Dynamic geological survey system and construction simulation method

Abstract

The invention discloses a dynamic geological survey system and a construction simulation method. The dynamic geological survey system comprises an intelligent sensing and data acquisition module and a multi-source data fusion processing module, the intelligent sensing and data acquisition module comprises a Raman scattering optical fiber strain sensor array, a three-component seismometer monitoring network and a distributed optical fiber pressure sensing subsystem; the Raman scattering optical fiber strain sensor array, the three-component seismometer monitoring network and the distributed optical fiber pressure sensing subsystem form a multi-dimensional sensing network for shield construction geological monitoring. By means of the integrated intelligent sensing and data acquisition module, geological changes in the shield construction process can be monitored. The Raman scattering optical fiber strain sensor array, the three-component seismometer monitoring network and the distributed optical fiber pressure sensing subsystem jointly form a multi-dimensional sensing network, geological strain, microseismic events, pore water pressure and other key information can be captured, and comprehensive geological data support is provided for construction.

Description

Technical Field

[0001] The present invention belongs to the technical field of underground space engineering construction equipment, and in particular relates to a dynamic geological survey system and a construction simulation method used in a shield construction process. Background Art

[0002] In the field of shield construction, the complexity and uncertainty of geological conditions have always been the key factors affecting engineering safety and quality. Traditional geological survey methods often rely on preliminary geological drilling and sampling analysis, which is not only time-consuming and labor-intensive, but also difficult to reflect the geological changes during the construction process in real time. Therefore, in actual construction, accidents such as shield machine jamming and ground collapse caused by geological mutations are often encountered, which brings huge safety hazards and economic losses to the project.

[0003] The patent with publication number CN118965814B discloses an online monitoring system for shield tunnels, which is mainly composed of a real-time data acquisition module. The module can collect the wear status of the shield machine cutter disc and geological condition data in real time based on the cutter disc wear status and geological data. By analyzing these data, the system can analyze the wear pattern when the cutter disc contacts the rock and soil, accurately identify the current construction environment, and obtain real-time geotechnical mechanical parameters and construction conditions accordingly. It can be seen that people are paying more and more attention to the safety of shield construction and have taken many measures.

[0004] In order to overcome the limitations of traditional geological survey methods, dynamic geological survey systems have emerged in recent years with the rapid development of sensor technology and data acquisition technology. Such systems can monitor the changes of geological parameters in the construction process in real time by integrating a variety of intelligent sensor modules, providing a scientific basis for construction decisions. However, the existing dynamic geological survey systems still have many shortcomings in the construction of perception networks and data processing, such as single perception dimension, low data acquisition accuracy, and insufficient intelligence of data processing algorithms, which makes it difficult to meet the comprehensive, accurate and real-time requirements of shield construction for geological data. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art, provide a dynamic geological survey system and construction simulation method, establish a multi-dimensional perception network, and provide geological data support for construction.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] Dynamic geological survey system, including intelligent sensing and data acquisition module and multi-source data fusion processing module;

[0008] The intelligent sensing and data acquisition module includes a Raman scattering optical fiber strain sensor array, a three-component seismometer monitoring network and a distributed optical fiber pressure sensing subsystem, wherein:

[0009] Raman scattering optical fiber strain sensor array: a group of sensor nodes is arranged every 10 meters along the axis of the shield, each group contains 3 orthogonally distributed sensing units, and wavelength division multiplexing technology is used to achieve single optical fiber 64-channel parallel transmission;

[0010] Three-component seismometer monitoring network: A three-dimensional array of 48 highly sensitive seismometers, each with a built-in XYZ three-axis MEMS accelerometer to achieve micron-level displacement monitoring;

[0011] Distributed optical fiber pressure sensing subsystem: a sensing optical fiber is laid every 30° along the circumferential direction of the shield segment;

[0012] The Raman scattering optical fiber strain sensor array, the three-component seismometer monitoring network, and the distributed optical fiber pressure sensing subsystem constitute a multi-dimensional sensing network for geological monitoring of shield construction, which monitors the state of the formation in all directions from the three physical dimensions of strain, vibration, and pressure.

[0013] Multi-protocol data convergence unit: built-in chip realizes synchronous acquisition of multi-source data, and the timestamp alignment error is ≤1μs;

[0014] The multi-source data fusion processing module includes:

[0015] Heterogeneous data processing engine: Equipped with a GPU computing cluster (8 GPUs per node, 320GB video memory), it adopts a streaming computing architecture and supports parallel processing of heterogeneous geological, hydrological, and mechanical data.

[0016] The spatiotemporal attention fusion algorithm acts on the heterogeneous data fusion stage to process the multimodal data of geological strain, microseismic events, and pore water pressure from the intelligent sensing and data acquisition modules; spatiotemporal feature fusion: Spatial dimension: Through the multi-head self-attention mechanism, the cross-modal correlation matrix of geological strain (Q), microseismic events (K), and pore water pressure (V) is calculated to capture the spatial correlation between formation deformation and microseismic activity.

[0017] Temporal dimension: The causal convolutional layer analyzes temporal dependencies and identifies the lag effects of construction parameters and formation responses.

[0018] Furthermore, it also includes a high-precision dynamic construction simulation module, which adopts a hybrid calculation model of finite element method and discrete element method, including:

[0019] Adaptive meshing function, with a minimum cell size of 1 cm;

[0020] The nonlinear constitutive model introduces strain softening and hardening parameters based on the modified Mohr-Coulomb criterion. Furthermore, the parameters of the nonlinear constitutive model are updated in real time through an intelligent optimization algorithm, including:

[0021] Dynamic adjustment range of cohesion c: 0.1kPa-500kPa; friction angle Update formula: φ t+1 =φ t +α·(c p -c threshold ), where α is the learning rate, c p is the current strain value, c threshold is the strain threshold.

[0022] Furthermore, it also includes an intelligent early warning decision module, which includes:

[0023] Multi-level early warning model, covering geological mutation, water pressure anomaly, construction parameter deviation risk indicators;

[0024] Deep reinforcement learning algorithm, based on a dual network architecture:

[0025] The policy network is a three-layer fully connected structure with 512, 256, and 128 neurons respectively;

[0026] The value network consists of two layers of convolution and one layer of full connection, with a convolution kernel size of 5×5 and a stride of 2.

[0027] Furthermore, the training process of the deep reinforcement learning algorithm includes:

[0028] The state space S includes geological parameters, water pressure values, and shield machine thrust;

[0029] The action space A is the adjustment amount of the cutter head speed and propulsion speed;

[0030] The reward function R is designed as:

[0031]

[0032] where w 1 、w 2 is the weight coefficient, the penalty coefficient w 3 .

[0033] Furthermore, the cloud collaborative management module adopts:

[0034] Distributed microservice architecture with storage capacity of 1PB;

[0035] Edge computing unit, the three-level architecture includes terminal layer, edge node layer and cloud layer;

[0036] A lightweight neural network model is deployed at the terminal layer for data preprocessing;

[0037] The edge node layer is equipped with high-performance computing units to implement local data analysis.

[0038] Furthermore, in the three-level architecture of the edge computing unit:

[0039] The terminal layer processor is ARM Cortex-A77, with a main frequency of 2.8GHz;

[0040] The edge node layer is equipped with NVIDIA Jetson AGX Xavier, with a computing power of 32TOPS;

[0041] The cloud layer uses Alibaba Cloud Shenlong server, with a single node memory of 768GB;

[0042] The data classification strategy is: real-time data is retained for ≤24 hours, and historical data is compressed and stored.

