Tunnel settlement real-time warning method, device, equipment and storage medium

Abstract

The present invention provides a method, device, equipment, and storage medium for real-time tunnel settlement early warning. The method comprises: obtaining input and output signals of an optical fiber installed in the vault of a target tunnel at multiple times, as well as temperature and seepage pressure data at different monitoring points at multiple times; obtaining a strain matrix at each time based on the input and output signals at that time; obtaining a spatial feature tensor at that time based on the strain matrix; concatenating the spatial feature tensors at multiple times according to time to obtain a time series matrix; obtaining a time series feature tensor based on the time series matrix; correcting the time series feature tensor based on the temperature and seepage pressure data at different monitoring points at multiple times to obtain a target feature tensor; inputting the target feature tensor into a tunnel settlement calculation model to obtain the settlement value of the target tunnel; and issuing an early warning based on the settlement value of the target tunnel. The present invention can improve the real-time performance and accuracy of tunnel settlement early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a tunnel settlement real-time early warning method, device, equipment and storage medium. Background Art

[0002] During the construction and operation and maintenance of tunnel projects, settlement monitoring is a key link in ensuring structural safety. However, current technologies have some limitations. First, the inversion accuracy is insufficient. Traditional BOTDR technology relies on parameterized models set by manual experience, resulting in inversion errors as high as ±2mm, which makes it difficult to meet the needs of high-precision shield control. Secondly, data processing efficiency is low. Traditional methods take more than 30 minutes to process 1 km of optical fiber data, making real-time monitoring difficult to achieve. Thirdly, there is a lag in early warning response. The existing system relies on manual interpretation of data, making it difficult to achieve early warning response in hours.

[0003] Currently, existing models are not yet able to adapt to complex geological conditions, such as the coupled effects of fluid permeability and temperature. Current technologies are limited in accuracy, real-time performance, and collaborative decision-making capabilities. Therefore, there is an urgent need to develop an intelligent monitoring and early warning system with high accuracy and good timeliness. Summary of the Invention

[0004] The embodiments of the present invention provide a tunnel settlement real-time early warning method, apparatus, device and storage medium to improve the real-time performance and accuracy of tunnel settlement early warning.

[0005] In a first aspect, an embodiment of the present invention provides a real-time early warning method for tunnel settlement, comprising:

[0006] The input signal and output signal of the optical fiber installed on the vault of the target tunnel at multiple times, as well as the temperature data and seepage pressure data at different monitoring points on the vault of the target tunnel at multiple times are obtained.

[0007] For each moment, the strain matrix at that moment is obtained based on the input signal and output signal at that moment; based on the strain matrix at that moment, the spatial feature tensor at that moment is obtained; wherein each element in the strain matrix represents the strain parameter at the monitoring point at different spatial positions.

[0008] The spatial feature tensors of multiple moments are spliced ​​according to time to obtain a time series matrix.

[0009] Based on the time series matrix, a time series feature tensor is obtained; based on the temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times, the time series feature tensor is corrected to obtain the target feature tensor.

[0010] The target characteristic tensor is input into the tunnel settlement calculation model to obtain the settlement value of the target tunnel; based on the settlement value of the target tunnel, an early warning is issued.

[0011] In one possible implementation, the optical fiber is laid in a spiral along the vault of the target tunnel.

[0012] According to the laying direction of the optical fiber, a monitoring point is set every preset length.

[0013] In one possible implementation, the spatial feature tensor at that moment is obtained based on the strain matrix at that moment, including:

[0014] Get the three-dimensional coordinates of each monitoring point.

[0015] According to the cosine function, the three-dimensional coordinates of each monitoring point are converted into the position code of each monitoring point.

[0016] Get the dimension of the location encoding of each monitoring point.

[0017] Based on the dimension of the position code of each monitoring point, each element in the strain matrix at that moment is dimensionally processed to obtain a target matrix; wherein the dimension of each element in the target matrix is ​​the same as the dimension of the position code of the corresponding monitoring point.

[0018] Each element in the target matrix is ​​added to the position code of the corresponding monitoring point to obtain the first matrix.

[0019] Based on the first matrix, the sampling offset of each monitoring point is generated through the fully connected network MLP.

[0020] Based on the deformable attention DeformAttn, the first matrix and the sampling offset of each monitoring point are aggregated to obtain the aggregated feature matrix.

[0021] The aggregated feature matrix is ​​normalized through a linear layer to obtain the spatial feature tensor at that moment.

[0022] In one possible implementation, a time series feature tensor is obtained based on the time series matrix, including:

[0023] Transpose the time series matrix to obtain the second matrix.

[0024] The second matrix is ​​input into the first bidirectional gated recurrent unit to extract minute-level temporal features to obtain a first output result.

[0025] The result after feature normalization of the second matrix is ​​recorded as the first normalization result.

[0026] The sum of the first output result and the first normalized result is input into the second bidirectional gated recurrent unit to extract the hourly time series features to obtain the second output result.

[0027] The sum of the first output result and the first normalized result is subjected to feature normalization, which is recorded as the second normalized result.

[0028] The sum of the second output result and the second normalized result is input into the third bidirectional gated recurrent unit to extract the day-level temporal features to obtain the third output result.

[0029] The sum of the second output result and the second normalized result is subjected to feature normalization, which is recorded as the third normalized result.

[0030] The sum of the third output result and the third normalized result is input into the fourth bidirectional gated recurrent unit to extract the cycle-level timing features to obtain a fourth output result.

[0031] Transpose the fourth output result to obtain the time series feature tensor.

[0032] In one possible implementation, based on the temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times, the time series feature tensor is corrected to obtain the target feature tensor, including:

[0033] The temperature data and the osmotic pressure data are linearly transformed and interpolated to obtain the temperature data correction matrix and the osmotic pressure data correction matrix.

