A GNSS-based self-driven high-precision sensing method and device for the main cable of a long bridge

Abstract

This invention belongs to the field of bridge main cable structural health monitoring, and provides a GNSS-based self-driven high-precision sensing method and device for the main cable of long bridges. The method includes acquiring raw GNSS displacement data, performing time-series preprocessing using the Median Absolute Deviation (MAD) algorithm; filtering out unstructured noise such as multipath effects using a CNN-LSTM hybrid model; constructing a baseline response model using an adversarial spatiotemporal Transformer network to accurately separate abnormal signals; adaptively adjusting the sampling and communication frequencies according to environmental conditions; and implementing online model optimization through an anomaly-driven incremental learning mechanism. The device includes data acquisition and preprocessing, edge intelligent computing, energy-saving management, and incremental learning modules. This invention achieves millimeter-level high-precision sensing, low-power self-driven operation, and long-term evolution capabilities, significantly improving the intelligence level of bridge safety monitoring.

Description

Technical Field

[0001] This invention belongs to the field of bridge structural health monitoring technology, specifically relating to a GNSS-based self-driven high-precision sensing method and device for the main cable of long bridges. Background Technology

[0002] Long bridges, such as cross-sea bridges and suspension bridges, are core hubs of modern transportation networks. Their main cables, as critical load-bearing components, directly affect the safety and durability of the overall structure. Long-term health monitoring is an important means of preventing catastrophic accidents and extending the life of bridges. However, traditional monitoring methods have many shortcomings: manual inspection is inefficient, relies on periodic inspections, cannot achieve real-time and continuous monitoring, and is highly subjective, making it difficult to quantify minor damage; wired sensor networks, while providing some data, are complex to deploy and have limited coverage, especially for main cables several kilometers long, where wiring costs are high, they are susceptible to environmental corrosion, and it is difficult to detect hidden damage. Statistics show that the annual cost of subsequent repairs due to inadequate main cable monitoring reaches hundreds of millions of yuan, while the socio-economic losses and safety risks caused by traffic interruptions are staggering. With technological advancements, Global Navigation Satellite Systems (GNSS) have been introduced into the field of displacement monitoring, but conventional GNSS technology still faces bottlenecks. Monitoring data is susceptible to interference from non-structural noise such as multipath effects and atmospheric delays, resulting in low signal-to-noise ratios. Centimeter-level noise often masks the true structural response at the millimeter level, making early damage diagnosis extremely difficult. Furthermore, achieving high precision requires continuous high-frequency sampling (e.g., 20Hz) and real-time data transmission, resulting in significant power consumption. In remote scenarios such as cross-sea bridges, power supply difficulties make long-cycle, low-power self-driven monitoring an urgent industry need. Existing technologies lack intelligent noise filtering and adaptive energy-saving mechanisms, leading to short battery life and insufficient accuracy in monitoring systems, failing to meet the intelligent management requirements of long bridge main cables. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a GNSS-based self-driven high-precision sensing method and device for the main cable of long-distance bridges, particularly suitable for long-term health monitoring of the main cable of cross-sea bridges or suspension bridges. This solves the problems of low monitoring efficiency, limited coverage, high noise interference, and high power consumption in existing technologies. The specific technical solution adopted is as follows:

[0004] In a first aspect, the present invention provides a GNSS-based self-driven high-precision sensing method for the main cable of a long bridge, comprising the following steps:

[0005] Raw displacement data of the main cable monitoring points are acquired by a Global Navigation Satellite System (GNSS) receiver based on real-time dynamic differential positioning (RTK); the raw displacement data is preprocessed in a time series manner, including data alignment, outlier removal, and missing value imputation, to form a regular displacement time series data.

[0006] The preprocessed displacement time series data is input into a CNN-LSTM hybrid deep learning model deployed on edge computing nodes to filter out unstructured noise caused by multipath effects and atmospheric delay, resulting in denoised displacement data.

[0007] The denoised displacement data is input into the baseline response model constructed by the adversarial spatiotemporal Transformer network to calculate the baseline displacement response under normal load. The difference between the denoised displacement data and the baseline displacement response is calculated to separate the abnormal signals triggered by structural damage or abnormal events.

[0008] The intelligent energy-saving management unit adaptively adjusts the sampling frequency of the GNSS receiver and the communication frequency of the LoRa wireless communication module based on the characteristics of the environmental state sensor data and the reference displacement response.