[0043] Furthermore, the attention mechanism algorithm realizes multi-source data fusion through dynamic weight adjustment:

[0044] The input feature matrix dimension is [N×D], where N is the number of time steps and D is the feature dimension;

[0045] The output dimension of the multi-head self-attention layer is [N×8×64], and after concatenation, the output is [N×512];

[0046] The cross-modal feature fusion layer uses a bidirectional GRU, and the hidden state update formula is:

[0047] h t =σ(W hz x t +W hh h t-1 +b h )

[0048] Where σ is the Sigmoid activation function, W hz , W hh is the weight matrix.

[0049] A shield construction dynamic geological survey and construction simulation method, comprising the following steps:

[0050] S1. Real-time collection and transmission of geological data:

[0051] An array of optical fiber strain sensors is laid out every 10 meters along the shield tunneling route;

[0052] Deploy 48 three-component seismometers and use an improved spectrum analysis algorithm to extract microseismic signals;

[0053] A distributed fiber-optic pressure sensing subsystem monitors water pressure in real time;

[0054] S2. Multi-source data fusion processing:

[0055] Use NVIDIA A100 GPU cluster to process heterogeneous data in parallel;

[0056] The geological, hydrological and mechanical data are integrated through the attention mechanism algorithm, and the fusion accuracy is required to be ≥95%; the multi-head self-attention layer calculates the feature correlation matrix;

[0057] The cross-modal fusion layer uses Bi-GRU to capture temporal dependencies;

[0058] S3. High-precision construction dynamic simulation:

[0059] Construct a finite element-discrete element hybrid model with a minimum grid size of 1 cm;

[0060] Dynamically update boundary conditions to simulate the coupling effect between stratum and shield machine;

[0061] S4. Intelligent early warning and decision-making:

[0062] Evaluate risk indicators based on deep reinforcement learning models;

[0063] When the prediction error probability P ≥ 0.9, a first-level warning is triggered and pushed to the terminal device through the 5G network;

[0064] S5. Cloud Collaborative Management:

[0065] Use edge computing to process 80% of data locally, reducing bandwidth requirements;

[0066] The visualization platform displays the three-dimensional formation stress field in real time.

[0067] Furthermore, the specific process of improving the attention mechanism algorithm in step S2 includes:

[0068] a. Input feature matrix Generate Q, K, V matrices through linear projection;

[0069] b. Calculate multi-head attention weights:

[0070]

[0071] where d k =64 is the key vector dimension;

[0072] c. Cross-modal fusion layer output:

[0073] h t =GRU(x t ,h t-1 )

[0074] Hidden layer dimension 256, time step N = 100;

[0075] d. The temporal dependency layer uses causal convolution to retain historical information constraints.

[0076] Furthermore, the adaptive grid division method of the hybrid computing model in step S3 includes:

[0077] The mesh encryption area is determined based on the curvature threshold θ = 0.05;

[0078] Dynamically adjust the cell size Δl:

[0079]

[0080] in is the displacement gradient;

[0081] The grid topology is updated every 30 minutes.

[0082] The present invention has at least the following beneficial effects:

[0083] The present invention can monitor geological changes during shield construction through integrated intelligent sensing and data acquisition modules. The Raman scattering fiber strain sensor array, three-component seismometer monitoring network and distributed fiber pressure sensing subsystem together constitute a multi-dimensional sensing network that can capture key information such as geological strain, microseismic events and pore water pressure, providing comprehensive geological data support for construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0085] Figure 1 The overall architecture diagram of the dynamic geological survey and construction simulation system of the present invention;

[0086] Figure 2 It is a schematic diagram of the layout of smart sensors;

[0087] Figure 3 This is the data fusion processing flow chart. DETAILED DESCRIPTION

[0088] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0089] Hereinafter, different embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that the present disclosure is not limited to the specific form disclosed herein, but rather, the present disclosure should be interpreted as covering different changes, equivalents and / or replacements of the embodiments of the present disclosure. When describing the accompanying drawings, similar reference numerals may be used to indicate similar components.

[0090] In the present disclosure, the terms are used to describe specific embodiments and do not limit the present disclosure. As used herein, the singular form is intended to also include the plural form, unless the content clearly indicates otherwise. In the specification, it should be understood that the term "including" or "having" represents the presence of a feature, number, step, operation, structural element, component or combination thereof, without excluding in advance the possibility of the presence or addition of one or more other features, numbers, steps, operations, structural elements, components or combinations thereof.

[0091] It should be understood that certain details are provided in the following description to facilitate a complete understanding of the example embodiments. However, one of ordinary skill in the art should understand that the example embodiments can be implemented without these specific details. For example, the system can be shown in a block diagram to avoid obscuring the example with unnecessary details. In other examples, well-known processes, structures, and techniques may not be shown in unnecessary details to avoid obscuring the example.

[0092] Dynamic geological survey system or shield tunnel detection system, including intelligent sensing and data acquisition module and multi-source data fusion processing module;

[0093] The intelligent sensing and data acquisition module includes a Raman scattering optical fiber strain sensor array, a three-component seismometer monitoring network and a distributed optical fiber pressure sensing subsystem, wherein:

[0094] Raman scattering fiber strain sensor array: a group of sensor nodes are arranged every 10 meters along the shield axis to form a continuous monitoring network. Each group contains 3 orthogonally distributed sensor units. Each sensor unit is arranged along the axial direction (excavation direction), circumferential direction (circumferential direction), and radial direction (vertical to the tunnel wall direction) of the shield tunnel. Axial direction: monitors the longitudinal tensile or compressive strain of the tunnel, such as tunnel deformation caused by stratum settlement; Circumferential direction: captures the deformation of the circumferential direction of the tunnel cross section, such as elliptical deformation caused by uneven soil pressure; Radial direction: detects the extrusion or expansion of the tunnel wall in the inner and outer directions, such as deformation caused by changes in groundwater pressure. The system uses a new generation of high-precision fiber strain sensor arrays for stratum monitoring. Each sensor unit is encapsulated in a titanium alloy shell, which has extremely strong compressive resistance and can withstand an environmental pressure of 200MPa. The sensor uses grating modulation technology, with a measurement accuracy of ±0.1 micron and a sampling frequency of 1000Hz. Wavelength division multiplexing technology is used to achieve single-fiber 64-channel parallel transmission.

[0095] Three-component seismometer monitoring network: It is composed of 48 selected high-sensitivity seismometers in a three-dimensional array. Each seismometer has a built-in XYZ three-axis MEMS acceleration sensor to achieve micron-level displacement monitoring. The seismometers are arranged in an equilateral triangle array, such as Figure 2 As shown in the yellow node in the middle, all-round three-dimensional monitoring of the construction area is achieved. The system is equipped with a PXI-5922 high-performance digital acquisition device from National Instruments, with a sampling rate of up to 10,000 times / second and a resolution of 24 bits, ensuring accurate collection of weak signals.

[0096] Distributed fiber optic pressure sensing subsystem: The groundwater pressure monitoring system uses distributed fiber optic pressure sensing technology, and the core sensing element uses the DPharp series pressure transmitter of Yokogawa Electric Corporation of Japan. The pressure sensor adopts a sapphire diaphragm design, which is highly corrosion-resistant and can maintain stable operation in both acidic and alkaline environments. A sensing fiber is laid every 30° along the circumferential direction of the shield segment; the system lays 3 groups of pressure sensors in each monitoring section, such as Figure 2 As shown in the blue nodes in the middle, a three-dimensional water pressure monitoring network is formed. When a certain line in Chengdu was crossing a fault zone, the system successfully warned of a sudden water inrush event and issued an alarm 15 minutes in advance, buying precious time for the construction workers to evacuate safely.