[0034] The temperature data correction matrix and the osmotic pressure data correction matrix are concatenated to obtain the environmental correction tensor.

[0035] The dimension of the environment correction tensor is reduced to obtain a corrected environment correction tensor; wherein the corrected environment correction tensor has the same dimension as the time series feature tensor.

[0036] The correlation between the corrected environment correction tensor and the time series feature tensor is calculated to obtain the correction weight.

[0037] Calculate the product of the corrected environmental correction tensor and the correction weight to obtain the environmental correction value corresponding to each monitoring point.

[0038] Based on the time series feature tensor and the environmental correction value corresponding to each monitoring point, the target feature tensor is obtained.

[0039] In one possible implementation, the training process of the tunnel settlement calculation model includes:

[0040] Obtain the target feature tensor and its corresponding real value of tunnel settlement in the historical database.

[0041] The target feature tensor in the historical database is input into the preset GRU model to obtain the simulation value corresponding to the target feature tensor in the historical database.

[0042] Based on the real value and the simulated value, the loss function of the preset GRU model is obtained.

[0043] Determine whether the result of the loss function of the preset GRU model is greater than a first threshold.

[0044] If the result of the loss function is greater than or equal to the first threshold, the parameters of the preset GRU model are updated and the simulation value corresponding to the target feature tensor in the historical database is recalculated; if the result of the loss function is less than the first threshold, the preset GRU model is defined as the tunnel settlement calculation model.

[0045] In one possible implementation, an early warning is issued based on the settlement value of the target tunnel, including:

[0046] If the settlement value of the target tunnel is greater than the first preset threshold and less than or equal to the second preset threshold, a visual warning is issued for the settlement value of the target tunnel and the warning is pushed to the staff.

[0047] If the settlement value of the target tunnel is greater than the second preset threshold and less than or equal to the third preset threshold, a visual early warning is issued for the settlement value of the target tunnel, and the drone is controlled to perform autonomous inspection.

[0048] If the settlement value of the target tunnel is greater than a third preset threshold, a tunnel reinforcement plan is dynamically generated.

[0049] In a second aspect, an embodiment of the present invention provides a real-time early warning device for tunnel subsidence, comprising:

[0050] The first processing module is used to obtain input signals and output signals of the optical fiber installed on the target tunnel vault at multiple times, as well as temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times.

[0051] The second processing module is used to obtain the strain matrix at each moment based on the input signal and output signal at that moment; based on the strain matrix at that moment, obtain the spatial feature tensor at that moment; wherein each element in the strain matrix represents the strain parameter at a monitoring point at a different spatial position.

[0052] The third processing module is used to splice the spatial feature tensors of multiple moments according to time to obtain a time series matrix.

[0053] The fourth processing module is used to obtain a time series feature tensor based on the time series matrix; based on the temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times, the time series feature tensor is corrected to obtain a target feature tensor.

[0054] The fifth processing module is used to input the target characteristic tensor into the tunnel settlement calculation model to obtain the settlement value of the target tunnel; and issue an early warning based on the settlement value of the target tunnel.

[0055] In a third aspect, an embodiment of the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method in the first aspect or any possible implementation of the first aspect is implemented.

[0056] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method in the first aspect or any possible implementation of the first aspect.

[0057] In a fifth aspect, an embodiment of the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the method in the first aspect or any possible implementation of the first aspect.

[0058] In this embodiment of the present invention, a strain matrix is ​​acquired through the input and output signals of optical fibers. The strain matrices are then spliced ​​together in time to form a time-series matrix. A time-series characteristic tensor derived from the time-series matrix is ​​then corrected using temperature and seepage pressure data at different monitoring points. Finally, the corrected target characteristic tensor is input into the tunnel settlement calculation model to determine the settlement value of the target tunnel and issue an early warning based on the settlement value. This solution fully considers the impact of environmental factors on strain data, achieves adaptive adaptation to complex geological conditions, and improves calculation accuracy. Furthermore, by avoiding the use of traditional strain sensors and instead using optical fibers, costs can be reduced and monitoring points can be set flexibly, freeing from the limitations of traditional strain sensors (which are costly and increase with more monitoring points). Furthermore, compared to solutions that collect and transmit data from a large number of strain sensors, this solution only requires the collection of input and output signals, resulting in less data collection and faster data transmission (data transmission efficiency from a large number of strain sensors is affected by bandwidth; excessive data volume can lead to insufficient bandwidth and transmission delays). This allows for more rapid settlement calculations and early warnings, ensuring the timeliness of the solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart for implementing a real-time early warning method for tunnel settlement provided by an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of optical fiber layout provided by an embodiment of the present invention;

[0061] Figure 3Schematic diagram of the structure of a real-time early warning device for tunnel settlement provided by an embodiment of the present invention;

[0062] Figure 4 is a schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0064] See also Figure 1 , which shows a flow chart for implementing a real-time early warning method for tunnel settlement provided by an embodiment of the present invention, as detailed below:

[0065] Step 101: acquiring input signals and output signals of an optical fiber installed on a target tunnel vault at multiple times, as well as temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times.

[0066] For example, the optical fiber installed on the vault of the target tunnel is a distributed sensing optical fiber, such as Figure 2 As shown in the figure, S1 is the fiber input port, S2 is the fiber output port, and the fiber is laid out in a spiral pattern along the vault of the target tunnel. Monitoring points are set at predetermined intervals along the fiber's layout. Specifically, the fiber is laid out along the fiber, from S1 to S2.

[0067] For example, the fiber helical winding density is 5 turns per meter, with a pitch of 20 cm. The fiber is primarily laid using a gelling agent such as epoxy resin, with both overall bonding and point fixation. During the laying process, a level is used to draw auxiliary lines and maintain the fiber in a pre-tensioned state. Finally, the fiber is connected in series to form a complete system and connected to the data monitoring module. The fiber model can be: SMF-28e + single-mode fiber. The strain sensitivity after pre-tensioning can be: ±0.35με (compared to ±1.0με for traditional linear layout).