[0009] When the abnormal signal is detected, an incremental learning mechanism is triggered to add the newly detected abnormal sample to the training set and to fine-tune and optimize the baseline response model online.

[0010] The outlier removal process in the time-series preprocessing is as follows:

[0011] For the original displacement sequence Using a statistical method based on the median absolute deviation (MAD), if a certain data point... satisfy If it is an outlier, it will be considered an outlier and removed. .

[0012] In the CNN-LSTM hybrid deep learning model:

[0013] The convolutional neural network (CNN) part is used to extract spatial features from the displacement data, and its convolution operation is represented as: ;

[0014] in, For the first The input feature map of the layer, and The first The kernel weights and biases of the layers. This represents the convolution operation. For activation functions;

[0015] The Long Short-Term Memory (LSTM) network is used to capture the time dependencies in the displacement data, and its cell state update formula is as follows:

[0016] ;

[0017] in, Input for the current time step. This is the hidden state from the previous time step. , These are the activation vectors for the forget gate, input gate, and output gate, respectively. The current cell state, This indicates element-wise multiplication.

[0018] The adversarial spatiotemporal Transformer network includes a spatiotemporal encoder and a discriminator, wherein:

[0019] The spatiotemporal encoder is used to learn the spatiotemporal characteristics of the main cable displacement under normal load conditions. The calculation formula for its multi-head self-attention mechanism is as follows:

[0020] ;

[0021] ;

[0022] ;

[0023] in, These are query, key, and value matrices, respectively. Let be the dimension of the key vector. , , , It is a learnable parameter matrix;

[0024] The discriminator is used to distinguish between the true normal displacement sequence and the reconstructed sequence, and its loss function is adversarial loss: ;

[0025] Where D is the discriminator and G is the generator. This is a normal sample displacement. These are latent variables.

[0026] The process of adaptively adjusting the sampling frequency and the communication frequency is as follows:

[0027] When abnormal signals or sudden changes in environmental conditions are detected, increase the GNSS sampling frequency and LoRa communication frequency;

[0028] When the environment is stable and the main cable response is within the normal range, reduce the GNSS sampling frequency and LoRa communication frequency to save energy.

[0029] The online fine-tuning and optimization process is as follows:

[0030] When a new abnormal sample is identified, its features are extracted and stored in the edge cache;

[0031] The baseline response model is fine-tuned using new samples from the cache, and the model update employs gradient descent. ;

[0032] in, For the current parameters of the model, Learning rate, loss function To be a weighted sum of reconstruction loss and feature loss, This is a new anomalous sample.

[0033] Secondly, the present invention also provides a GNSS-based self-driven high-precision sensing device for the main cable of a long bridge, comprising:

[0034] The data acquisition and preprocessing module is used to acquire the raw displacement data of the main cable monitoring point obtained by the RTK-based GNSS receiver, and to perform time-series preprocessing on the raw displacement data.

[0035] The edge intelligent computing module is used to filter out unstructured noise from the preprocessed displacement time series data using a CNN-LSTM hybrid deep learning model, and to build a baseline response model using an adversarial spatiotemporal Transformer network to separate normal displacement from abnormal signals.

[0036] The energy-saving management module is used to adaptively adjust the GNSS sampling frequency and LoRa communication frequency based on environmental conditions and main cable response characteristics.

[0037] The incremental learning module is used to trigger the incremental learning mechanism when an abnormal signal is detected, and to optimize the baseline response model online.

[0038] Thirdly, the present invention also provides a computer-readable storage medium, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0039] The beneficial effects of the technical solution of this invention mainly include:

[0040] 1. The monitoring accuracy is significantly improved. The CNN-LSTM hybrid model effectively filters out centimeter-level unstructured noise in GNSS data, so that the millimeter-level real structural response can be clearly presented, which solves the core pain point of the low signal-to-noise ratio of traditional methods.

[0041] 2. Breakthrough in anomaly diagnosis capabilities: Utilizing the powerful spatiotemporal modeling capabilities of the adversarial spatiotemporal Transformer network, a high-fidelity baseline response model was constructed, which can extremely sensitively separate abnormal signals caused by damage, achieving a leap from "monitoring" to "intelligent diagnosis".

[0042] 3. Intelligent balance between power consumption and performance: Through an adaptive sampling and communication strategy driven by environmental and response characteristics, the system is in a "sleep" mode with extremely low power consumption in most stable states, and only activates a high-performance mode when necessary. This makes it possible to achieve long-term continuous monitoring for several years under solar power conditions, thus solving the deployment power supply problem.

