Subway tunnel vault displacement monitoring method and system based on multi-source data fusion

Abstract

The application discloses a subway tunnel vault displacement monitoring method and system based on multi-source data fusion, and the method comprises the following steps: synchronously collecting multi-source vault data, converting non-point cloud data into point cloud format and unifying to a tunnel construction coordinate system, performing random rotation, Gauss noise addition and local clipping on original point cloud to expand a data set; a 3D PointNet++ model performs end-to-end processing on the pretreated mixed point cloud, selects key points through a sampling layer, aggregates features through a feature aggregation layer, processes through an attention mechanism layer, and outputs vault displacement characteristic values; an LSTM-Transformer hybrid model inputs displacement time series data and environmental parameters, captures displacement long-period rules, identifies local anomalies, and outputs anomaly probabilities and generates anomaly data correction values; the vault displacement and rate change are predicted by using the corrected data, different levels of early warnings are triggered, and corresponding operations are performed. The application provides reliable guarantee for subway operation safety.

Description

Technical Field

[0001] This invention belongs to the field of subway tunnel monitoring technology, and more specifically, relates to a method and system for monitoring the displacement of the tunnel arch based on multi-source data fusion. Background Technology

[0002] As a vital urban transportation infrastructure, the structural safety of subway tunnels is directly related to urban operation and the safety of people's lives and property. During the long-term operation of subway tunnels, crown displacement is one of the key indicators reflecting structural stability. However, current monitoring of subway tunnel crown displacement faces many practical challenges.

[0003] On the one hand, tunnel arch displacement is influenced by multiple factors, including geological conditions, construction techniques, operational loads, and environmental factors, exhibiting complex spatiotemporal variation characteristics. Traditional monitoring methods, such as manual measurement and single-point sensor monitoring, suffer from limited coverage, single data types, and poor spatiotemporal synchronization, making it difficult to comprehensively and in real-time capture the three-dimensional displacement state of the arch. For example, single-point sensors can only acquire displacement information at local points and cannot reflect the overall deformation trend of the arch. Furthermore, the inconsistent acquisition times and coordinate systems of different monitoring devices make it difficult to effectively fuse and analyze multi-source data.

[0004] On the other hand, the tunnel operating environment is complex, and the crown displacement data often contains noise interference and outliers. Traditional data analysis methods struggle to accurately identify valid information and abnormal data, let alone accurately predict displacement trends. Failure to detect abnormal crown displacement and issue early warnings in a timely manner may lead to serious accidents such as tunnel structural cracking, water leakage, or even collapse, which will not only affect the normal operation of the subway but also cause huge economic losses and social impacts.

[0005] Furthermore, with the rapid development of urban rail transit and the ever-expanding subway network, the requirements for intelligent and automated tunnel monitoring are becoming increasingly stringent. Existing monitoring technologies, in terms of data fusion, feature extraction, time-series analysis, and intelligent early warning, are insufficient to meet the demands of modern subway tunnel safety operation and maintenance. Therefore, there is an urgent need for a method for monitoring tunnel arch displacement that can achieve multi-source data fusion, accurately extract displacement features, effectively identify anomalies, and provide intelligent prediction and early warning. This would improve the level of intelligent monitoring of subway tunnel structural safety and ensure the safe operation of subways, which is of significant practical importance for the sustainable development of urban rail transit. Summary of the Invention

[0006] This invention aims to solve the problems of single data type, poor spatiotemporal synchronization, and difficulty in effectively integrating multi-source data in existing technologies; at the same time, it overcomes the shortcomings of traditional methods in terms of inaccurate displacement feature extraction, weak ability to identify abnormal data, and inability to accurately predict displacement trends, so as to achieve comprehensive perception, accurate analysis, and intelligent early warning of displacement of the tunnel arch, improve the intelligent level of tunnel structure safety monitoring, and provide reliable protection for the safety of subway operation.

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion, comprising: S1. The preprocessing process realizes the spatiotemporal synchronous acquisition of multi-source arch data, converts non-point cloud data into point cloud format and unifies it to the tunnel construction coordinate system, and then expands the dataset size by performing operations on the original point cloud through the 3DPointNet++ front-end module, including random rotation, Gaussian noise addition, and local clipping. The S2.3DPointNet++ model performs end-to-end processing on the preprocessed hybrid point cloud data. It filters key points through a sampling layer, fuses local and global information through a feature aggregation layer, and dynamically allocates weights through an attention mechanism layer, outputting the fused arch displacement feature value. The S3.LSTM-Transformer hybrid model takes displacement time series data and environmental parameters as input, captures long-period patterns of displacement data through LSTM layer, identifies local anomalies through Transformer layer and jointly outputs anomaly probability, and then generates anomaly correction values ​​based on the time series distribution characteristics of normal data. The S4.LSTM-Transformer model uses corrected effective data to predict the trend of crown displacement and rate change, and triggers different levels of early warning and executes corresponding response operations based on the prediction results.

[0008] Furthermore, the multi-source arch data in S1 includes: three-dimensional point cloud data of the arch collected by a 3D laser scanner, distance change data of specific points on the arch collected by a laser rangefinder, coordinate data of characteristic points on the arch collected by a total station, and displacement data of tunnel segment joints collected by a displacement sensor.