[0068] Specifically, the strain matrix dimension is M×512 (M monitoring points, each with 512 channels of spectral data). Further monitoring points can be set up: along the fiber's layout direction, a monitoring point is set every 30 cm, starting from the fiber's input port and continuing to the fiber's output port.

[0069] Step 102: For each moment, based on the input signal and output signal at that moment, a strain matrix at that moment is obtained; based on the strain matrix at that moment, a spatial feature tensor at that moment is obtained; wherein each element in the strain matrix represents a strain parameter at a monitoring point at a different spatial position.

[0070] Exemplarily, the processing of input and output signals, that is, the process of obtaining a strain matrix, falls within the scope of optical fiber monitoring technology. For example, the strain matrix is ​​obtained by processing the input and output signals through BOTDA, BOTDR, BOFDA, and other technologies.

[0071] In one possible implementation, obtaining the spatial feature tensor at the moment based on the strain matrix at the moment may include:

[0072] Get the three-dimensional coordinates of each monitoring point.

[0073] According to the cosine function, the three-dimensional coordinates of each monitoring point are converted into the position code of each monitoring point.

[0074] Get the dimension of the location encoding of each monitoring point.

[0075] Based on the dimension of the position code of each monitoring point, each element in the strain matrix at that moment is dimensionally processed to obtain a target matrix; wherein the dimension of each element in the target matrix is ​​the same as the dimension of the position code of the corresponding monitoring point.

[0076] Each element in the target matrix is ​​added to the position code of the corresponding monitoring point to obtain the first matrix.

[0077] Based on the first matrix, the sampling offset of each monitoring point is generated through the fully connected network MLP.

[0078] Based on the deformable attention DeformAttn, the first matrix and the sampling offset of each monitoring point are aggregated to obtain the aggregated feature matrix.

[0079] The aggregated feature matrix is ​​normalized through a linear layer to obtain the spatial feature tensor at that moment.

[0080] For example, the traditional strain matrix is ​​converted into a spatial feature tensor, which can fully capture local deformation in space and solve the problem that traditional methods (only collecting strain matrices) cannot dynamically capture local deformation. This solution can improve the strain detection sensitivity of joints and abnormal formation areas.

[0081] Exemplarily, based on the first matrix, a sampling offset of each monitoring point is generated by a fully connected network MLP, including:

[0082] The first matrix is ​​input into the fully connected network MLP to generate the sampling offset of each monitoring point.

[0083] The calculation formula for the sampling offset of each monitoring point is:

[0084]

[0085] in, Indicates the The sampling offset of each monitoring point, represents the MLP function, Indicates the Monitoring points in the first matrix The corresponding eigenvector in .

[0086] The calculation formula of the aggregate feature matrix is:

[0087]

[0088] in, represents the aggregate feature matrix, Represents the DeformAttn function, which indicates the sampling offset.

[0089] The calculation formula of the spatial feature tensor at this moment is:

[0090]

[0091] in, Represents a spatial feature tensor.

[0092] In addition to the above-mentioned spatial feature tensor calculation scheme without attention mechanism, attention mechanism can also be introduced. The specific calculation logic is as follows:

[0093] 1. Offset generation and sampling point calculation

[0094] Offset learning: Generate the sampling offset of each monitoring point through a fully connected network .

[0095] Among them: 8 attention heads (using 8-head DeformableTransformer modules), 4 sampling points per head, 2 represents the 2D plane offset (the cross section x, y direction of the target tunnel) .

[0096] Sampling point coordinates: reference point After superimposing the offset, the dynamic sampling point is obtained:

[0097]

[0098] The significance of dynamic sampling points: In the water-rich sand layer, the model automatically learns to offset the sampling points to the abnormal seepage pressure area (such as K1+500) to capture the local strain concentration characteristics.

[0099] 2. Attention weight calculation

[0100] Feature similarity measurement: Calculate monitoring points With sampling point The feature association weight of

[0101]

[0102] Among them, 64 is the single-head feature dimension (512 dimensions / 8 heads).

[0103] The physical meaning of weight: At the seams, weight will increase significantly, reflecting the contribution of this area to the overall settlement (such as the strong correlation between joint opening and settlement).

[0104] 3. Feature Sampling and Aggregation

[0105] Bilinear interpolation sampling: for non-grid sampling points , obtain the sampling value through feature interpolation of the four surrounding real monitoring points:

[0106] in, is the interpolation weight, Number the adjacent monitoring points.

[0107] Multi-head feature aggregation:

[0108] Each of the eight heads captures a different feature—for example, heads 1-2 focus on hoop strain (fiber segment roundness), heads 3-4 focus on longitudinal strain gradient (tunnel axis bending), and heads 5-8 capture temperature coupling effects.

[0109] 4. Multi-head feature fusion and output

[0110] Concatenate the 8-head output into a 512-dimensional feature vector: .

[0111] The significance of splicing: This operation integrates multi-dimensional features such as segment structure (headers 1-2), formation interaction (headers 3-4), and environmental interference (headers 5-8) to form a comprehensive deformation representation.

[0112] 5. Linear Projection and Layer Normalization

[0113] Adjust the feature distribution through the linear layer:

[0114] in, is the weight matrix, For bias.

[0115] Finally, as a spatial feature tensor.

[0116] Step 103: splice the spatial feature tensors of multiple moments according to time to obtain a time series matrix.

[0117] Step 104 : obtaining a time series characteristic tensor based on the time series matrix; and correcting the time series characteristic tensor based on the temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times to obtain a target characteristic tensor.