[0043] 4. The system has evolutionary capabilities. Its unique anomaly-driven incremental learning mechanism enables the system to continuously optimize its own model by utilizing newly discovered anomaly samples, thereby adapting to the degradation of bridge structures over time and unknown new failure modes, significantly improving the long-term applicability and robustness of the system. Attached Figure Description

[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0045] Figure 1 The diagram shows the overall architecture of the system, illustrating the connection relationships between the GNSS displacement monitoring module, edge intelligent computing module, energy-saving management module, and incremental learning module.

[0046] Figure 2 A disassembly diagram of the bridge main cable monitoring device shows the sensor deployment locations and hardware components;

[0047] Figure 3 The diagram illustrates the calculation process of the multi-head self-attention mechanism in Transformer attention.

[0048] Figure 4 This is an energy-saving management status transition diagram, which describes the transition conditions between normal, early warning, and alarm states.

[0049] Figure 5 This is a schematic diagram of incremental learning updates, illustrating the process of storing outlier samples and fine-tuning the model;

[0050] Figure 6 The displacement monitoring curve of the main cable of the long bridge was used to compare the displacement data and abnormal signals before and after noise reduction.

[0051] Figure 7 The results and comparison charts show the performance test results of the device system, including data comparing accuracy, power consumption, and traditional methods. Detailed Implementation

[0052] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0053] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 Please refer to the overall architecture diagram provided in this embodiment of the invention. Figure 2 A disassembly diagram of the bridge main cable monitoring device provided in this embodiment of the invention. The method specifically includes the following steps:

[0054] Step 1: Collect raw displacement data of the main cable monitoring points using a Global Navigation Satellite System (GNSS) receiver based on real-time dynamic differential positioning (RTK); use the median absolute deviation (MAD) algorithm to detect and remove outliers from the raw displacement data, and use the cubic spline interpolation algorithm to fill in missing data to form a regular displacement time series data.

[0055] In one specific embodiment of the present invention, three-dimensional raw displacement data of the main cable monitoring point is acquired by a Global Navigation Satellite System (GNSS) receiver based on Real-Time Dynamic Differential Positioning (RTK) at an adjustable sampling frequency of 1-20Hz. The raw displacement data undergoes time-series preprocessing, which includes data alignment, outlier removal, and missing value imputation. Outlier removal employs a method based on the absolute deviation of the median: if data point X... i Satisfy |X i If -median(X)| > 3 × MAD(X) (MAD(X) = 1.4826 × median(|X - median(X)|)), then it is discarded. Missing value imputation uses cubic spline interpolation algorithm to construct interpolation function based on 4 adjacent data points to form regular displacement time series data.

[0056] Step 2: Input the preprocessed displacement time series data into the CNN-LSTM hybrid deep learning model deployed on the edge computing node. The CNN part uses 32 convolutional kernels of length 3 to extract spatial features, and the LSTM part uses 64 hidden units to capture temporal dependencies, effectively filtering out unstructured noise caused by multipath effects and atmospheric delay, and outputting clean displacement data with improved signal-to-noise ratio.

[0057] In a specific embodiment of the present invention, the edge intelligent computing and noise filtering steps are as follows: the preprocessed displacement time series data is input into a CNN-LSTM hybrid deep learning model deployed on edge computing nodes, wherein the CNN part uses 32 convolutional kernels of length 3 for spatial feature extraction, and the convolution operation is represented as F[i]=max(0,∑ j X[i+j-1] W[j]+b) (step size is 1, activation function is ReLU), the LSTM part uses 64 hidden units to capture time dependencies, the unit state update formula includes forget gate, input gate and output gate calculation, filters out unstructured noise caused by multipath effect and atmospheric delay, and outputs displacement data with improved signal-to-noise ratio.

[0058] Step 3: Please refer to Figure 3 A schematic diagram of Transformer attention computation is shown. An adversarial spatiotemporal Transformer network is constructed as the baseline response model, in which the spatiotemporal encoder adopts an 8-head self-attention mechanism (d). model =512,d k =d v =64) Learn the displacement response law under normal load conditions, optimize the model parameters through adversarial training of the discriminator, and achieve high-precision separation of normal displacement and abnormal signal;

[0059] In a specific embodiment of the present invention, the benchmark modeling and anomaly separation steps involve inputting the denoised displacement data into a benchmark response model constructed by an adversarial spatiotemporal Transformer network. The network includes a spatiotemporal encoder and a discriminator, and the spatiotemporal encoder employs an 8-head self-attention mechanism (d). model =512, d k =d v =64) Learning the spatiotemporal characteristics under normal load conditions, the attention calculation formula is: The discriminator distinguishes between real and reconstructed sequences using an adversarial loss function, calculates the baseline displacement response under conventional loads (such as wind, vehicles, and water flow), and subtracts the denoised displacement data from the baseline displacement response to obtain the residual signal. The discriminator then separates the abnormal signals triggered by structural damage or abnormal events using a dynamic threshold (adaptively adjusted based on historical residual statistics).

