than the set rate, the threshold decreases exponentially. If the absolute value of the speed change does not exceed the set rate, the threshold remains unchanged; The adjusted threshold is immediately written to the adaptive threshold comparator and takes effect.

5. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 1, characterized in that, The risk avoidance path for the tunneling machine generated based on the adjusted parameters includes: The tunnel axis is discretized into a set of nodes at fixed intervals; For each node, call the inrush water risk model to calculate the maximum inrush water risk level of that node over a certain number of future time periods. If the maximum risk level of sudden water inrush exceeds the preset level, the node will be marked as a risk node. Construct a cost function, which is a weighted sum of the node risk level and the Euclidean distance from the node to the current cutterhead position; The A* algorithm is used to search for the minimum cost path from the current cutterhead position to the end of the tunnel design axis in the node set. The cost of risk nodes is multiplied by a penalty coefficient, which increases monotonically with the risk level. The minimum cost paths are connected in sequence to generate a continuous risk avoidance path, which is then sent to the tunneling machine navigation system for execution.

6. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 2, characterized in that, The sudden water inrush risk model library is constructed in the following ways and also includes: In the GTSNX platform, for each type of sandy stratum tunnel—burial depth—geological age ternary set, a three-dimensional simulation scene is established that fully couples the physical model of the sandy stratum tunnel, the physical model of the tunnel boring machine, and the groundwater seepage field. The set of working conditions covering the parameter space is generated by Latin hypercube sampling. The input of each working condition is the radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, remaining range and auxiliary variables, and the output is the true label of the sudden water inrush. The simulation was run to extract the time of the sudden water inrush, the inrush volume, and the peak pore water pressure to form training samples. Bayesian optimization is used to globally optimize the hyperparameters of the LSTM-Attention architecture, with cross-entropy loss as the optimization objective; The optimal hyperparameters are solidified into the ternary water inrush risk model and stored in the water inrush risk model library with the stratigraphic type-burial depth-geological age as the key. At the same time, the model weights and normalized parameters are saved.

7. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 1, characterized in that, The acquisition of the auxiliary variable sequence, including the cutterhead torque fluctuation index, includes: A torque sensor is installed at the spindle of the cutter head drive motor to continuously collect the raw torque signal of the cutter head at a fixed frequency; The original signal is subjected to moving average filtering to obtain the filtered torque signal; The filtered torque signal is segmented into segments of fixed duration and the variance of each segment is calculated to obtain a short-time variance sequence. The short-time variance sequence is subjected to minimum-maximum normalization to obtain the tool head torque fluctuation index; When the cutterhead torque fluctuation index exceeds the set limit, an early warning signal is triggered and the tunneling cycle number is recorded.

8. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 7, characterized in that, Establishing a three-field coupled simulation environment for sandy strata tunnels, tunnel boring machines, and groundwater in GTSNX includes: A three-dimensional geological model of a tunnel in sandy strata was established in GTSNX and nonlinear elastoplastic constitutive and permeability parameters were assigned. Establish a model of the tunneling machine casing and set the boundary conditions for the cutterhead, screw conveyor, and grouting holes; Establish a groundwater seepage field and set the hydraulic head boundary and permeability coefficient; The three-field model is meshed and a coupled solver is set up. The coupling variables include displacement, pore water pressure and seepage velocity. Retrieving target models from the sudden water inrush risk model database corresponding to the current type, depth, and geological age of tunnels in sandy strata also includes: if no perfectly matching model exists, performing similarity retrieval and transfer learning. Calculate the cosine similarity between the current stratigraphic feature vector and the corresponding stratigraphic feature vectors of all models in the library; The model with the highest similarity is selected as the source model, its attention layer weights are frozen, and only the LSTM layer weights are fine-tuned. The LSTM layer weights are updated online using real-time data collected from the current tunneling section. After the update is completed, the model is fixed and stored in the database.

9. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 8, characterized in that, Online updates of LSTM layer weights using real-time data collected from the current tunneling section include: Construct a sliding time window of fixed length, with a step size of one tunneling cycle; Within the time window, outlier removal and missing value filling are performed on the real-time acquired radar instantaneous energy spectrum, microseismic instantaneous energy spectrum, remaining distance, and auxiliary variables; The preprocessed data is input into the source model of the frozen attention layer to obtain the predicted risk level of sudden water inrush, and the cross-entropy loss is calculated by comparing it with the measured level on site. An adaptive optimizer is used to update the gradient of the LSTM layer weights, and the learning rate adaptively decreases with the sample variance within the time window. Add a resilient weighting term to the loss function to prevent catastrophic forgetting; When the cross-entropy loss of several consecutive time windows is lower than the set threshold and the decrease in loss is lower than the set proportion, the update stops, the current weights are solidified to form a new inrush water risk model, and the original reservoir record is overwritten with the current stratum type-burial depth-geological age as the key.

10. The method for predicting the risk of sudden water inrush in tunnels in sandy strata according to claim 1, characterized in that, The output includes the risk levels of sudden water inrush for multiple future time periods, and also generates construction instructions, including: Map the risk level of sudden water inrush to a risk color code; Risk color codes are overlaid on the tunnel BIM model to form a visual risk cloud map; Based on the risk cloud map, the shield machine control commands are automatically generated. The commands include quantitative values ​​of cutterhead speed, thrust, synchronous grouting pressure and screw conveyor speed. Control commands are sent to the tunnel boring machine's PLC via industrial Ethernet for execution.

Images