[0118] For example, temperature data is acquired using FBG temperature sensors. One sensor is welded every meter, for a total of 2000 sensors. The wavelength range is 1525-1565 nm, and the temperature resolution is ±0.05°C. Osmotic pressure data is acquired using FBG osmotic pressure sensors. One sensor is implanted every 5 meters, for a total of 400 sensors. The sensors are buried 5 cm behind the tunnel lining, with a sampling frequency of 1 Hz.

[0119] Specifically: FBG temperature sensors and osmotic pressure sensors both measure data within a range. Therefore, the temperature data or osmotic pressure data at the installation site of the FBG temperature sensor and osmotic pressure sensor can represent the temperature data and osmotic pressure data of multiple monitoring points. If you want to correct the time series feature tensor, you need to interpolate the data first to align the data quantity.

[0120] In one possible implementation, obtaining a time series feature tensor based on a time series matrix may include:

[0121] Transpose the timing matrix to obtain the second matrix.

[0122] The second matrix is ​​input into the first bidirectional gated recurrent unit to extract minute-level temporal features to obtain a first output result.

[0123] The result after feature normalization of the second matrix is ​​recorded as the first normalization result.

[0124] The sum of the first output result and the first normalized result is input into the second bidirectional gated recurrent unit to extract the hourly time series features to obtain the second output result.

[0125] The sum of the first output result and the first normalized result is subjected to feature normalization, which is recorded as the second normalized result.

[0126] The sum of the second output result and the second normalized result is input into the third bidirectional gated recurrent unit to extract the day-level temporal features to obtain the third output result.

[0127] The sum of the second output result and the second normalized result is subjected to feature normalization, which is recorded as the third normalized result.

[0128] The sum of the third output result and the third normalized result is input into the fourth bidirectional gated recurrent unit to extract the cycle-level timing features to obtain a fourth output result.

[0129] Transpose the fourth output result to obtain the time series feature tensor.

[0130] For example, through processing multiple layers of bidirectional gated recurrent units, temporal features from minute to weekly levels are extracted, enabling full-cycle settlement evolution analysis from minute to weekly levels and reducing long-term creep prediction errors. It is important to note that starting with the second layer of bidirectional gated recurrent units, the input data becomes the residual of the input and output data of the previous layer (for the second bidirectional gated recurrent unit, for example, its input is the sum of the first output result and the first normalized result). This preserves transient characteristics and improves the accuracy of settlement inversion at geological interfaces.

[0131] For example, in addition to the above process, the following steps may also be included:

[0132] Before inputting the second matrix into the first bidirectional gated recurrent unit, the second matrix is ​​also time-normalized.

[0133] The calculation formula of the normalized second matrix is:

[0134]

[0135] in, represents the second matrix after time series normalization, represents the second matrix, represents the eigenmean of the second matrix, represents the characteristic standard deviation of the second matrix.

[0136] Specifically, the core formula of the bidirectional GRU (taking the lth layer as an example) includes:

[0137] Update Gate:

[0138] Reset the gate:

[0139] Candidate hidden states:

[0140] Hide status update:

[0141] Bidirectional output: (weighted sum of forward / backward hidden states)

[0142] Parameter configuration: input dimension: 512, hidden layer size: 256 (unidirectional), bidirectional output dimension = 256.

[0143] Activation function: σ is Sigmoid, tanh is hyperbolic tangent.

[0144] In one possible implementation, based on temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times, correcting the time series feature tensor to obtain the target feature tensor may include:

[0145] The temperature data and the osmotic pressure data are linearly transformed and interpolated to obtain the temperature data correction matrix and the osmotic pressure data correction matrix.

[0146] The temperature data correction matrix and the osmotic pressure data correction matrix are concatenated to obtain the environmental correction tensor.

[0147] The dimension of the environment correction tensor is reduced to obtain a corrected environment correction tensor; wherein the corrected environment correction tensor has the same dimension as the time series feature tensor.

[0148] The correlation between the corrected environment correction tensor and the time series feature tensor is calculated to obtain the correction weight.

[0149] Calculate the product of the corrected environmental correction tensor and the correction weight to obtain the environmental correction value corresponding to each monitoring point.

[0150] Based on the time series feature tensor and the environmental correction value corresponding to each monitoring point, the target feature tensor is obtained.

[0151] For example, the correlation between the modified environmental correction tensor and the time series feature tensor can be calculated using a dot product. When the correlation is greater than 0.7, it corresponds to a first preset correction weight. When the correlation is less than or equal to 0.7, it corresponds to a second preset correction weight. The first preset correction weight is greater than the second preset correction weight. Highly correlated regions are assigned a high weight, suppressing the strain value in that region.

[0152] During winter construction, the tunnel vault temperature dropped from 20°C to 5°C. After correction, the temperature-related component of the strain characteristic dropped from 0.5με to below 0.1με, avoiding the risk of misjudging it as structural settlement.

[0153] The temperature and osmotic pressure physical fields are converted into computable characteristic weights to achieve dynamic correction of environmental interference in optical fiber strain monitoring, ultimately reducing the error in the calculation of the target characteristic tensor.

[0154] Step 105: Input the target characteristic tensor into the tunnel settlement calculation model to obtain the settlement value of the target tunnel; and issue an early warning based on the settlement value of the target tunnel.

[0155] In one possible implementation, the training process of the tunnel settlement calculation model may include:

[0156] Obtain the target feature tensor and its corresponding real value of tunnel settlement in the historical database.

[0157] The target feature tensor in the historical database is input into the preset GRU model to obtain the simulation value corresponding to the target feature tensor in the historical database.

[0158] Based on the real value and the simulated value, the loss function of the preset GRU model is obtained.

[0159] Determine whether the result of the loss function of the preset GRU model is greater than a first threshold.

[0160] If the result of the loss function is greater than or equal to the first threshold, the parameters of the preset GRU model are updated and the simulation value corresponding to the target feature tensor in the historical database is recalculated; if the result of the loss function is less than the first threshold, the preset GRU model is defined as the tunnel settlement calculation model.