[0009] Furthermore, the acquisition process of multi-source arch data in S1 is as follows: the arch monitoring area is synchronously acquired through a 3D laser scanner, laser range sensor, total station and displacement sensor, and the acquisition time of each device is aligned through a time synchronization mechanism to form a multi-dimensional raw data set including three-dimensional point cloud, distance change, coordinate and joint displacement. The time synchronization mechanism is as follows: a unified time reference signal is sent to all acquisition devices through a time synchronization module deployed in the tunnel monitoring system. Each device calibrates its internal clock according to the reference signal, so that all data acquisition operations are performed on the same time scale, thus completing the matching of the acquisition times of multi-source data.

[0010] Furthermore, the specific process of unifying to the tunnel construction coordinate system in S1 is as follows: Let the coordinates of the original point cloud data in the coordinate system of the 3D laser scanner be... The initial coordinates of the collected non-point cloud data after conversion are: , Representing different non-point cloud data types, the origin of the tunnel construction coordinate system is... Its X-axis is along the tunnel axis, its Y-axis is perpendicular to the tunnel axis in the horizontal direction, and its Z-axis is along the tunnel vertical direction; First, by using pre-set known coordinate reference points within the tunnel, the transformation parameters between the coordinate systems of each acquisition device and the tunnel construction coordinate system are obtained; the origin of the 3D laser scanner coordinate system is set relative to the origin of the tunnel construction coordinate system. The translation vector is The rotation matrix between the two coordinate systems is Then the original point cloud data is transformed into coordinates in the tunnel construction coordinate system. satisfy: For non-point cloud data, the initial coordinates are first determined based on the spatial geometric relationship between the sensor installation location and the reference point. Translation vector relative to the tunnel construction coordinate system and rotation matrix Coordinates after transformation satisfy: Through the matrix transformation and translation operations based on spatial geometric relationships described above, multi-source data can be unified into the tunnel construction coordinate system, ensuring the consistency of data under the same spatial reference.

[0011] Furthermore, the method for selecting key points in S2 is as follows: Let the hybrid point cloud dataset be... ,in The first in the representative point cloud The coordinates of a point in the tunnel construction coordinate system. This represents the total number of point clouds; key point selection is achieved through the following process: Initialize the key point set From point cloud datasets Randomly select a point As the initial key point, at this time ; Iterative selection of key points; for points not included in the point cloud dataset. any point Calculate the relationship between this point and Calculate the Euclidean distances to all selected key points and take the minimum value as the distance from that point to the target key. The distance is given by the formula: in, For point Three-dimensional coordinates in the tunnel construction coordinate system; It is a set of key points Any key point in the data; each iteration selects... The largest point Add S, update ; Determine the screening termination condition; when the number of points in S satisfies the condition of being able to completely represent the geometric profile of the vault, or when new points are added. When the value is less than the minimum structural feature size of the arch, the iteration stops, and S is the set of key points selected.

[0012] Furthermore, the method for calculating the characteristic value of the arch displacement in S2 is as follows: Let the set of key points obtained after screening be... ,in For the first The current coordinates of the key points in the tunnel construction coordinate system, and their coordinates at the initial acquisition time are: ; For each key point First, calculate its absolute displacement vector in three-dimensional space. : Then, a feature aggregation layer is used to extract all points in the neighborhood of that point. The relative displacement characteristics, for The local neighborhood point set, defining the points within the neighborhood. relatively displacement deviation for: in, For point The absolute displacement vector, The Euclidean norm of a vector; All The mean of is used as a local consistency index, and Combined to form the comprehensive feature vector of this key point: in, Indicates the first The comprehensive feature vector of each key point is used to characterize the displacement and local consistency information of that key point. Indicates the number of points in the neighborhood; Finally, for all key locations Spatial correlation analysis is performed, and the overall change trend of each dimension is taken as the final arch displacement characteristic value, which fully reflects the deformation state of the arch in three-dimensional space and the correlation characteristics between the local and the whole.

[0013] Furthermore, the LSTM layer in S3 specifically comprises: Let S2 output the time series of arch displacement characteristic values. , The time step corresponds to the sequence of environmental parameters at that time. Including influencing factors such as temperature and humidity, the LSTM layer processes timing information through a three-gating mechanism: The input gate is based on the current feature value. With environmental parameters fusion vector The formula for determining whether new information should be included in the cell state is as follows: ,in, Indicates the input gate output of the LSTM layer. For the sigmoid function, This represents the cell state at the previous moment. , The gate matrix is ​​generated during model training. The forget gate is based on the state of historical cells. With current input Preserving long-term patterns, the calculation formula is: , , Forget gate matrix, Indicates the output of the forget gate; Cell status updated to , For element-wise product, Generate candidate states. The state matrix, This represents the current state of the cell. The output gate combines the updated cell state with the current input to generate the hidden state. ,in, This represents the weight matrix of the output gate. This represents the cyclic weight matrix of the output gate. It incorporates both historical cycle characteristics and current environmental influences, and is transmitted to the Transformer layer as an output of long-term cycle patterns.