[0060] Step 4: Please refer to Figure 4 An energy-saving management state transition diagram is constructed, and an intelligent energy-saving management strategy based on a finite state machine is designed. Based on parameters such as anomaly score (threshold 0.5 / 0.8), wind speed (threshold 20m / s), and temperature change rate (threshold 5°C / h), the GNSS sampling frequency (adjustable 1 / 5 / 20Hz, including discrete or continuous adjustment) and LoRa communication interval (3600 / 600 / 60 seconds) are adaptively adjusted to achieve the optimal balance between power consumption and performance.

[0061] In a specific embodiment of the present invention, the adaptive energy-saving management step is as follows: The intelligent energy-saving management unit, based on the characteristics of environmental state sensor data (such as wind speed and temperature change rate) and the reference displacement response (such as anomaly score), adaptively adjusts the sampling frequency of the GNSS receiver (1Hz for normal state, 5Hz for warning state, and 20Hz for alarm state) and the communication frequency of the LoRa wireless communication module (intervals of 3600s, 600s, and 60s, respectively) through a finite state machine. The state transition conditions include: if the anomaly score is >0.8 or the wind speed is >20m / s, the alarm state is entered; if the anomaly score is >0.5 or the temperature change rate is >5°C / h, the warning state is entered; otherwise, the normal state is maintained.

[0062] Step 5: Please refer to Figure 5 The incremental learning update flowchart establishes an anomaly-driven incremental learning mechanism. When a new anomaly sample is identified, it is stored in an edge cache with a fixed capacity (1000 samples). The baseline response model is fine-tuned online using a stochastic gradient descent algorithm with a momentum of 0.9 (learning rate 0.001) to continuously improve the adaptability to unknown fault modes.

[0063] In one specific embodiment of the present invention, when the abnormal signal is detected, an incremental learning mechanism is triggered, the newly detected abnormal samples are stored in the edge cache (fixed capacity 1000 samples, using a circular queue strategy), and the baseline response model is fine-tuned online using a stochastic gradient descent algorithm with a momentum of (0.9) and a learning rate of 0.001, with the loss function being L=L reconstruction +0.5×L feature The batch size is 32 and the number of training rounds is 50 to improve the ability to adapt to unknown failure modes.

[0064] The specific implementation process for each step is given below:

[0065] Step 1: Data acquisition and preprocessing for displacement monitoring of long bridges;

[0066] This example uses a U-blox ZED-F9P RTK GNSS receiver to collect three-dimensional displacement data of the main cable monitoring point at an adjustable sampling frequency of 1-20Hz, achieving a positioning accuracy of millimeters.

[0067] For the original displacement sequence X=[x1,x2,...,x n The outlier detection process is as follows:

[0068] Calculate the median of the sequence: median = median(X);

[0069] Calculate the absolute deviation MAD = 1.4826 × median(|X - median|);

[0070] Set the upper limit of the anomaly threshold to median+3×MAD and the lower limit to median-3×MAD; remove abnormal data points that exceed this threshold range.

[0071] The missing data was filled using a cubic spline interpolation algorithm.

[0072] Specifically for missing positions A cubic spline function is constructed based on four adjacent data points;

[0073] The system vector is determined by solving a system of linear equations. .

[0074] Step 2: Calculation of the CNN-LSTM hybrid deep learning intelligent noise cancellation model;

[0075] The CNN feature extraction part processes the input as a preprocessed displacement sequence with shape (n,1,1).

[0076] Convolution operation: ;

[0077] The convolution kernel W has a size of 3, a number of 32, a stride of 1, and the activation function is ReLU.

[0078] The LSTM time series modeling part is used to capture long-term dependencies in the sequence:

[0079] The cell state update formula is as follows:

[0080] ;

[0081] The hidden unit number is 64, the forget gate bias is initialized to 1.0, and the timing information is transmitted through the unit state update mechanism.