[0161] Exemplarily, the first threshold may be 0.05.

[0162] For example, the loss function of the preset GRU model includes three parts: the mean square error between the real value and the simulated value, Hinge loss (penalizing sudden changes in sedimentation rate with a threshold of ±0.5 mm / day), and TV regularization (suppressing spatial noise and ensuring smooth sedimentation of adjacent points).

[0163] Specifically, the calculation function of the loss function is:

[0164]

[0165] in, represents the mean square error, represents Hinge loss, represents TV regularization.

[0166] For example, to ensure the applicability of subsequent tunnel settlement calculation models, it is necessary to integrate the corresponding tunnel settlement calculation models for each target tunnel. First, knowledge distillation is performed on the models using a KL divergence-constrained distillation loss function to obtain multiple distilled models. The Cross-Silo federated averaging algorithm is then applied to these multiple distilled models, with differential privacy protection (DP noise equal to 0.8) to obtain multiple differential models. Finally, these differential models are aggregated to obtain a tunnel settlement calculation model applicable to multiple target tunnels.

[0167] The tunnel settlement calculation model applicable to multiple target tunnels has wider applicability. When the tunnel settlement calculation model for a certain target tunnel does not exist in the database, the tunnel settlement calculation model applicable to multiple target tunnels can be used to calculate the corresponding settlement value.

[0168] In one possible implementation, the early warning based on the settlement value of the target tunnel may include:

[0169] If the settlement value of the target tunnel is greater than the first preset threshold and less than or equal to the second preset threshold, a visual warning is issued for the settlement value of the target tunnel and the warning is pushed to the staff.

[0170] If the settlement value of the target tunnel is greater than the second preset threshold and less than or equal to the third preset threshold, a visual early warning is issued for the settlement value of the target tunnel, and the drone is controlled to perform autonomous inspection.

[0171] If the settlement value of the target tunnel is greater than a third preset threshold, a tunnel reinforcement plan is dynamically generated.

[0172] For example, the first preset threshold may be 8 mm, the second preset threshold may be 12 mm, and the third preset threshold may be 15 mm.

[0173] For example, in order to better reflect the urgency of the settlement value, a three-dimensional visual modeling of the target tunnel was also carried out, and a digital twin of the tunnel was constructed with a three-dimensional grid resolution of 1 cm³. Sensor data was synchronously updated every 10 minutes to achieve visual dynamic deduction of settlement deformation.

[0174] Specifically, the three-dimensional visual modeling engine may be: FEM-PSO algorithm.

[0175] Finite element mesh used for modeling: 0.1m density, 2 km tunnel is divided into 20,000 elements (20,000 × 0.1m = 2,000m).

[0176] Particle swarm optimization: 100 iterations / cycle, a single iteration takes 15 seconds, and the residual converges to within 5%.

[0177] Input: Settlement values ​​predicted by the model.

[0178] Output: 3D settlement coordinates, accuracy ±1mm.

[0179] The particle swarm algorithm is used to repeatedly optimize the modeling results. When the residual converges to within 5%, it can be considered that the accuracy of the 3D visualization modeling meets the conditions and the displayed 3D model is not much different from the actual situation.

[0180] For example, the tunnel reinforcement plan is based on the NSGA-III multi-objective optimization algorithm, which comprehensively considers factors such as maintenance cost, construction safety, and construction period impact to generate a Pareto optimal solution set (coverage ≥ 90%), supporting intelligent decision-making for preventive maintenance plans.

[0181] Specifically, the calculation process of the reinforcement plan may include:

[0182] 1. Data preparation:

[0183] Collect historical cases and build a database (such as MySQL or MongoDB). Each case contains multiple fields, such as: project name, geological conditions (soil type, water content, burial depth, etc.), cross-sectional dimensions, deformation, reinforcement plan (grouting pressure, grouting material, steel support model, spacing, etc.), construction results, etc.

[0184] The key clauses of relevant specifications (such as GB50157-2013) are extracted, stored in the database, and associated with the cases (for example, the reinforcement plan of a case complies with a certain clause of the specification).

[0185] 2. Build a knowledge base:

[0186] DeepSeek-R1 is used to vectorize the descriptive text of each historical case (such as geological conditions, engineering problems, etc.) to obtain a high-dimensional vector (embedding), which is stored in a vector database (such as Milvus, FAISS, ChromaDB, etc.).

[0187] Similarly, the specification terms are also vectorized and stored.

[0188] 3. Case matching engine:

[0189] The user enters a description of the current case (such as geological conditions, engineering problems, etc.).

[0190] The system converts the current case description into a vector through DeepSeek-R1, and performs similarity search (such as cosine similarity) in the vector database to find historical cases with a similarity ≥ 82%.

[0191] At the same time, the system matches the current case with the regulatory clauses to find the applicable regulatory clauses.

[0192] 4. Solution generation:

[0193] If historical cases with a similarity of ≥82% are matched, the reinforcement parameters of these cases (grouting pressure, steel support layout, etc.) are extracted and verified and adjusted in combination with the code provisions.

[0194] If no sufficiently similar historical cases are matched, a solution is generated based on the specification terms and domain knowledge (generated through LLM).

[0195] Grouting pressure plan: provide pressure range, grouting materials, grouting steps, etc.

[0196] Steel support layout plan: Provide support model, layout spacing, connection method, etc., and generate a schematic diagram (requires graphics library support, such as Matplotlib or front-end graphics library).

[0197] 5. Integrate LLM for knowledge reasoning:

[0198] During the matching process, LLM can be used to compare and analyze the current case and the matched historical cases, explain the reasons for the matching, and provide a matching solution (reinforcement solution).

[0199] For example, the solution can be deployed on the corresponding edge detection terminal of each target tunnel, and the cloud platform only displays the three-dimensional visual modeling and settlement value calculation results.