[0014] Furthermore, the specific process by which the Transformer layer in S3 identifies local anomalies and jointly outputs the anomaly probability is as follows: Let the hidden state sequence output by the LSTM layer be... , For time steps, It is the first In the hidden state at any given time, the Transformer layer identifies local anomalies and outputs the anomaly probability through the following process: Multi-head attention correlation calculation; for each attention head, Mapped to query vector Key vector Value vector Calculate any two time points and Relevance: in, As a vector dimension, through the correlation matrix Capture temporal correlations; Local deviation metric; for time Calculate its average correlation with all other times. Define the deviation : The greater the deviation, the more significant the difference between the data at that moment and the overall pattern. Anomaly probability generation; combining the long-period regularity characteristics of the LSTM layer output, time intervals are calculated. abnormal probability : in, The denominator is defined as the radius of the time window, and the minimum correlation between neighboring times is introduced into the denominator to strengthen the influence of local continuity on anomaly detection; when When the threshold is exceeded based on the distribution of historical normal data, it is marked as a local anomaly.

[0015] As a second aspect of the present invention, a subway tunnel arch displacement monitoring system based on multi-source data fusion is also provided, comprising: The multi-source data preprocessing unit is used to realize the spatiotemporal synchronous acquisition of multi-source arch data in the preprocessing process, convert non-point cloud data into point cloud format and unify it to the tunnel construction coordinate system, and then expand the dataset size by performing operations on the original point cloud through the 3DPointNet++ front-end module, including random rotation, Gaussian noise addition, and local clipping. The displacement feature value output unit is used by the 3DPointNet++ model to perform end-to-end processing on the preprocessed hybrid point cloud data. It filters key points through the sampling layer, fuses local and global information through the feature aggregation layer, and dynamically allocates weights through the attention mechanism layer, and outputs the fused arch displacement feature value. Anomaly identification and correction unit is used to input displacement time series data and environmental parameters into the LSTM-Transformer hybrid model. It captures long-period patterns of displacement data through the LSTM layer, identifies local anomalies through the Transformer layer and jointly outputs the anomaly probability, and then generates anomaly data correction values ​​based on the time series distribution characteristics of normal data. The displacement early warning response unit is used by the LSTM-Transformer model to predict the displacement and rate change trend of the arch using the corrected effective data, and to trigger different levels of early warning and execute corresponding response operations based on the prediction results.

[0016] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor of any step of the method for monitoring displacement of the arch of a subway tunnel based on multi-source data fusion.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The subway tunnel arch displacement monitoring method based on multi-source data fusion of this invention achieves spatiotemporal synchronous acquisition of multi-source arch data through a preprocessing process, converting non-point cloud data into point cloud format and unifying it to the tunnel construction coordinate system. Then, the 3DPointNet++ pre-processing module is used to expand the dataset size by performing random rotation, Gaussian noise addition, and local cropping operations on the original point cloud. This technical feature solves the problem of data integration difficulties caused by inconsistent multi-source data formats and large differences in coordinate systems in traditional monitoring. Simultaneously, the dataset expansion improves the adaptability of subsequent model data processing, ensuring effective fusion of multi-source data and providing a comprehensive and consistent data foundation for subsequent displacement feature extraction.

[0018] 2. The subway tunnel arch displacement monitoring method based on multi-source data fusion of the present invention uses a 3DPointNet++ model to perform end-to-end processing on preprocessed mixed point cloud data. Key points are selected through a sampling layer, local and global information is fused through a feature aggregation layer, and weights are dynamically allocated through an attention mechanism layer, outputting the fused arch displacement feature values. This technique avoids the problems of missing key information and fragmentation of local and global information in traditional feature extraction. Through the model's end-to-end processing capabilities, it accurately extracts the core features of arch displacement, enabling the output displacement feature values ​​to comprehensively reflect the spatial deformation state of the arch, providing accurate feature data support for subsequent time-series analysis and anomaly identification.

[0019] 3. The multi-source data fusion-based method for monitoring the displacement of the tunnel arch in this invention uses an LSTM-Transformer hybrid model to input displacement time-series data and environmental parameters. The LSTM layer captures long-term patterns in the displacement data, the Transformer layer identifies local anomalies and outputs anomaly probabilities, and then generates corrected values ​​for the anomaly data based on the normal data time-series distribution. Finally, the corrected data is used to predict the trend of arch displacement and rate changes and trigger early warnings. This technical feature solves the problems of insufficient long-term pattern capture, large interference from anomaly data, and low prediction accuracy in traditional time-series analysis. It achieves accurate analysis, anomaly correction, and trend prediction of displacement data, and can promptly trigger early warnings of different levels, providing timely and reliable decision-making basis for the safe operation and maintenance of tunnel structures. Attached Figure Description