[0082] Step 3: Abnormal signal separation;

[0083] A multi-head self-attention mechanism is employed for signal separation. Input dimension d model =512, number of heads h=8, each dimension d k =d v =64;

[0084] For each head i, the calculation formula is: ;

[0085] in , The learnable parameter matrix; the output is .

[0086] During adversarial training, the generator loss function includes reconstruction error and a regularization term, while the discriminator loss function is used to distinguish between real and generated signals. The formula used for the generator loss is: +λ ;

[0087] The formula used for discriminator loss is: - ;

[0088] The training rounds are 100, and the batch size is 32.

[0089] Step 4: Adaptive energy-saving management;

[0090] Design a state machine to achieve adaptive energy-saving management, specifically including normal state, early warning state, and alarm state:

[0091] Normal conditions: GNSS sampling rate 1Hz, LoRa communication interval 3600s;

[0092] Warning status: GNSS sampling rate 5Hz, LoRa communication interval 600s;

[0093] Alarm status: GNSS sampling rate 20Hz, LoRa communication interval 60s.

[0094] The state transition conditions are determined based on parameters such as anomaly score, wind speed, and temperature change rate, and the specific rules are as follows:

[0095] If the abnormal score is greater than 0.8 or the wind speed exceeds 20 m / s, the state machine switches to the alarm state.

[0096] If the alarm conditions are not met, but the anomaly score is greater than 0.5 or the temperature change rate is greater than 5°C / h, then switch to the warning state.

[0097] If none of the above conditions are met, the system will remain in normal condition.

[0098] Step 5: Incremental learning optimization;

[0099] The sample storage uses a fixed-capacity circular queue structure, with a maximum sample size of 1000. When the queue is full, the earliest-added sample is automatically evicted.

[0100] During online learning, the Stochastic Gradient Descent (SGD) algorithm with momentum was used as the optimizer, with a learning rate of 0.001 and a momentum parameter of 0.9.

[0101] The loss function is defined as the weighted sum of the reconstruction error and the feature loss, i.e., L = L reconstruction +0.5×L feature ; where L reconstruction For reconstruction error (such as mean square error), L feature The loss function is for features (such as cosine similarity); the batch size during training is 32, and the number of training rounds is 50.

[0102] Please see Figure 6 The main cable displacement monitoring curve graph for long bridges outputs monitoring curves under normal and abnormal scenarios. Please refer to [link / reference]. Figure 7 A field deployment test comparison of a GNSS-based high-precision sensing device for the main cable of a long bridge, used to implement the method.

[0103] This invention also provides a GNSS-based self-driven high-precision sensing device for the main cable of a long bridge. The device includes a data acquisition and preprocessing module, an edge intelligent computing module, an energy-saving management module, and an incremental learning module.

[0104] The data acquisition and preprocessing module is used to acquire the raw displacement data of the main cable monitoring point obtained by the RTK-based GNSS receiver at an adjustable sampling frequency of multiple levels (such as 1Hz, 5Hz, 20Hz), and to perform time-series preprocessing on the raw displacement data.

[0105] The edge intelligent computing module is used to filter out unstructured noise from the preprocessed displacement time series data using a CNN-LSTM hybrid deep learning model, and to build a baseline response model using an adversarial spatiotemporal Transformer network to separate normal displacement from abnormal signals.

[0106] The energy-saving management module is used to adaptively adjust the GNSS sampling frequency and LoRa communication frequency based on environmental conditions and main cable response characteristics;

[0107] The incremental learning module is used to trigger the incremental learning mechanism when an abnormal signal is detected, and to optimize the baseline response model online.

[0108] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described.