[0200] Pruning strategy: Structured pruning is performed on GRU layer neurons, removing 35% of redundant connections, reducing the number of parameters and reducing computational complexity by 40%.

[0201] Quantization solution: FP32 floating-point parameters are converted to INT8 fixed-point numbers, and the exponential moving average (EMA) is used to calibrate the error, with a compensation rate of 6% to ensure accuracy loss of less than 1%.

[0202] The edge terminal uses Jetson Orin NX (32GB RAM, 200TOPS computing power), with a single-channel inference delay of less than 0.7 seconds, supporting the processing of full-section data of an 8km tunnel within 56 seconds, meeting real-time monitoring needs.

[0203] The above-mentioned real-time tunnel settlement early warning method uses optical fiber input and output signals to acquire a strain matrix. This strain matrix is ​​then spliced ​​together to form a time-series matrix. The time-series characteristic tensor derived from the time-series matrix is ​​then corrected using temperature and seepage pressure data at different monitoring points. Finally, the corrected target characteristic tensor is input into the tunnel settlement calculation model to determine the settlement value of the target tunnel and issue an early warning based on the settlement value. This solution fully considers the influence of environmental factors on strain data, achieves adaptive adaptation to complex geological conditions, and improves calculation accuracy. It also avoids the use of traditional strain sensors and instead uses optical fiber, which reduces costs and allows for flexible monitoring point configuration, freeing it from the limitations of traditional strain sensors (which are costly and increase with more monitoring points). Furthermore, compared to solutions that collect and transmit data from a large number of strain sensors, this solution only requires the collection of input and output signals, resulting in less data collection and faster data transmission (data transmission efficiency from a large number of strain sensors is affected by bandwidth; excessive data volume can lead to insufficient bandwidth and transmission delays). This allows for more rapid settlement calculations and early warnings, ensuring the timeliness of the solution.

[0204] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0205] The following are device embodiments of the present invention. For details not fully described therein, reference may be made to the corresponding method embodiments described above.

[0206] Figure 3 The following is a schematic diagram showing the structure of a real-time early warning device for tunnel settlement according to an embodiment of the present invention. For ease of explanation, only the parts related to the embodiment of the present invention are shown, which are described in detail as follows:

[0207] like Figure 3 As shown, the tunnel settlement real-time early warning device includes:

[0208] The first processing module 201 is used to obtain input signals and output signals of the optical fiber installed on the target tunnel vault at multiple times, as well as temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times.

[0209] The second processing module 202 is used to obtain the strain matrix at each moment based on the input signal and output signal at that moment; based on the strain matrix at that moment, obtain the spatial feature tensor at that moment; wherein each element in the strain matrix represents the strain parameter at a monitoring point at a different spatial position.

[0210] The third processing module 203 is used to splice the spatial feature tensors of multiple moments according to time to obtain a time series matrix.

[0211] The fourth processing module 204 is configured to obtain a time series feature tensor based on the time series matrix; and to modify the time series feature tensor based on temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times to obtain a target feature tensor.

[0212] The fifth processing module 205 is configured to input the target characteristic tensor into the tunnel settlement calculation model to obtain the settlement value of the target tunnel; and issue an early warning based on the settlement value of the target tunnel.

[0213] In one possible implementation, the optical fiber is laid in a spiral along the vault of the target tunnel.

[0214] According to the laying direction of the optical fiber, a monitoring point is set every preset length.

[0215] In a possible implementation, the second processing module 202 may be configured to:

[0216] Get the three-dimensional coordinates of each monitoring point.

[0217] According to the cosine function, the three-dimensional coordinates of each monitoring point are converted into the position code of each monitoring point.

[0218] Get the dimension of the location encoding of each monitoring point.

[0219] Based on the dimension of the position code of each monitoring point, each element in the strain matrix at that moment is dimensionally processed to obtain a target matrix; wherein the dimension of each element in the target matrix is ​​the same as the dimension of the position code of the corresponding monitoring point.

[0220] Each element in the target matrix is ​​added to the position code of the corresponding monitoring point to obtain the first matrix.

[0221] Based on the first matrix, the sampling offset of each monitoring point is generated through the fully connected network MLP.

[0222] Based on the deformable attention DeformAttn, the first matrix and the sampling offset of each monitoring point are aggregated to obtain the aggregated feature matrix.

[0223] The aggregated feature matrix is ​​normalized through a linear layer to obtain the spatial feature tensor at that moment.

[0224] In a possible implementation, the fourth processing module 204 may be configured to:

[0225] Transpose the timing matrix to obtain the second matrix.

[0226] The second matrix is ​​input into the first bidirectional gated recurrent unit to extract minute-level temporal features to obtain a first output result.

[0227] The result after feature normalization of the second matrix is ​​recorded as the first normalization result.

[0228] The sum of the first output result and the first normalized result is input into the second bidirectional gated recurrent unit to extract the hourly time series features to obtain the second output result.

[0229] The sum of the first output result and the first normalized result is subjected to feature normalization, which is recorded as the second normalized result.

[0230] The sum of the second output result and the second normalized result is input into the third bidirectional gated recurrent unit to extract the day-level temporal features to obtain the third output result.

[0231] The sum of the second output result and the second normalized result is subjected to feature normalization, which is recorded as the third normalized result.

[0232] The sum of the third output result and the third normalized result is input into the fourth bidirectional gated recurrent unit to extract the cycle-level timing features to obtain a fourth output result.

[0233] Transpose the fourth output result to obtain the time series feature tensor.

[0234] In a possible implementation, the fourth processing module 204 may be configured to:

[0235] The temperature data and the osmotic pressure data are linearly transformed and interpolated to obtain the temperature data correction matrix and the osmotic pressure data correction matrix.