[0020] Figure 1 This is a flowchart of a subway tunnel arch displacement monitoring method based on multi-source data fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a coordinate system processing procedure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the key point selection process for the three-dimensional point cloud of the tunnel arch in an embodiment of the present invention. Figure 4 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Example 1 Please refer to Figure 1This embodiment 1 provides a method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion, including: S1. The preprocessing process realizes the spatiotemporal synchronous acquisition of multi-source arch data, converts non-point cloud data into point cloud format and unifies it to the tunnel construction coordinate system, and then expands the dataset size by performing operations on the original point cloud through the 3D PointNet++ front-end module, including random rotation, Gaussian noise addition, and local clipping. The S2.3D PointNet++ model performs end-to-end processing on the preprocessed hybrid point cloud data. It filters key points through a sampling layer, fuses local and global information through a feature aggregation layer, and dynamically allocates weights through an attention mechanism layer, outputting the fused arch displacement feature value. The S3.LSTM-Transformer hybrid model takes displacement time series data and environmental parameters as input, captures long-period patterns of displacement data through LSTM layer, identifies local anomalies through Transformer layer and jointly outputs anomaly probability, and then generates anomaly correction values ​​based on the time series distribution characteristics of normal data. The S4.LSTM-Transformer model uses corrected effective data to predict the trend of crown displacement and rate change, and triggers different levels of early warning and executes corresponding response operations based on the prediction results.

[0023] This embodiment 1 further elaborates on the above steps.

[0024] (1) Multi-source data preprocessing In monitoring the displacement of the tunnel arch, the effective integration of multi-source data is the foundation for accurate analysis. However, traditional monitoring often leads to analytical biases due to the dispersed data sources and inconsistent spatiotemporal benchmarks. Therefore, the preprocessing process needs to first solve the problem of collaborative acquisition and unification of multi-source data.

[0025] This process first targets the arch monitoring scenario, integrating 3D point cloud data of the arch collected by a 3D laser scanner, distance change data of specific points on the arch collected by a laser rangefinder, coordinate data of characteristic points on the arch collected by a total station, and displacement data of tunnel segment joints collected by a displacement sensor. During the acquisition process, a unified time reference signal is sent to all acquisition devices through a time synchronization module deployed in the tunnel monitoring system. Each device calibrates its internal clock according to this reference signal, ensuring that all data acquisition operations are performed on the same time scale, completing the matching of the acquisition times of multi-source data; forming a raw data set containing multi-dimensional information, laying the foundation for time synchronization for subsequent data fusion.

[0026] Please refer to Figure 2Considering the differences in coordinate systems between different devices, direct use would lead to spatial inconsistencies. Therefore, the preprocessing workflow needs to further unify the spatial distribution of the data. Let the coordinates of the original point cloud data in the 3D laser scanner's own coordinate system be... The initial coordinates of the collected non-point cloud data after conversion are: , Representing different non-point cloud data types, the origin of the tunnel construction coordinate system is... Its X-axis is along the tunnel axis, its Y-axis is perpendicular to the tunnel axis in the horizontal direction, and its Z-axis is along the tunnel vertical direction; First, by using pre-set known coordinate reference points within the tunnel, the transformation parameters between the coordinate systems of each acquisition device and the tunnel construction coordinate system are obtained; the origin of the 3D laser scanner coordinate system is set relative to the origin of the tunnel construction coordinate system. The translation vector is The rotation matrix between the two coordinate systems is Then the original point cloud data is transformed into coordinates in the tunnel construction coordinate system. satisfy: For non-point cloud data, the initial coordinates are first determined based on the spatial geometric relationship between the sensor installation location and the reference point. Translation vector relative to the tunnel construction coordinate system and rotation matrix Coordinates after transformation satisfy: Through the matrix transformation and translation operations based on spatial geometric relationships described above, multi-source data can be unified into the tunnel construction coordinate system, ensuring the consistency of data under the same spatial reference.

[0027] Furthermore, the original point cloud data may lack sufficient scale and diversity, affecting the subsequent model's learning effect on the arch deformation features. The preprocessing workflow uses the 3D PointNet++ pre-module to perform random rotation, Gaussian noise addition, and local cropping operations on the unified original point cloud. This simulates the data form under different monitoring angles, environmental interference, and local occlusion scenarios, expanding the scale and diversity of the dataset. This ensures that the subsequent model can fully learn various features of the arch displacement, improves its adaptability to complex deformation situations, and provides high-quality, highly adaptable data support for subsequent end-to-end feature extraction and analysis.

[0028] (2) Output of displacement characteristic value The preprocessed hybrid point cloud data provides a unified spatiotemporal benchmark and a rich sample base for the 3D PointNet++ model. The 3D PointNet++ model performs end-to-end processing on the preprocessed hybrid point cloud data, forming a coherent feature extraction process. First, the sampling layer filters key points from the point cloud; it randomly selects initial points and iteratively calculates the minimum distance from unselected points to selected key points, adding the points with the largest distances to the set until the points are sufficient to characterize the arch outline or the distance of newly selected points is less than the minimum structural feature size, thus solving the data redundancy problem and unifying the coordinates with the preprocessed coordinates.

[0029] Next, the feature aggregation layer fuses the information; it calculates the three-dimensional absolute displacement vector of each key point, and combines it with the mean of the relative displacement deviation of the neighboring points to form a comprehensive feature vector, which utilizes the integrity of the preprocessed data and solves the problem of isolated information at a single point.