[0097] The Raman scattering optical fiber strain sensor array, the three-component seismometer monitoring network, and the distributed optical fiber pressure sensing subsystem constitute a multi-dimensional sensing network for geological monitoring of shield construction, which monitors the state of the formation in all directions from the three physical dimensions of strain, vibration, and pressure.

[0098] Multi-protocol data convergence unit: built-in FPGA chip realizes synchronous acquisition of multi-source data, timestamp alignment error ≤1μs, equipped with 5G communication module, working frequency band 3.3-3.8GHz, single base station downlink peak rate can reach 20Gbps. The network coverage radius reaches 500 meters, and a single base station can support 1,000 terminals to access at the same time. The system adopts a dual-channel redundant design, the main channel uses 5G network, and the backup channel uses industrial Ethernet. Millisecond-level switching can be achieved between the two channels. By deploying MEC edge computing nodes, the network delay is controlled within 10 milliseconds and the packet loss rate is less than 0.01%.

[0099] The multi-source data fusion processing module includes:

[0100] Heterogeneous data processing engine: Equipped with NVIDIA A100 GPU computing cluster (8 GPUs per node), it adopts streaming computing architecture and supports parallel processing of geological, hydrological, and mechanical heterogeneous data; the data processing platform adopts streaming computing architecture and is equipped with NVIDIA DGX A100 artificial intelligence computing system, including 8 A100 GPUs. The system uses an improved multi-head attention mechanism algorithm for data fusion, and in practice has achieved real-time fusion analysis of more than 100 different types of sensor data, with a processing delay of less than 50 milliseconds.

[0101] The data preprocessing module uses an adaptive noise reduction algorithm based on wavelet transform, combined with Kalman filtering technology, to effectively remove the influence of environmental noise and improve the signal-to-noise ratio by 15dB. The system performs multi-scale decomposition on the original data and processes the noise coefficient through the soft threshold method to achieve effective reconstruction of the signal.

[0102] The present invention builds a multi-dimensional sensing network for geological monitoring of shield construction by integrating Raman scattering optical fiber strain sensor array, three-component seismometer monitoring network and distributed optical fiber pressure sensing subsystem. The network can conduct all-round and high-precision real-time monitoring of the stratum state from three physical dimensions of strain, vibration and pressure, significantly improving the geological monitoring capability during the construction process.

[0103] Furthermore, the arrangement structure of the Raman scattering optical fiber strain sensor array is axially arranged: a group of sensor nodes is arranged every 10 meters along the shield tunneling axis, and each group contains 3 orthogonally distributed sensor units, which are arranged along the axial direction (tunneling direction), circumferential direction (circumferential direction), and radial direction (vertical to the tunnel wall direction) of the shield tunnel (see Figure 2 , Figure 2 The cross-sectional diagram does not fully display the relevant arrangement). Axial unit: monitors the longitudinal tensile or compressive strain of the tunnel (such as tunnel deformation caused by ground settlement); Circumferential unit: captures the deformation of the tunnel cross section in the circumferential direction (such as elliptical deformation caused by uneven soil pressure); Radial unit: detects the extrusion or expansion of the tunnel wall in the inner and outer directions (such as deformation caused by changes in groundwater pressure). Packaging and performance: The sensor is packaged in a titanium alloy shell with a compressive performance of ≥200MPa, a measurement accuracy of ±0.1 micron, and a sampling frequency of 1000Hz. Data transmission: Wavelength division multiplexing technology is used to achieve parallel transmission of 64 channels on a single optical fiber. Each channel independently modulates the wavelength (range 1525-1565nm) to avoid signal crosstalk.

[0104] Furthermore, the three-component seismometer monitoring network layout structure. Stereo array design: A three-dimensional monitoring network composed of 48 high-sensitivity seismometers covers the entire section of the shield construction area ( Figure 2Yellow nodes in the middle are used as the schematic diagram). Horizontal layer layout: one layer is laid out every 10 meters / 20 meters along the tunnel axis, and each layer contains 6 seismometers arranged in a regular hexagon; vertical layering: the spacing between each layer is 5 meters, with a total of 8 layers, forming a three-dimensional monitoring grid; sensor configuration: each seismometer has a built-in XYZ three-axis MEMS acceleration sensor, with a displacement resolution of 1 micron and a frequency response range of 0.1-200Hz. Signal acquisition: using NIPXI-5922 digital collector, with a sampling rate of 10,000 times / second, a dynamic range of 120dB, and supports microseismic signals (energy ≤10 -6 J)’s precise capture.

[0105] Layout structure of distributed optical fiber pressure sensing subsystem: Circumferential layout: a sensing optical fiber is laid out every 30° along the circumferential direction of the shield segment, and 3 groups of pressure sensors are laid out in each monitoring section ( Figure 2 The blue nodes in the middle are schematically shown) to form a three-dimensional water pressure monitoring network. Fiber type: Sapphire diaphragm pressure sensing fiber is used, with corrosion resistance up to IP68, measurement range 0-10MPa, accuracy ±0.01MPa; Redundancy design: Each set of sensors is equipped with dual-channel optical fiber, which automatically switches to the backup channel when the main channel fails, with reliability ≥99.99%. Installation method: Soft soil layer: The optical fiber is embedded in the outer buffer layer of the shield segment (thickness 50mm) to avoid mechanical damage; Hard rock layer: The optical fiber is covered with a stainless steel armored sheath, with impact resistance ≥50J.

[0106] The collaborative mechanism of the multi-dimensional sensing network, data synchronization: the timestamp alignment of the three-system data is achieved through the FPGA chip of the multi-protocol data convergence unit (error ≤ 1μs); spatial coverage density: strain monitoring: axial density 10 meters / group, circumferential density 30° / line; microseismic monitoring: horizontal density 10 meters or 20 meters / layer, vertical density 5 meters / layer; water pressure monitoring: circumferential density 30° / line, radial density 3 groups / section. Electromagnetic shielding optical fiber (shielding effectiveness ≥ 60dB) and grounding isolation technology are used to suppress electromagnetic interference of the shield machine; the pressure sensor has a built-in temperature compensation module (range -20℃~80℃) to eliminate the influence of environmental temperature drift.

[0107] The layout of the Raman scattering optical fiber strain sensor array ensures that the deformation of the tunnel in all directions can be accurately captured. Its high precision and anti-interference ability provide a strong guarantee for safety monitoring during the construction process. The three-component seismometer monitoring network realizes all-round three-dimensional monitoring of the construction area, and can accurately capture weak signals such as seismic waves, providing a reliable basis for geological early warning. The distributed optical fiber pressure sensing subsystem effectively monitors the changes in groundwater pressure and provides important data support for preventing geological disasters such as sudden water inrush.

[0108] The multi-protocol data convergence unit realizes the synchronous collection and efficient transmission of multi-source data, ensuring the real-time and accuracy of the data. At the same time, the application of 5G communication modules also improves the speed and stability of data transmission, providing strong support for remote monitoring and real-time data analysis. The dual-channel redundant design further improves the reliability and stability of the system, ensuring continuous monitoring during the construction process.

[0109] The multi-source data fusion processing module realizes the real-time fusion analysis of geological, hydrological and mechanical heterogeneous data by carrying high-performance computing clusters. The data preprocessing module effectively removes the influence of environmental noise and improves the signal-to-noise ratio and accuracy of the data. The application of these technical means enables the system to achieve accurate prediction and early warning of the formation state, providing a strong guarantee for the construction safety of shield tunnels.