[0236] The temperature data correction matrix and the osmotic pressure data correction matrix are concatenated to obtain the environmental correction tensor.

[0237] The dimension of the environment correction tensor is reduced to obtain a corrected environment correction tensor; wherein the corrected environment correction tensor has the same dimension as the time series feature tensor.

[0238] The correlation between the corrected environment correction tensor and the time series feature tensor is calculated to obtain the correction weight.

[0239] Calculate the product of the corrected environmental correction tensor and the correction weight to obtain the environmental correction value corresponding to each monitoring point.

[0240] Based on the time series feature tensor and the environmental correction value corresponding to each monitoring point, the target feature tensor is obtained.

[0241] In one possible implementation, the training process of the tunnel settlement calculation model includes:

[0242] Obtain the target feature tensor and its corresponding real value of tunnel settlement in the historical database.

[0243] The target feature tensor in the historical database is input into the preset GRU model to obtain the simulation value corresponding to the target feature tensor in the historical database.

[0244] Based on the real value and the simulated value, the loss function of the preset GRU model is obtained.

[0245] Determine whether the result of the loss function of the preset GRU model is greater than a first threshold.

[0246] If the result of the loss function is greater than or equal to the first threshold, the parameters of the preset GRU model are updated and the simulation value corresponding to the target feature tensor in the historical database is recalculated; if the result of the loss function is less than the first threshold, the preset GRU model is defined as the tunnel settlement calculation model.

[0247] In a possible implementation, the fifth processing module 205 may be configured to:

[0248] If the settlement value of the target tunnel is greater than the first preset threshold and less than or equal to the second preset threshold, a visual warning is issued for the settlement value of the target tunnel and the warning is pushed to the staff.

[0249] If the settlement value of the target tunnel is greater than the second preset threshold and less than or equal to the third preset threshold, a visual early warning is issued for the settlement value of the target tunnel, and the drone is controlled to perform autonomous inspection.

[0250] If the settlement value of the target tunnel is greater than a third preset threshold, a tunnel reinforcement plan is dynamically generated.

[0251] Figure 4 Schematic diagram of an electronic device provided by an embodiment of the present invention. Figure 4 As shown, the electronic device 5 of this embodiment includes: a processor 50 and a memory 51. The memory 51 stores a computer program 52. When the processor 50 executes the computer program 52, the steps of the above-mentioned method embodiments are implemented. Alternatively, when the processor 50 executes the computer program 52, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0252] For example, the computer program 52 may be divided into one or more modules / units, which are stored in the memory 51 and executed by the processor 50 to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program 52 in the electronic device 5.

[0253] The electronic device 5 may include, but is not limited to, a processor 50 and a memory 51. Those skilled in the art will appreciate that Figure 4 It is only an example of the electronic device 5 and does not constitute a limitation of the electronic device 5. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device 5 may also include input and output devices, network access devices, buses, etc.

[0254] The processor 50 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0255] The memory 51 can be an internal storage unit of the electronic device 5, such as the hard drive or memory of the electronic device 5. The memory 51 can also be an external storage device of the electronic device 5, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. equipped on the electronic device 5. Furthermore, the memory 51 can include both the internal storage unit of the electronic device 5 and an external storage device. The memory 51 is used to store the computer program 52 and other programs and data required by the electronic device 5. The memory 51 can also be used to temporarily store data that has been output or is about to be output.

[0256] For the sake of convenience and brevity, the division of the above functional modules / units is only used as an example. In actual applications, the above functions can be assigned to different functional modules / units as needed. The above modules / units can be implemented in the form of hardware, software, or a combination of hardware and software.

[0257] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods in the above-mentioned method embodiments.

[0258] An embodiment of the present invention further provides a computer program product, including a computer program, which, when executed by a processor, implements the methods in the above-mentioned method embodiments.

[0259] The term "computer program" includes computer program code, which may be in source code form, object code form, executable file, or some intermediate form. Computer-readable media may include any entity or device capable of carrying computer program code, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunications signals, and software distribution media.

[0260] In the above embodiments, the descriptions of each embodiment have their own focus. For parts not described or recorded in detail in one embodiment, please refer to the relevant descriptions of other embodiments. Unless otherwise specified or there is a logical conflict, the terms and / or descriptions between different embodiments are consistent and can be referenced to each other. The technical features of different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0261] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A real-time early warning method for tunnel settlement, characterized in that: include: Acquiring input signals and output signals of an optical fiber installed on a target tunnel vault at multiple times, as well as temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times; For each moment, a strain matrix at that moment is obtained based on the input signal and output signal at that moment; based on the strain matrix at that moment, a spatial feature tensor at that moment is obtained; wherein each element in the strain matrix represents a strain parameter at a monitoring point at a different spatial position; The spatial feature tensors of multiple moments are spliced ​​according to time to obtain a time series matrix; Based on the time series matrix, a time series characteristic tensor is obtained; based on the temperature data and seepage pressure data at different monitoring points of the target tunnel vault at the multiple time points, the time series characteristic tensor is corrected to obtain a target characteristic tensor; Inputting the target characteristic tensor into a tunnel settlement calculation model to obtain a settlement value of the target tunnel; and issuing an early warning based on the settlement value of the target tunnel; The step of correcting the time series characteristic tensor based on the temperature data and seepage pressure data at different monitoring points on the target tunnel vault at the multiple time points to obtain the target characteristic tensor includes: Performing linear transformation and interpolation processing on the temperature data and the osmotic pressure data to obtain a temperature data correction matrix and an osmotic pressure data correction matrix; Concatenating the temperature data correction matrix and the osmotic pressure data correction matrix to obtain an environmental correction tensor; Performing dimensionality reduction on the environment correction tensor to obtain a corrected environment correction tensor; wherein the corrected environment correction tensor has the same dimension as the time series feature tensor; Calculating the correlation between the corrected environment correction tensor and the time series feature tensor to obtain a correction weight; Calculating the product of the corrected environment correction tensor and the correction weight to obtain the environment correction value corresponding to each monitoring point; The target feature tensor is obtained based on the time series feature tensor and the environmental correction value corresponding to each monitoring point.