[0030] Finally, the attention mechanism layer dynamically allocates weights, extracts the overall trend through spatial correlation analysis, and outputs the arch displacement feature value, providing high-quality input for subsequent time series analysis and avoiding information loss in step-by-step processing.

[0031] like Figure 3 As shown, specifically, the method for selecting key points is as follows: Let the hybrid point cloud dataset be... ,in The first in the representative point cloud The coordinates of a point in the tunnel construction coordinate system. This represents the total number of point clouds; key point selection is achieved through the following process: Initialize the key point set From point cloud datasets Randomly select a point As the initial key point, at this time .

[0032] Iterative selection of key points; for points not included in the point cloud dataset. any point Calculate the relationship between this point and Calculate the Euclidean distances to all selected key points and take the minimum value as the distance from that point to the target key. The distance is given by the formula: in, For point Three-dimensional coordinates in the tunnel construction coordinate system; It is a set of key points Any key point in the data; each iteration selects... The largest point Add S, update ; Determine the screening termination condition; when the number of points in S satisfies the condition of being able to completely represent the geometric profile of the vault, or when new points are added. When the value is less than the minimum structural feature size of the arch, the iteration stops, and S is the set of key points selected.

[0033] Furthermore, based on the selected key points, the model calculates feature values: for each key point, it calculates the three-dimensional absolute displacement vector between the current and initial coordinates, and then combines this with the average relative displacement deviation of its neighborhood points to form a comprehensive feature vector. The specific calculation method for the arch displacement feature values ​​is as follows: Let the set of key points obtained after screening be... ,in For the first The current coordinates of the key points in the tunnel construction coordinate system, and their coordinates at the initial acquisition time are: ; For each key point First, calculate its absolute displacement vector in three-dimensional space. : Then, a feature aggregation layer is used to extract all points in the neighborhood of that point. The relative displacement characteristics, for The local neighborhood point set, defining the points within the neighborhood. relatively displacement deviation for: in, For point The absolute displacement vector, The Euclidean norm of a vector is denoted by .

[0034] All The mean of is used as a local consistency index, and Combined to form the comprehensive feature vector of this key point: in, Indicates the first The comprehensive feature vector of each key point is used to characterize the displacement and local consistency information of that key point. Indicates the number of points in the neighborhood; Finally, for all key locations Spatial correlation analysis is performed, and the overall change trend of each dimension is taken as the final arch displacement characteristic value, which fully reflects the deformation state of the arch in three-dimensional space and the correlation characteristics between the local and the whole.

[0035] This process not only continues the spatial consistency advantage brought by the unified coordinate system in the preprocessing stage, but also solves the problem that traditional single-point monitoring cannot reflect the relationship between local and overall deformation. It enables the feature values ​​to reflect both the deformation amount at specific points and the deformation coordination of local areas, providing feature data that takes into account both details and the whole for subsequent analysis.

[0036] Finally, the model dynamically allocates weights through an attention mechanism layer, obtains the overall trend of change through spatial correlation analysis, and outputs the fused arch displacement feature values. This step strengthens the influence of important features, solves the problem of insufficient prominence of key information caused by equal feature weights in different regions, and its output feature values ​​not only inherit the information from the preprocessed multi-source data, but also provide high-quality input for subsequent temporal pattern capture and anomaly identification through accurate extraction, forming a complete link from data integration to feature extraction, laying the foundation for a comprehensive understanding of the arch displacement state.

[0037] (3) Anomaly identification and correction After receiving the displacement time-series data and corresponding environmental parameters output by the 3D PointNet++ model, the LSTM-Transformer hybrid model first processes the data through a three-gating mechanism of the LSTM layer. The displacement feature values ​​and environmental parameters are fused into a vector. The input gate filters whether new information should be included in the cell state, the forget gate retains long-term historical patterns (such as seasonal displacement cycles caused by temperature changes), the cell state integrates historical and current valid information, and the output gate generates a hidden state that contains both historical periodic features and integrates current environmental influences. Specifically, in a preferred embodiment, the LSTM layer is as follows: Let S2 output the time series of arch displacement characteristic values. , The time step corresponds to the sequence of environmental parameters at that time. Including influencing factors such as temperature and humidity, the LSTM layer processes timing information through a three-gating mechanism: The input gate is based on the current feature value. With environmental parameters fusion vector The formula for determining whether new information should be included in the cell state is as follows: ,in, Indicates the input gate output of the LSTM layer. For the sigmoid function, This represents the cell state at the previous moment. , The gate matrix is ​​generated during model training. The forget gate is based on the state of historical cells. With current input Preserving long-term patterns, the calculation formula is: , , Forget gate matrix, Indicates the output of the forget gate; Cell status updated to , For element-wise product, Generate candidate states. The state matrix, This represents the current state of the cell. The output gate combines the updated cell state with the current input to generate the hidden state. ,in, This represents the weight matrix of the output gate. This represents the cyclic weight matrix of the output gate. It incorporates both historical cycle characteristics and current environmental influences, and is transmitted to the Transformer layer as an output of long-term cycle patterns.