[0110] The spatiotemporal attention fusion algorithm acts on the heterogeneous data fusion stage, processing the multimodal data of geological strain, microseismic events, and pore water pressure from the intelligent sensing and data acquisition modules; spatiotemporal feature fusion: Spatial dimension: Through the multi-head self-attention mechanism, the cross-modal correlation matrix of geological strain (Q'), microseismic events (K'), and pore water pressure (V') is calculated to capture the spatial correlation between formation deformation and microseismic activity. Example: When the annular strain increases suddenly, the microseismic events (such as fault activation) and water pressure anomaly areas related to the location are automatically focused.

[0111] The spatiotemporal attention fusion algorithm is a multimodal data fusion method that combines spatial correlation and temporal dependency. Its core consists of two parts: multi-head self-attention mechanism (spatial dimension) and causal convolution + bidirectional GRU (temporal dimension). Input data → spatial attention (cross-modal correlation) → temporal convolution (temporal modeling) → bidirectional GRU (feature fusion) → output fusion features.

[0112] The built-in chip of the multi-protocol data convergence unit realizes the synchronous collection of multi-source data, ensuring the timeliness and accuracy of the data. The system can efficiently process data from different sensors and monitoring networks, avoiding data delay or loss.

[0113] The heterogeneous data processing engine in the multi-source data fusion processing module is equipped with a GPU computing cluster and adopts a streaming computing architecture, which can efficiently perform parallel processing of heterogeneous data such as geology, hydrology, mechanics, etc. This not only improves the speed of data processing, but also enhances the system's data processing capabilities, enabling it to cope with large-scale and complex geological data.

[0114] The spatiotemporal attention fusion algorithm plays an important role in the heterogeneous data fusion stage. The algorithm can intelligently process multimodal data such as geological strain, microseismic events, pore water pressure, etc. from the intelligent sensing and data acquisition module, and provide more accurate and comprehensive geological survey results for shield construction by fusing this information. This helps the construction unit to take necessary measures in a timely manner to ensure the safety and smooth progress of the construction.

[0115] Spatial dimension: Multi-head self-attention mechanism. Goal: Capture the spatial correlation between geological strain, microseismic events, and pore water pressure (for example: whether a sudden increase in strain in a certain area is related to a nearby microseismic event). Implementation steps: Input feature matrix: geological strain (Q'), microseismic events (K'), pore water pressure (V') data are aligned according to spatial position to form the input matrix Where N is the number of time steps and D is the feature dimension (eg, D=256).

[0116] Linear projection generates Q / K / V: through the weight matrix W Q ,W K ,W V Map the input to the query (Q), key (K), value (V) space: Q = XW Q ,K=XW K ,V=XW V .

[0117] Multi-head attention calculation: 8 parallel attention heads are used (each head has dimension d k =64), calculate the cross-modal correlation matrix:

[0118] Temperature coefficient d k : Prevent gradient explosion or disappearance and make Softmax weight distribution more stable.

[0119] Softmax: Normalizes the relevance weights to generate attention distribution (weight range 0-1).

[0120] Output concatenation: The 8-head attention results are concatenated and reduced to R N×512 .

[0121] QK T : Calculate the cross-modal correlation between geological strain and microseismic events (such as the strong correlation between sudden strain increases in fault zones and adjacent microseisms).

[0122] Spatial correlation extraction: When the annular strain increases suddenly, the system automatically focuses on microseismic events (such as fault activation) and water pressure anomalies (such as sudden increase in pore water pressure) in related areas.

[0123] Fusion feature output: Perform weighted summation of V according to the QK correlation weights to extract key abnormal signals (such as abnormal water pressure areas before sudden water surges).

[0124] In the spatiotemporal attention fusion algorithm, the causal convolution layer is the core module used to model the temporal dependency between construction parameters and formation responses. Its core goal is to capture the hysteresis effect between the adjustment of operating parameters (such as cutterhead speed and grouting pressure) and formation responses (such as surface settlement and stress redistribution) during shield construction, thereby providing a basis for real-time warning and construction optimization. Causal convolution is a one-way time series modeling method, which is characterized by using only historical information (data at the current moment and before) for calculation to avoid future data leakage.

[0125] In shield construction, there is often a time lag in the response of the stratum. For example, the cutterhead speed increases → the stratum disturbance intensifies → the surface settlement accelerates after 15 minutes; the grouting pressure is adjusted → the pore water pressure changes → the stress field is redistributed after 30 minutes. Causal convolution quantifies such lag relationships through multi-scale time series feature extraction. The specific implementation is as follows: (1) Parameter configuration, number of convolution kernels: 33, covering the long-term historical window; expansion rate: d = 2, gradually expanding the receptive field to 2 n (n is the number of layers); step size: 1, to ensure that the time resolution is not reduced. (2) Hysteresis effect modeling process, input data: time series of construction parameters (cutterhead speed, advancement speed) and formation response (surface settlement, pore water pressure); convolution operation: extract features of different time spans through dilated causal convolution; feature fusion: combine the self-attention mechanism to locate the key lag interval (such as 15-30 minutes after the speed adjustment).

[0126] Bidirectional gated feature fusion layer: Bi-GRU network (hidden dimension 256, forget gate initial bias 0.7) is used to fuse spatiotemporal features and extract 128-dimensional high-order feature vectors; Bidirectional GRU network: It consists of two GRUs, forward and backward, which process sequence data from the forward and reverse time directions respectively. Hidden dimension 256: The hidden state dimension of each GRU is 256, which is used to capture long-term and short-term dependencies. Forget gate initial bias 0.7: The "forget gate" here is actually the update gate of GRU, and the bias is initialized to 0.7, which aims to control the retention ratio of historical information (such as retaining 70% of historical states and updating 30% of new inputs). The hidden states of the forward and backward GRUs are concatenated and reduced to 128 dimensions through a fully connected layer to extract high-order spatiotemporal features. Bi-GRU associates the delayed response of formation strain during shield advancement (such as formation deformation 30 minutes after the cutterhead speed is adjusted) with the spatial distribution of microseismic events through bidirectional processing, thereby enhancing the spatiotemporal consistency of risk prediction. The initial bias of 0.7 is designed to filter out high-frequency mechanical vibration noise.

[0127] Causal convolutional temporal modeling layer: Configure a 3×3 convolution kernel (step size 1, dilation rate 2) to build a deep separable convolutional network to capture the long-term dependencies of the shield advancement process; use dilated causal convolution to ensure that the output only depends on current and historical inputs to avoid future data leakage.

[0128] Real-time data fusion service: data fusion accuracy ≥ 95%, processing delay ≤ 50 milliseconds, and output multi-dimensional fusion feature tensor (dimension: time step × spatial coordinate × 32 channels). Through the collaboration of multi-head self-attention mechanism and Bi-GRU, the spatiotemporal deviations of multimodal data (geological strain, microseismic, water pressure) are effectively aligned. Relying on the streaming computing architecture of GPU cluster (8×NVIDIA A100), end-to-end real-time processing is achieved. Each spatiotemporal position contains 32-dimensional fusion features, which can be used for downstream tasks (such as risk prediction, construction simulation). This embodiment realizes data alignment: through timestamp synchronization (error ≤ 1μs) and spatial interpolation, the difference in sensor sampling rate is solved. Feature fusion: multi-head self-attention layer calculates cross-modal correlation weights (Softmax normalization); Bi-GRU and causal convolution respectively extract spatiotemporal dependency features; feature splicing is reduced to 32 channels. Combining deep learning models (Bi-GRU, causal convolution) with engineering monitoring needs, high-precision real-time fusion of multimodal data is achieved, providing core technical support for shield construction under complex geological conditions. For relevant procedures, see Figure 3 .