2. The tunnel settlement real-time early warning method according to claim 1 is characterized in that: The optical fiber is laid out in a spiral manner along the vault of the target tunnel; According to the laying direction of the optical fiber, a monitoring point is set every preset length.

3. The real-time early warning method for tunnel subsidence according to claim 1, characterized in that: The spatial feature tensor at the moment is obtained based on the strain matrix at the moment, including: Obtain the three-dimensional coordinates of each monitoring point; According to the cosine function, the three-dimensional coordinates of each monitoring point are converted into the position code of each monitoring point; Get the dimension of the location code of each monitoring point; Performing dimension processing on each element in the strain matrix at the moment based on the dimension of the position code of each monitoring point to obtain a target matrix; wherein the dimension of each element in the target matrix is ​​the same as the dimension of the position code of the corresponding monitoring point; Adding each element in the target matrix to the position code of the corresponding monitoring point to obtain a first matrix; Based on the first matrix, a sampling offset of each monitoring point is generated through a fully connected network MLP; Based on the deformable attention DeformAttn, the first matrix and the sampling offset of each monitoring point are aggregated to obtain an aggregated feature matrix; The aggregated feature matrix is ​​normalized through a linear layer to obtain the spatial feature tensor at that moment.

4. The real-time early warning method for tunnel subsidence according to claim 1, characterized in that: The obtaining of a time series feature tensor based on the time series matrix includes: Transposing the timing matrix to obtain a second matrix; Inputting the second matrix into a first bidirectional gated recurrent unit to extract minute-level temporal features to obtain a first output result; The result after feature normalization of the second matrix is ​​recorded as the first normalization result; Inputting the sum of the first output result and the first normalized result into a second bidirectional gated recurrent unit to extract hourly time series features to obtain a second output result; performing feature normalization on the sum of the first output result and the first normalized result, and recording the result as a second normalized result; Inputting the sum of the second output result and the second normalized result into a third bidirectional gated recurrent unit to extract the day-level time series features to obtain a third output result; performing feature normalization on the sum of the second output result and the second normalized result, and recording the result as a third normalized result; Inputting the sum of the third output result and the third normalized result into a fourth bidirectional gated recurrent unit to extract cycle-level temporal features to obtain a fourth output result; Transpose the fourth output result to obtain the time series feature tensor.

5. The real-time early warning method for tunnel subsidence according to claim 1, characterized in that: The training process of the tunnel settlement calculation model includes: Obtain the target feature tensor and its corresponding real value of tunnel settlement in the historical database; Inputting the target feature tensor in the historical database into a preset GRU model to obtain a simulation value corresponding to the target feature tensor in the historical database; Based on the real value and the simulated value, obtaining a loss function of the preset GRU model; Determine whether a result of the loss function of the preset GRU model is greater than a first threshold; If the result of the loss function is greater than or equal to the first threshold, the parameters of the preset GRU model are updated and the simulation value corresponding to the target feature tensor in the historical database is recalculated; if the result of the loss function is less than the first threshold, the preset GRU model is defined as the tunnel settlement calculation model.

6. The real-time early warning method for tunnel subsidence according to claim 1, characterized in that: The issuing of an early warning based on the settlement value of the target tunnel includes: If the settlement value of the target tunnel is greater than the first preset threshold and less than or equal to the second preset threshold, a visual warning is issued for the settlement value of the target tunnel, and the warning is pushed to the staff; If the settlement value of the target tunnel is greater than the second preset threshold and less than or equal to the third preset threshold, a visual early warning is issued for the settlement value of the target tunnel, and the drone is controlled to perform an autonomous inspection; If the settlement value of the target tunnel is greater than the third preset threshold, a tunnel reinforcement plan is dynamically generated.

7. A real-time early warning device for tunnel settlement, characterized in that: include: A first processing module is configured to obtain input signals and output signals of an optical fiber installed on a target tunnel vault at multiple times, as well as temperature data and seepage pressure data at different monitoring points on the target tunnel vault at multiple times; A second processing module is configured to obtain, for each moment, a strain matrix at that moment based on the input signal and output signal at that moment; and obtain a spatial feature tensor at that moment based on the strain matrix at that moment; wherein each element in the strain matrix represents a strain parameter at a monitoring point at a different spatial position; The third processing module is used to splice the spatial feature tensors of multiple moments according to time to obtain a time series matrix; a fourth processing module, configured to obtain a time series characteristic tensor based on the time series matrix; and to modify the time series characteristic tensor based on the temperature data and seepage pressure data at different monitoring points on the target tunnel vault at the multiple time points to obtain a target characteristic tensor; a fifth processing module, configured to input the target characteristic tensor into a tunnel settlement calculation model to obtain a settlement value of the target tunnel; and issue an early warning based on the settlement value of the target tunnel; The fourth processing module is further configured to: Performing linear transformation and interpolation processing on the temperature data and the osmotic pressure data to obtain a temperature data correction matrix and an osmotic pressure data correction matrix; Concatenating the temperature data correction matrix and the osmotic pressure data correction matrix to obtain an environmental correction tensor; Performing dimensionality reduction on the environment correction tensor to obtain a corrected environment correction tensor; wherein the corrected environment correction tensor has the same dimension as the time series feature tensor; Calculating the correlation between the corrected environment correction tensor and the time series feature tensor to obtain a correction weight; Calculating the product of the corrected environment correction tensor and the correction weight to obtain the environment correction value corresponding to each monitoring point; The target feature tensor is obtained based on the time series feature tensor and the environmental correction value corresponding to each monitoring point.

8. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images