[0038] Next, the Transformer layer performs local anomaly identification based on the hidden state sequence output by the LSTM. The specific process by which the Transformer layer identifies local anomalies and jointly outputs the anomaly probability is as follows: Let the hidden state sequence output by the LSTM layer be... , For time steps, It is the first In the hidden state at any given time, the Transformer layer identifies local anomalies and outputs the anomaly probability through the following process: Multi-head attention correlation calculation; for each attention head, Mapped to query vector Key vector Value vector Calculate any two time points and Relevance: in, As a vector dimension, through the correlation matrix Capture temporal correlations; Local deviation metric; for time Calculate its average correlation with all other times. Define the deviation : The greater the deviation, the more significant the difference between the data at that moment and the overall pattern. Anomaly probability generation; combining the long-period regularity characteristics of the LSTM layer output, time intervals are calculated. abnormal probability : in, The denominator is defined as the radius of the time window, and the minimum correlation between neighboring times is introduced into the denominator to strengthen the influence of local continuity on anomaly detection; when When the threshold is exceeded based on the distribution of historical normal data, it is marked as a local anomaly.

[0039] This operation connects to the long-period pattern output of the LSTM layer, solving the problem that local anomalies are easily masked by long-period patterns. It can accurately identify sudden displacement anomalies (such as sudden deformation of segment joints) and avoid affecting tunnel safety judgment due to anomaly omission.

[0040] Finally, the model jointly outputs the anomaly probability and generates anomaly correction values ​​based on the time-series distribution characteristics of normal data. This step not only builds upon the anomaly identification results mentioned earlier but also clears data obstacles for subsequent displacement trend prediction, solving the problem of anomaly data interference analysis. This ensures that subsequent prediction results can accurately reflect the arch displacement state, forming a complete link from feature input to data correction, and providing coherent and reliable technical support for tunnel safety operation and maintenance.

[0041] (4) Displacement early warning response The LSTM-Transformer model uses corrected effective data for prediction. The LSTM layer continues to capture long-term trends and combines the correlation between environmental factors and displacement in historical data. The Transformer layer focuses on recent key deformation nodes through an attention mechanism and pays attention to the changes in displacement rate in local areas. The two work together to generate the three-dimensional displacement of the arch and the rate change curves in each direction for a future preset period.

[0042] The predicted results are compared with a multi-level threshold system based on structural safety limits, covering absolute displacement, rate change, and trend acceleration thresholds. When all indicators are normal, the system records data and maintains routine monitoring; when a single indicator approaches the threshold, a level-one warning is triggered, key areas are marked, and an inspection list is pushed out; when multiple indicators exceed limits and the trend deteriorates, a level-two warning is activated, an emergency assessment report is generated and pushed to relevant departments, and measures such as train speed limits and section closures are implemented, and a professional dispatch team is prepared for on-site handling.

[0043] This process is based on corrected data, and uses models to accurately grasp short-term and long-term trends to achieve advanced predictions, transforming post-event response into pre-event intervention, forming a closed loop from data correction to trend prediction to graded handling, thereby improving the timeliness and pertinence of tunnel safety management.

[0044] The multi-source data fusion-based method for monitoring the displacement of subway tunnel arches in Example 1 has broad application prospects in the field of urban rail transit operation and maintenance. Currently, most subway tunnels in Chinese cities are in the mid-to-late stages of operation, and factors such as structural aging and changes in the geological environment are leading to an increasing risk of arch displacement year by year. Traditional monitoring methods suffer from problems such as difficulty in data integration and delayed early warning. This method, through a complete chain of multi-source data spatiotemporal synchronization, accurate feature extraction using 3D PointNet++, and prediction and early warning using LSTM-Transformer, can adapt to different geological conditions (such as soft soil and rock strata) and tunnel structure types. It can be directly applied to the daily safety monitoring of subway tunnels, and is especially suitable for core lines with busy operations where frequent manual inspections are difficult. It helps the operation and maintenance team to grasp the arch deformation status in real time and avoid potential structural safety hazards in advance.

[0045] Meanwhile, the technical logic of this method can also be extended to similar underground projects, such as river-crossing tunnels and integrated utility tunnels. These projects also face common challenges such as difficulty in integrating multi-source monitoring data and identifying local anomalies. The technical framework of multi-source data preprocessing, point cloud feature extraction, and time-series prediction in this embodiment can be reused after adaptive adjustments, reducing the R&D costs of monitoring systems for similar projects. Furthermore, with the advancement of intelligent transformation in urban rail transit, the precise displacement data and early warning results output by this method can also be integrated into urban intelligent transportation management platforms, providing data support for tunnel lifecycle health assessments and operation and maintenance decision optimization, and helping to build an integrated "monitoring-early warning-response" underground engineering safety management system.