[0129] Specifically, the multi-source data fusion processing module includes:

[0130] NVIDIA A100 GPU computing cluster, with a daily data processing capacity of 100TB;

[0131] Improved attention mechanism algorithm, including multi-head self-attention layer, cross-modal feature fusion layer and temporal dependency modeling layer;

[0132] The multi-head self-attention layer uses 8 attention heads, and the dimension of each attention head is 64;

[0133] The cross-modal feature fusion layer uses a bidirectional gated recurrent unit network (Bi-GRU) with a hidden layer dimension of 256;

[0134] The temporal dependency modeling layer uses a causal convolutional network with a convolution kernel size of 3×3 and a step size of 1;

[0135] Data fusion accuracy ≥ 95%, processing delay ≤ 100 milliseconds.

[0136] In one embodiment, a high-precision dynamic construction simulation module is further included. The high-precision dynamic construction simulation module adopts a hybrid calculation model of finite element method and discrete element method, including:

[0137] The meshing module uses adaptive octree meshing technology to automatically encrypt meshes in key areas such as around the cutterhead of the shield machine, with a minimum unit size of 1 cm. The system automatically determines the mesh encryption area based on curvature and displacement gradient, ensuring the optimal balance between calculation accuracy and efficiency.

[0138] Adaptive octree meshing technology: Octree mesh generation, with the shield machine cutterhead as the center, the initial mesh size is 10cm, and local encryption is achieved through recursive subdivision (up to 4 layers), with the minimum unit size reaching 1cm. The subdivision conditions include: curvature threshold: when the grid unit curvature θ ≥ 0.05, the subdivision is triggered; displacement gradient threshold: if Dynamic adjustment mechanism: The grid topology is updated every 30 minutes based on real-time monitoring data (such as formation stress and displacement), with priority given to high-risk areas (such as fault zones and water-rich layers).

[0139] The specific process includes:

[0140] Initial coarse grid generation: covering the shield construction influence area (radius 50m);

[0141] Local curvature calculation: Based on real-time displacement field data, the unit curvature is calculated using the second-order difference method;

[0142] Subdivision judgment: If the curvature or displacement gradient threshold is met, octree subdivision is performed;

[0143] Merge optimization: Merge adjacent grids in stable areas (curvature < 0.03 and gradient < 0.05) to improve computational efficiency.

[0144] In the dynamic simulation of shield construction, meshing is the core step of discretizing the continuous geological body or structure into a finite number of small units for numerical calculation.

[0145] The strata around the cutterhead of the shield machine are prone to high stress concentration (such as shear bands and crack expansion) due to mechanical disturbance. By encrypting the grid to 1 cm, the local stress distribution can be accurately simulated (such as stress peak error ≤ 5%). The grid is automatically subdivided to ensure the calculation accuracy of high gradient areas (such as fault activation areas). Through recursive subdivision and merging, the grid topology is dynamically adjusted to avoid the waste of resources of traditional uniform grids. With traditional methods, uniform grids require fine division of the entire model, which increases the amount of calculation and cannot dynamically respond to geological changes.

[0146] The grid is updated every 30 minutes based on the fiber optic sensor data (strain, displacement) to adapt to dynamic changes in the formation (such as stress redistribution after grouting). Boundary condition coupling: Viscoelastic artificial boundaries and adaptive grids work together to eliminate false wave reflections and improve the accuracy of dynamic response simulation. Macro-micro coupling: Coarse grids simulate overall formation deformation (such as surface settlement), and fine grids characterize micromechanical behaviors (such as friction between particles and crack expansion). Multi-physics field coupling: Through the hybrid finite element-discrete element model, the interaction between continuous media (soil layer stress) and discontinuous media (rock block crushing) is simulated simultaneously.

[0147] The nonlinear constitutive model introduces strain softening and hardening parameters based on the modified Mohr-Coulomb criterion. The nonlinear constitutive model is a mathematical model that describes the mechanical behavior of materials under complex stress states, and its stress-strain relationship does not satisfy the linear proportion. In shield construction, the mechanical properties of stratum materials (such as soil and rock) will change dynamically with construction disturbances, and traditional linear models cannot accurately simulate them. The modified Mohr-Coulomb criterion introduces strain softening and hardening mechanisms on the basis of the classical model to dynamically reflect the strength changes of the stratum at different deformation stages: strain softening: the strength of the material gradually decreases after reaching the yield point (such as the decrease in cohesion after shear failure of the soil); strain hardening: the strength of the material increases due to deformation and compaction (such as the increase in friction angle after compression of sand).

[0148] The parameters of the nonlinear constitutive model are updated in real time by the Adam optimization algorithm (learning rate = 0.001), including:

[0149] Dynamic adjustment range of cohesion c: 0.1kPa (sand) - 500kPa (hard rock), based on real-time strain energy density feedback;

[0150] Friction angle Update formula: φ t+1 =φ t +α·(c p -c threshold )

[0151] Where α is the learning rate, c p is the current strain value (collected by the optical fiber sensor), c threshold is the strain threshold (corresponding to the formation yield point).

[0152] Strain softening of cohesion c: When strain ε ≥ ε yield When , the cohesion decays exponentially:

[0153]

[0154] c 0 : Initial cohesion (range 0.1kPa–500kPa)

[0155] β=0.05: Softening coefficient, controls the attenuation rate.

[0156] ε yield is the yield strain of the material, which indicates the critical strain value at which the material transitions from the elastic deformation stage to the plastic deformation stage. yield Distinguish between elastic and plastic stages, and accurately simulate the progressive destruction process of the strata during shield construction (such as the formation of soil shear bands and rock strata crushing). Combined with the strain data monitored in real time by the optical fiber sensor, the constitutive model parameters are dynamically adjusted to improve the prediction accuracy.

[0157] Strain hardening of the friction angle φ: The friction angle changes with the plastic strain ε p Linear growth:

[0158] φ=φ 0 +α·ε p

[0159] φ 0 : Initial friction angle (range 15°–45°); α value is 0.3: Hardening rate, which determines the speed of strength improvement.

[0160] Update mechanism: Based on real-time fiber optic sensor data (strain, displacement), the model parameters are adjusted every 5 minutes through the Adam optimizer (learning rate = 0.001);

[0161] Constraints: cohesion c∈[0.1kPa,500kPa], friction φ∈[15°,45°].

[0162] (3) Modified Mohr-Coulomb criterion

[0163] τ=c+σ n tanφ

[0164] τ: shear strength; σ n :Normal stress

[0165] Dynamic correction: c and φ change in real time with strain, breaking through the static assumptions of traditional models.

[0166] The Adam optimization algorithm is used to update the parameters of the nonlinear constitutive model (cohesion c and friction angle φ) in real time to adapt it to the dynamic changes in the mechanical properties of the strata during shield construction. Data-driven optimization: Through real-time collection of strain, water pressure and other data (such as fiber optic sensor monitoring values), the model parameters are dynamically corrected to solve the defect that the traditional static model cannot reflect the actual working conditions. The intelligent sensing module (optical fiber, seismometer, pressure sensor) collects stratum data in real time as the input of the Adam algorithm. The updated constitutive model parameters are input into the finite element-discrete element hybrid model to dynamically simulate the interaction between the stratum and the shield machine (such as stress field and displacement field). Adaptive meshing (minimum 1cm) is combined with dynamic parameters to accurately capture local high gradient phenomena (such as fault fissure expansion). The simulation results are input into the deep reinforcement learning model to evaluate risk indicators (such as the probability of sudden water inrush P≥0.9) and trigger early warnings.

[0167] In one embodiment, an intelligent early warning decision module is also included, and the intelligent early warning decision module includes:

[0168] The multi-level early warning model covers risk indicators such as geological mutation, water pressure anomaly, and construction parameter deviation; the early warning information is released using a multi-level distribution mechanism, supporting SMS, mobile application push, sound and light alarm, etc. The system sets different levels of early warning thresholds and automatically adjusts the release scope and frequency of early warning information according to the risk level.