[0046] Example 2 Please refer to Figure 4 This embodiment 2 provides a subway tunnel arch displacement monitoring system based on multi-source data fusion, including: The multi-source data preprocessing unit is used to realize the spatiotemporal synchronous acquisition of multi-source arch data in the preprocessing process, convert non-point cloud data into point cloud format and unify it to the tunnel construction coordinate system, and then expand the dataset size by performing operations on the original point cloud through the 3D PointNet++ front-end module, including random rotation, Gaussian noise addition, and local clipping. The displacement feature value output unit is used by the 3D PointNet++ model to perform end-to-end processing on the pre-processed hybrid point cloud data. It filters key points through the sampling layer, fuses local and global information through the feature aggregation layer, and dynamically allocates weights through the attention mechanism layer, and outputs the fused arch displacement feature value. Anomaly identification and correction unit is used to input displacement time series data and environmental parameters into the LSTM-Transformer hybrid model. It captures long-period patterns of displacement data through the LSTM layer, identifies local anomalies through the Transformer layer and jointly outputs the anomaly probability, and then generates anomaly data correction values ​​based on the time series distribution characteristics of normal data. The displacement early warning response unit is used by the LSTM-Transformer model to predict the displacement and rate change trend of the arch using the corrected effective data, and to trigger different levels of early warning and execute corresponding response operations based on the prediction results.

[0047] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of a UAV inspection method for bridge force-measuring supports.

[0048] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0049] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0050] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion, characterized in that, include: S1. The preprocessing process realizes the spatiotemporal synchronous acquisition of multi-source arch data, converts non-point cloud data into point cloud format and unifies it to the tunnel construction coordinate system, and then expands the dataset size by performing operations on the original point cloud through the 3D PointNet++ front-end module, including random rotation, Gaussian noise addition, and local clipping. The S2.3D PointNet++ model performs end-to-end processing on the preprocessed hybrid point cloud data. It filters key points through a sampling layer, fuses local and global information through a feature aggregation layer, and dynamically allocates weights through an attention mechanism layer, outputting the fused arch displacement feature value. The S3.LSTM-Transformer hybrid model takes displacement time series data and environmental parameters as input, captures long-period patterns of displacement data through LSTM layer, identifies local anomalies through Transformer layer and jointly outputs anomaly probability, and then generates anomaly correction values ​​based on the time series distribution characteristics of normal data. The S4.LSTM-Transformer model uses the corrected effective data to predict the trend of crown displacement and rate change, and triggers different levels of early warning and executes corresponding response operations based on the prediction results. The specific process by which the Transformer layer in S3 identifies local anomalies and jointly outputs the anomaly probability is as follows: Let the hidden state sequence output by the LSTM layer be... , For time steps, It is the first In the hidden state at any given time, the Transformer layer identifies local anomalies and outputs the anomaly probability through the following process: Multi-head attention correlation calculation; for each attention head, Mapped to query vector Key vector Value vector Calculate any two time points and Relevance: in, As a vector dimension, through the correlation matrix Capture temporal correlations; Local deviation metric; for time Calculate its average correlation with all other times. Define the deviation : The greater the deviation, the more significant the difference between the data at that moment and the overall pattern. Anomaly probability generation; combining the long-period regularity characteristics of the LSTM layer output, time intervals are calculated. abnormal probability : in, The denominator is defined as the radius of the time window, and the minimum correlation between neighboring times is introduced into the denominator to strengthen the influence of local continuity on anomaly detection; when When the threshold is exceeded based on the distribution of historical normal data, it is marked as a local anomaly.

2. The method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion according to claim 1, characterized in that, The multi-source arch data in S1 includes: three-dimensional point cloud data of the arch collected by a 3D laser scanner, distance change data of specific points on the arch collected by a laser rangefinder, coordinate data of characteristic points on the arch collected by a total station, and displacement data of tunnel segment joints collected by a displacement sensor.

3. The method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion according to claim 1, characterized in that, The acquisition process of multi-source arch data in S1 is as follows: synchronous acquisition of the arch monitoring area is carried out by 3D laser scanner, laser range sensor, total station and displacement sensor, and the acquisition time of each device is aligned by time synchronization mechanism to form a multi-dimensional raw data set including three-dimensional point cloud, distance change, coordinate and joint displacement. The time synchronization mechanism is as follows: a unified time reference signal is sent to all acquisition devices through a time synchronization module deployed in the tunnel monitoring system. Each device calibrates its internal clock according to the reference signal, so that all data acquisition operations are performed on the same time scale, thus completing the matching of the acquisition times of multi-source data.

4. The method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion according to claim 1, characterized in that, The specific process of unifying to the tunnel construction coordinate system in S1 is as follows: Let the coordinates of the original point cloud data in the coordinate system of the 3D laser scanner be... The initial coordinates of the collected non-point cloud data after conversion are: , Representing different non-point cloud data types, the origin of the tunnel construction coordinate system is... Its X-axis is along the tunnel axis, its Y-axis is perpendicular to the tunnel axis in the horizontal direction, and its Z-axis is along the tunnel vertical direction; First, by using pre-set known coordinate reference points within the tunnel, the transformation parameters between the coordinate systems of each acquisition device and the tunnel construction coordinate system are obtained; the origin of the 3D laser scanner coordinate system is set relative to the origin of the tunnel construction coordinate system. The translation vector is The rotation matrix between the two coordinate systems is Then the original point cloud data is transformed into coordinates in the tunnel construction coordinate system. satisfy: For non-point cloud data, the initial coordinates are first determined based on the spatial geometric relationship between the sensor installation location and the reference point. Translation vector relative to the tunnel construction coordinate system and rotation matrix Coordinates after transformation satisfy: By performing matrix transformations and translation operations based on spatial geometric relationships, multi-source data can be unified into the tunnel construction coordinate system, ensuring data consistency under the same spatial reference.

5. The method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion according to claim 1, characterized in that, The method for selecting key points in S2 is as follows: Let the hybrid point cloud dataset be... ,in The first in the representative point cloud The coordinates of a point in the tunnel construction coordinate system. This represents the total number of point clouds; key point selection is achieved through the following process: Initialize the key point set From point cloud datasets Randomly select a point As the initial key point, at this time ; Iterative selection of key points; for points not included in the point cloud dataset. any point Calculate the relationship between this point and Calculate the Euclidean distances to all selected key points and take the minimum value as the distance from that point to the target key. The distance is given by the formula: in, For point Three-dimensional coordinates in the tunnel construction coordinate system; It is a set of key points Any key point in the data; each iteration selects... The largest point Add S, update ; Determine the screening termination condition; when the number of points in S satisfies the condition of being able to completely represent the geometric profile of the vault, or when new points are added. When the value is less than the minimum structural feature size of the arch, the iteration stops, and S is the set of key points selected.

6. The method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion according to claim 1, characterized in that, The method for calculating the characteristic value of the arch displacement in S2 is as follows: Let the set of key points obtained after screening be... ,in For the first The current coordinates of the key points in the tunnel construction coordinate system, and their coordinates at the initial acquisition time are: ; For each key point First, calculate its absolute displacement vector in three-dimensional space. : Then, a feature aggregation layer is used to extract all points in the neighborhood of that point. The relative displacement characteristics, for The local neighborhood point set, defining the points within the neighborhood. relatively displacement deviation for: in, For point The absolute displacement vector, The Euclidean norm of a vector; All The mean of is used as a local consistency index, and Combined to form the comprehensive feature vector of this key point: in, Indicates the first The comprehensive feature vector of each key point is used to characterize the displacement and local consistency information of that key point. Indicates the number of points in the neighborhood; Finally, for all key locations Spatial correlation analysis is performed, and the overall change trend of each dimension is taken as the final arch displacement characteristic value, which fully reflects the deformation state of the arch in three-dimensional space and the correlation characteristics between the local and the whole.

7. The method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion according to claim 1, characterized in that, The LSTM layer in S3 is specifically as follows: Let S2 output the time series of arch displacement characteristic values. , The time step corresponds to the sequence of environmental parameters at that time. Including influencing factors such as temperature and humidity, the LSTM layer processes timing information through a three-gating mechanism: The input gate is based on the current feature value. With environmental parameters fusion vector The formula for determining whether new information should be included in the cell state is as follows: ,in, Indicates the input gate output of the LSTM layer. For the sigmoid function, This represents the cell state at the previous moment. , The gate matrix is ​​generated during model training. The forget gate is based on the state of historical cells. With current input Preserving long-term patterns, the calculation formula is: , , Forget gate matrix, Indicates the output of the forget gate; Cell status updated to , For element-wise product, Generate candidate states. The state matrix, This represents the current state of the cell. The output gate combines the updated cell state with the current input to generate the hidden state. ,in, This represents the weight matrix of the output gate. This represents the cyclic weight matrix of the output gate. It incorporates both historical cycle characteristics and current environmental influences, and is transmitted to the Transformer layer as an output of long-term cycle patterns.

8. A subway tunnel arch displacement monitoring system based on multi-source data fusion, used to implement the subway tunnel arch displacement monitoring method based on multi-source data fusion as described in claim 1, characterized in that, include: The multi-source data preprocessing unit is used to realize the spatiotemporal synchronous acquisition of multi-source arch data in the preprocessing process, convert non-point cloud data into point cloud format and unify it to the tunnel construction coordinate system, and then expand the dataset size by performing operations on the original point cloud through the 3D PointNet++ front-end module, including random rotation, Gaussian noise addition, and local clipping. The displacement feature value output unit is used by the 3D PointNet++ model to perform end-to-end processing on the pre-processed hybrid point cloud data. It filters key points through the sampling layer, fuses local and global information through the feature aggregation layer, and dynamically allocates weights through the attention mechanism layer, and outputs the fused arch displacement feature value. Anomaly identification and correction unit is used to input displacement time series data and environmental parameters into the LSTM-Transformer hybrid model. It captures long-period patterns of displacement data through the LSTM layer, identifies local anomalies through the Transformer layer and jointly outputs the anomaly probability, and then generates anomaly data correction values ​​based on the time series distribution characteristics of normal data. The displacement early warning response unit is used by the LSTM-Transformer model to predict the displacement and rate change trend of the arch using the corrected effective data, and to trigger different levels of early warning and execute corresponding response operations based on the prediction results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-7: the method for monitoring the displacement of the arch of a subway tunnel based on multi-source data fusion.

Images