[0169] Classification of risk indicators: covers risk indicators such as geological mutations (such as fault activation, karst collapse), abnormal water pressure (such as precursors of sudden water surges), and deviations from construction parameters (excessive cutter torque, abnormal advancement speed). Three levels of warning thresholds are set for each type of indicator.

[0170] Level 1 warning (red): risk probability P ≥ 0.9, triggering immediate shutdown and emergency response;

[0171] Level 2 warning (orange): 0.7≤P<0.9, speed limit construction and start reinforcement measures;

[0172] Level 3 warning (yellow): 0.5≤P<0.7), manual review and adjustment of parameters.

[0173] Deep reinforcement learning algorithm, based on a dual network architecture:

[0174] The strategy network is a three-layer fully connected structure with 512, 256, and 128 neurons respectively; the three-layer fully connected network (512-256-128 neurons), the activation function is ReLU, and the output layer uses Tanh to constrain the action range. Input: The state space S includes geological parameters (stratum stress σ, pore water pressure P), shield machine state (cutter head torque T, propulsion speed v); Output: The action space A is the cutter head speed adjustment Δn∈[-0.5,0.5]rpm, the propulsion speed adjustment Δv∈[-2,2]mm / min.

[0175] The value network is a two-layer convolution and one-layer fully connected structure, with a convolution kernel size of 5×5 and a step size of 2. Input layer → convolution layer (5×5 kernel, step size 2) → fully connected layer (256 neurons) → output layer; evaluate the Q value of the state-action pair to guide the optimization direction of the policy network. In deep reinforcement learning, the Q value is a core indicator used to quantify "the cumulative reward expected to be obtained after performing an action in a given state". The training process of the deep reinforcement learning algorithm includes:

[0176] The state space S includes geological parameters, water pressure values, and shield machine thrust;

[0177] The action space A is the adjustment amount of the cutter head speed and propulsion speed;

[0178] The reward function R is designed as:

[0179]

[0180] where w 1 、w 2 is the weight coefficient, weight w 1 =0.6, w 2 =0.4.

[0181] Penalty coefficient w 3 =0.2.

[0182] Δs: surface settlement deviation (mm); v: propulsion speed (mm / min); |T-Topt|: cutter head torque deviation from the optimal value (kN·m).

[0183] PPO is used for optimization, with a learning rate of 0.0003 and a discount factor γ = 0.99.

[0184] In one embodiment, the cloud collaborative management module adopts:

[0185] Distributed microservice architecture with storage capacity up to 1PB;

[0186] Edge computing unit, the three-level architecture includes terminal layer, edge node layer and cloud layer;

[0187] A lightweight neural network model is deployed at the terminal layer for data preprocessing and feature extraction;

[0188] The edge node layer is equipped with high-performance computing units to achieve local data analysis and decision-making;

[0189] The cloud layer is responsible for global optimization and model updating.

[0190] The cloud-based collaborative management module is the key support for efficient processing of geological data. Characteristics of geological data: During shield construction, Raman optical fiber, seismometers, pressure sensors and other equipment can generate TB-level data (such as strain data sampled at 1000Hz, 10000Hz microseismic signals). The cloud-based collaborative management module supports long-term storage of high-resolution geological data (such as 1 year of historical data). Data processing, analysis, and storage are split into independent microservices (such as data cleaning and feature extraction) to improve system scalability and fault tolerance. Three-level architecture of edge computing: Terminal layer (data preprocessing): Lightweight model: Filter noise (such as mechanical vibration) and extract key features (such as a sudden increase of 500με in strain) at the sensor end; Reduce transmission burden: Only upload feature vectors (not raw waveform data). Edge node layer (local decision-making): High-performance computing: Analyze the deformation trend of the formation (LSTM prediction error ≤ 3mm), trigger local warnings (such as identification of precursors of sudden water inrush); Cloud layer (global optimization): Model training and updating: Aggregate data from multiple construction sites, train a global reinforcement learning model (PPO algorithm), and synchronize to edge nodes every 6 hours; Cross-project knowledge transfer: Water-rich layer construction experience can be quickly adapted to new projects. Operations such as shield machine cutter head speed adjustment and grouting pressure control can respond to formation changes. Through layered processing, the system completes the closed loop from data collection to warning issuance in a short time (such as surface settlement exceeding the limit and cutter head torque abnormality). Edge computing reduces cloud dependence, and the response speed of dynamic adjustment of shield parameters is improved; the cloud global model is continuously optimized to support rapid adaptation of complex geological conditions (such as karst and faults).

[0191] Preferably, in the three-level architecture of the edge computing unit:

[0192] The terminal layer processor is ARM Cortex-A77, with a main frequency of 2.8GHz;

[0193] The edge node layer is equipped with NVIDIA Jetson AGX Xavier, with a computing power of 32TOPS;

[0194] The cloud layer uses Alibaba Cloud Shenlong server, with a single node memory of 768GB;

[0195] The data classification strategy is: real-time data is retained for ≤24 hours, and historical data is compressed and stored.

[0196] In one embodiment, the attention mechanism algorithm realizes multi-source data fusion through dynamic weight adjustment: the input feature matrix dimension is [N×D], where N is the number of time steps, N=100 (100 sampling points within 10 seconds), and D is the feature dimension, D=256 (covering 256-dimensional features such as geological strain, microseismic energy, and water pressure value);

[0197] The output dimension of the multi-head self-attention layer is [N×8×64], and after concatenation, the output is [N×512];

[0198] The cross-modal feature fusion layer uses a bidirectional GRU. The forward GRU captures historical dependencies, and the backward GRU predicts future trends. The hidden state update formula is:

[0199] h t =σ(W hz x t +W hh h t-1 +b h )

[0200] Where σ is the Sigmoid activation function, W hz is the input weight matrix, W hh is the hidden state weight matrix, x t is the input of the current time step, h t-1 is the hidden state of the previous time step, b h is the hidden state bias term.

[0201] Time series feature fusion: Through W hh h t-1 Retain historical information to capture the lag effect of ground response during shield construction (such as surface settlement 30 minutes after grouting pressure adjustment). Multimodal data integration: Input x t It contains the fusion features of geological strain, microseismic events, and water pressure, and realizes the dynamic association of cross-modal information. Noise resistance: The Sigmoid function suppresses noise interference (such as mechanical vibration).

[0202] The spatiotemporal attention features are integrated to extract high-order feature vectors (128 dimensions) for downstream risk prediction and construction parameter optimization.

[0203] Data: Geological strain: axial / circumferential / radial strain collected in real time by fiber optic sensors; microseismic events: vibration energy and temporal and spatial distribution monitored by seismometers; pore water pressure: water pressure change data of distributed fiber optic pressure sensing subsystem.

[0204] Cross-modal correlation of multi-head self-attention layer. Multi-head structure: 8 parallel attention heads (each head dimension 64), calculating the cross-modal correlation matrix of geological strain (Q), microseismic events (K), and pore water pressure (V):

[0205]

[0206] Q (geological strain): As an active query signal, it captures key areas of stratum deformation (such as a sudden increase of 500με in hoop strain);

[0207] K (microseismic events): Correlate matching signals to locate stratum activity (such as microseismic events caused by fault activation);

[0208] V (pore water pressure): information-carrying signal, weighted output of abnormal water pressure areas (such as precursors to sudden water inrush).

[0209] Dynamic weight: Attention weights are generated through Softmax normalization to quantify the spatial correlation strength of different modal data (for example, the strong correlation weight between a sudden increase in strain in a certain area and nearby microseisms is 0.9).

[0210] The present invention dynamically aligns multi-source data through multi-head self-attention weights to eliminate the timing misalignment caused by sampling rate differences. The attention mechanism focuses on key areas (such as fault zones), suppresses noise interference (such as mechanical vibration), and improves fusion accuracy. The GPU-accelerated streaming computing architecture (NVIDIA A100) works in conjunction with the lightweight GRU model to achieve end-to-end processing. The fused high-order features (128 dimensions) are input into the finite element-discrete element hybrid model, and the grid and boundary conditions are dynamically updated to improve simulation accuracy.

[0211] In one embodiment, an intelligent adjustment device is also included, including: a core controller: an industrial-grade ARM processor with a main frequency of 3.0GHz, 16 cores and 32 threads; an actuator: including a servo motor group, a hydraulic drive and a precision displacement platform; a feedback unit: including a high-precision displacement sensor, a force sensor and an angle encoder to achieve real-time intelligent adjustment of the cutter head speed, thrust and posture of the shield machine. The core controller calls the deep reinforcement learning model (strategy network) based on the feature vector to generate action instructions (such as the cutter head speed ± 0.2rpm). For example, when the formation strain suddenly increases, the controller reduces the speed to reduce the disturbance, and verifies the adjustment effect through the feedback unit. The prediction results provided by the construction simulation module (such as the formation stress field distribution) are input into the core controller to optimize the control parameters. If the simulation predicts that the cutter head torque exceeds the limit (error ≤ 5%), the controller adjusts the propulsion speed in advance to avoid equipment overload. When the early warning module triggers a first-level warning (P ≥ 0.9), the controller immediately executes the emergency stop command and starts grouting reinforcement through the hydraulic drive. The displacement sensor monitors the formation deformation after grouting, and the data is transmitted back to the early warning module to update the risk assessment model. Controller status data (such as speed and thrust) are uploaded to the cloud in real time to support global construction progress visualization. The cloud sends optimized reinforcement learning model parameters every 6 hours to improve the adaptability of local control strategies. The intelligent adjustment device is the "nerve endings" of the dynamic geological survey system. Through the closed-loop architecture of the core controller-actuator-feedback unit, it seamlessly connects data analysis, risk warning and construction operations.

[0212] A shield construction dynamic geological survey and construction simulation method, comprising the following steps:

[0213] S1. Real-time collection and transmission of geological data:

[0214] An optical fiber strain sensor array is deployed every 10 meters along the shield tunneling route, with a sampling frequency of 1000 Hz, and data is transmitted through a 5G private network;

[0215] 48 three-component seismometers were deployed, and an improved spectrum analysis algorithm was used to extract microseismic signals with a displacement resolution of 1 micron;

[0216] The distributed optical fiber pressure sensing subsystem monitors water pressure in real time with an accuracy of ±0.01MPa;

[0217] S2. Multi-source data fusion processing:

[0218] NVIDIA A100 GPU cluster is used to process heterogeneous data in parallel, with a daily processing capacity of up to 100TB;

[0219] The attention mechanism algorithm is used to integrate geological, hydrological and mechanical data to improve the fusion accuracy;

[0220] The multi-head self-attention layer calculates the feature correlation matrix;

[0221] The cross-modal fusion layer uses Bi-GRU to capture temporal dependencies;

[0222] S3. High-precision construction dynamic simulation:

[0223] Construct a finite element-discrete element hybrid model with a minimum grid size of 1 cm;

[0224] Dynamically update boundary conditions to simulate the coupling effect of stratum and shield machine; the boundary condition processing module adopts intelligent dynamic boundary technology to update the boundary state of the calculation domain in real time according to the field monitoring data. The system automatically identifies the boundary type, including displacement boundary, force boundary and mixed boundary, and uses viscoelastic artificial boundary model to eliminate false wave reflection.

[0225] S4. Intelligent early warning and decision-making:

[0226] Evaluate risk indicators based on deep reinforcement learning models;

[0227] When the prediction error probability P ≥ 0.9, a first-level warning is triggered and pushed to the terminal device through the 5G network;

[0228] S5. Cloud Collaborative Management:

[0229] Use edge computing to process 80% of data locally, reducing bandwidth requirements;

[0230] The visualization platform displays the three-dimensional formation stress field in real time.

[0231] Figure 1 The overall architecture diagram of the dynamic geological survey and construction simulation system of the present invention is shown.

[0232] Preferably, the specific process of the attention mechanism algorithm in step S2 includes:

[0233] a. Input feature matrix Generate Q, K, V matrices through linear projection;

[0234] b. Calculate multi-head attention weights:

[0235]

[0236] where d k =64 is the key vector dimension;

[0237] c. Cross-modal fusion layer output:

[0238] h t =GRU(x t ,h t-1 )

[0239] Hidden layer dimension 256, time step N = 100;

[0240] d. The temporal dependency layer uses causal convolution to retain historical information constraints.

[0241] Furthermore, the adaptive grid division method of the hybrid computing model in step S3 includes:

[0242] Based on engineering experience, the mesh encryption area is determined according to the curvature threshold θ = 0.05;

[0243] Dynamically adjust the cell size Δl:

[0244]

[0245] in is the displacement gradient;

[0246] The grid topology is updated every 30 minutes based on the curvature threshold (θ = 0.05) and the displacement gradient Refine the mesh.

[0247] Furthermore, the training method of the deep reinforcement learning model in step S4 includes:

[0248] The state space S contains: formation stress σ (MPa), water pressure P (MPa), cutterhead torque T (kN·m);

[0249] The action space A is: propulsion speed v∈[10,20]mm / min, cutter head speed n∈[1,3]rpm;

[0250] Reward function R calculation:

[0251]

[0252] Where Δs is the settlement deviation (mm), T opt is the optimal torque;

[0253] The PPO algorithm is used to update the policy network, with a learning rate of 0.0003, the target network is updated every 1000 iterations, and the experience replay pool capacity is 1×10^6.

[0254] Furthermore, the data processing strategy of edge computing in step S5 includes:

[0255] The terminal layer performs data downsampling, reducing the sampling rate from 1000Hz to 200Hz;

[0256] The edge node layer runs a lightweight LSTM model to predict the deformation trend of the stratum;

[0257] The cloud layer synchronizes global model parameters every 6 hours;

[0258] That is, the edge computing data flow: terminal layer → edge node (LSTM prediction) → cloud (global model update, cycle 6 hours).

[0259] Data tiered storage strategy: The data type is real-time monitoring, the retention period is ≤24 hours, and the storage medium is NVMeSSD; the data type is historical data, the retention period is ≥1 year, and the storage medium is HDD array.

[0260] In a preferred embodiment, the risk probability P is obtained by dynamically evaluating the deep reinforcement learning model in combination with real-time monitoring data and historical risk patterns.

[0261] Input state space (S)

[0262] The evaluation of risk probability P is based on multidimensional state parameters, including:

[0263] Geological parameters: formation stress, strain, and microseismic energy;

[0264] Hydrological parameters: pore water pressure, permeability;

[0265] Shield machine status: cutter head torque, propulsion speed, and thrust.

[0266] Deep reinforcement learning model architecture

[0267] Strategy network: Structure: 3-layer fully connected network (512-256-128 neurons), output action space (such as cutter head speed adjustment, propulsion speed adjustment); generate control strategy to guide shield machine operation.

[0268] Value network: Structure: 2 layers of convolution (5×5 kernel, stride 2) + 1 layer of full connection (256 neurons), outputting the value of the state-action pair.

[0269] Calculation process of risk probability P

[0270] State feature extraction: The spatiotemporal attention fusion algorithm extracts a 128-dimensional high-order feature vector (including the correlation weights of formation deformation, microseismic activity, and water pressure anomaly); the Bi-GRU network is used to capture temporal dependencies (such as the lag effect of surface settlement 30 minutes after grouting pressure adjustment).

[0271] Q-value evaluation: The value network calculates the Q-value of the current state-action pair, reflecting the expectation of future cumulative rewards: The Q value is mapped to the risk probability P through the Softmax function:

[0272]

[0273] When P ≥ 0.9, it is judged as a high-risk event (such as sudden water inrush, ground collapse), triggering a level 1 warning.

[0274] The reward function R is used to quantify the balance between construction safety and efficiency. The corrected formula is:

[0275]

[0276] The PPO algorithm (Proximal Policy Optimization) is used to update the policy network parameters in real time to adapt to complex geological conditions. The risk probability P uses a deep reinforcement learning model to perform spatiotemporal correlation analysis on multi-source data, combining Q-value evaluation with reward function optimization to dynamically quantify construction risks.

[0277] The present invention adopts a layered architecture design, including a perception layer, a data layer, an analysis layer, a control layer, and an application layer. The perception layer is composed of a distributed sensor network to achieve all-round monitoring of the geological environment; the data layer is responsible for data collection, cleaning, and storage; the analysis layer performs data fusion and model calculation; the control layer realizes intelligent regulation of the shield machine; and the application layer provides a visual interface and remote management functions. The various levels of the system are seamlessly connected through standardized interface protocols to ensure the real-time and reliability of data flow.

[0278] Take a subway tunnel project in Chengdu as an example. The maximum burial depth of the project is 57.1 meters, and the geological conditions are extremely complex. The shield tunneling construction faces huge challenges due to the large excavation surface, the large burial depth of the tunnel after entering the main urban area, and the frequent intersection with the existing lines. By deploying this system, intelligent monitoring of the entire construction process is realized. The system has deployed 228 sets of fiber optic strain sensors and 48 three-component seismometers to build a complete monitoring network. During the construction process, the system successfully predicted 5 dangerous geological bodies and identified 2 confined water enrichment areas in advance, providing an important basis for the optimization of the construction plan.

[0279] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Dynamic geological survey system, characterized by: Including intelligent sensing and data acquisition module and multi-source data fusion processing module; The intelligent sensing and data acquisition module includes a Raman scattering optical fiber strain sensor array, a three-component seismometer monitoring network and a distributed optical fiber pressure sensing subsystem. Raman scattering optical fiber strain sensor array, three-component seismometer monitoring network, and distributed optical fiber pressure sensing subsystem constitute a multi-dimensional sensing network for geological monitoring of shield construction; Multi-protocol data convergence unit: built-in chip realizes synchronous acquisition of multi-source data; The multi-source data fusion processing module includes: heterogeneous data processing engine: equipped with a GPU computing cluster, adopting a streaming computing architecture, supporting parallel processing of geological, hydrological, and mechanical heterogeneous data; The spatiotemporal attention fusion algorithm acts on the heterogeneous data fusion stage to process the multimodal data of geological strain, microseismic events, and pore water pressure from the intelligent sensing and data acquisition modules.

2. The dynamic geological survey system according to claim 1, characterized in that: It also includes a high-precision dynamic construction simulation module, which adopts a hybrid calculation model of the finite element method and the discrete element method, including: an adaptive meshing function with a minimum unit size of 1 cm; a nonlinear constitutive model based on the modified Mohr-Coulomb criterion, introducing strain softening and hardening parameters.

3. The dynamic geological survey system according to claim 1, characterized in that: The parameters of the nonlinear constitutive model are updated in real time through an intelligent optimization algorithm, including: dynamic adjustment range of cohesion c: 0.1kPa-500kPa; friction angle Update formula: φ t+1 =φ t +α·(c p -c threshold ), where α is the learning rate, c p is the current strain value, c threshold is the strain threshold.

4. The dynamic geological survey system according to claim 1, characterized in that: It also includes an intelligent early warning decision module, which includes: Multi-level early warning model, covering geological mutation, water pressure anomaly, construction parameter deviation risk indicators; The deep reinforcement learning algorithm is based on a dual network architecture: the policy network is a three-layer fully connected structure with 512, 256, and 128 neurons respectively; the value network is a two-layer convolution and one-layer fully connected structure with a convolution kernel size of 5×5 and a step size of 2.

5. The dynamic geological survey system according to claim 1, characterized in that: The training process of the deep reinforcement learning algorithm includes: the state space S contains geological parameters, water pressure values, and shield machine thrust; the action space A is the adjustment amount of the cutter head speed and propulsion speed; the reward function R is designed as: Among them, w1 and w2 are weight coefficients, and the penalty coefficient is w3.

6. The dynamic geological survey system according to claim 1, characterized in that: It also includes a cloud collaborative management module, which adopts: a distributed microservice architecture; an edge computing unit: a three-level architecture including a terminal layer, an edge node layer, and a cloud layer; A lightweight neural network model is deployed at the terminal layer for data preprocessing; The edge node layer is equipped with high-performance computing units to implement local data analysis.

7. The dynamic geological survey system according to claim 1, characterized in that: The attention mechanism algorithm realizes multi-source data fusion through dynamic weight adjustment: The input feature matrix dimension is [N×D], where N is the number of time steps and D is the feature dimension; The output dimension of the multi-head self-attention layer is [N×8×64], and after concatenation, the output is [N×512]; The cross-modal feature fusion layer uses a bidirectional GRU, and the hidden state update formula is: h t =σ(W hz x t +W hh h t-1 +b h ) Where σ is the Sigmoid activation function, W hz , W hh is the weight matrix.

8. A method for dynamic geological survey and construction simulation of shield construction, characterized in that: The following steps are involved: S1. Real-time collection and transmission of geological data: An array of optical fiber strain sensors is laid out every 10 meters along the shield tunneling route; Deploy 48 three-component seismometers and use an improved spectrum analysis algorithm to extract microseismic signals; A distributed fiber-optic pressure sensing subsystem monitors water pressure in real time; S2. Multi-source data fusion processing: Use GPU clusters to process heterogeneous data in parallel; Fusion of geological, hydrological, and mechanical data through attention mechanism algorithms; The multi-head self-attention layer calculates the feature correlation matrix; The cross-modal fusion layer uses Bi-GRU to capture temporal dependencies; S3. High-precision construction dynamic simulation: Construct a finite element-discrete element hybrid model with a minimum grid size of 1 cm; Dynamically update boundary conditions to simulate the coupling effect between stratum and shield machine; S4. Intelligent early warning and decision-making: Evaluate risk indicators based on deep reinforcement learning models; When the prediction error probability P≥0.9, a first-level warning is triggered and pushed to the terminal device through the network; S5. Cloud Collaborative Management: Use edge computing to process 80% of data locally, reducing bandwidth requirements; The visualization platform displays the three-dimensional formation stress field in real time.

9. A shield construction dynamic geological survey and construction simulation method according to claim 8, characterized in that: The specific process of the attention mechanism algorithm in step S2 includes: a. Input feature matrix Generate Q, K, V matrices through linear projection; b. Calculate multi-head attention weights: where d k =64 is the key vector dimension; c. Cross-modal fusion layer output: h t =GRU(x t ,h t-1 ) Hidden layer dimension 256, time step N = 100; d. The temporal dependency layer uses causal convolution to retain historical information constraints.

10. A shield construction dynamic geological survey and construction simulation method according to claim 8, characterized in that: The adaptive meshing method of the hybrid model in step S3 includes: Determine the mesh encryption area based on the curvature threshold θ=0.05; dynamically adjust the unit size Δl: in is the displacement gradient; the grid topology is updated every 30 minutes.

Images