[0109] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A GNSS-based self-driven high-precision sensing method for the main cable of a long bridge, characterized in that, Includes the following steps: Step 1: Collect raw displacement data of the main cable monitoring points obtained by the Global Navigation Satellite System receiver based on real-time dynamic differential positioning; The original displacement data undergoes time-series preprocessing, which includes data alignment, outlier removal, and missing value imputation to form a regular displacement time series data. Step 2: Input the preprocessed displacement time series data into the CNN-LSTM hybrid deep learning model deployed on the edge computing nodes to filter out unstructured noise caused by multipath effects and atmospheric delay, and obtain the denoised displacement data. Step 3: Input the denoised displacement data into the baseline response model constructed by the adversarial spatiotemporal Transformer network to calculate the baseline displacement response under normal load. Subtract the denoised displacement data from the baseline displacement response to separate anomalous signals triggered by structural damage or abnormal events. The adversarial spatiotemporal Transformer network includes a spatiotemporal encoder and a discriminator. The spatiotemporal encoder is used to learn the spatiotemporal characteristics of the main cable displacement under normal load conditions. The calculation formula for its multi-head self-attention mechanism is as follows: ； in, These are query, key, and value matrices, respectively. Let be the dimension of the key vector. , , , It is a learnable parameter matrix; The discriminator is used to distinguish between the true normal displacement sequence and the reconstructed sequence. Its loss function is adversarial loss, and the formula is: ; Where D is the discriminator and G is the generator. This is a normal sample displacement. As latent variables; Step 4: Based on the characteristics of the environmental state sensor data and the reference displacement response, adaptively adjust the sampling frequency of the GNSS receiver and the communication frequency of the LoRa wireless communication module. Step 5: When an abnormal signal is detected, the incremental learning mechanism is triggered to add the newly detected abnormal samples to the training set and fine-tune and optimize the baseline response model online.

2. The GNSS-based self-driven high-precision sensing method for the main cable of long bridges according to claim 1, characterized in that, The outlier removal process in time-series preprocessing includes applying a statistical method based on the absolute deviation of the median to the original displacement sequence. If a data point meets the condition that the absolute deviation from the median of the sequence is greater than three times the absolute deviation of the median, it is considered an outlier and removed. The calculation of the absolute deviation of the median is based on the median of the absolute difference between the sequence data and the median.

3. The GNSS-based self-driven high-precision sensing method for the main cable of long bridges according to claim 1, characterized in that, Missing value imputation uses cubic spline interpolation algorithm, which constructs an interpolation function based on adjacent data points to fill in the missing values; the convolutional neural network part in the CNN-LSTM hybrid deep learning model is used to extract spatial features in the displacement data, and its convolution operation uses multiple convolutional kernels to perform feature mapping and outputs through activation function; The Long Short-Term Memory (LSTM) network is used to capture the temporal dependencies in the displacement data. Its cell state updates include the calculation of forget gates, input gates, and output gates, as well as the updating of cell states.

4. The GNSS-based self-driven high-precision sensing method for the main cable of long bridges according to claim 1, characterized in that, The sampling frequency and communication frequency are adaptively adjusted in the following ways: When the edge intelligent computing module detects abnormal signals or sudden changes in environmental conditions, it increases the GNSS sampling frequency and the LoRa communication frequency. When the environment is stable and the main cable response is within the normal range, reduce the GNSS sampling frequency and LoRa communication frequency to save energy.

5. The GNSS-based self-driven high-precision sensing method for the main cable of long bridges according to claim 4, characterized in that, The adaptive adjustment is implemented based on a finite state machine, and the states include normal state, early warning state, and alarm state. The normal state corresponds to a global navigation satellite system sampling frequency of 1Hz and a LoRa communication interval of 3600 seconds; The aforementioned warning status corresponds to a global navigation satellite system sampling frequency of 5Hz and a LoRa communication interval of 600 seconds; The alarm status corresponds to a global navigation satellite system sampling frequency of 20Hz and a LoRa communication interval of 60 seconds.

6. The GNSS-based self-driven high-precision sensing method for the main cable of long bridges according to claim 1, characterized in that, The online fine-tuning and optimization includes extracting the features of a new abnormal sample and storing them in the edge cache when a new abnormal sample is identified. The baseline response model is fine-tuned using new samples from the cache. The model update employs gradient descent, and the formula is as follows: ; Where the loss function To be a weighted sum of reconstruction loss and feature loss, For the current parameters of the model, For learning rate, This is a new anomalous sample; The updated model is used for subsequent anomaly detection, improving its adaptability to unknown failure modes.

7. A GNSS-based self-driven high-precision sensing device for the main cable of a long bridge, characterized in that, The apparatus for the GNSS-based self-driven high-precision sensing method for the main cable of a long bridge as described in any one of claims 1-6 includes: The data acquisition and preprocessing module is used to acquire the raw displacement data of the main cable monitoring point obtained by the RTK-based GNSS receiver, and to perform time-series preprocessing on the raw displacement data. The edge intelligent computing module filters out non-structural noise from the preprocessed displacement time series data using a CNN-LSTM hybrid deep learning model, and constructs a baseline response model using an adversarial spatiotemporal Transformer network to separate normal displacement from anomalous signals. The adversarial spatiotemporal Transformer network includes a spatiotemporal encoder and a discriminator. The spatiotemporal encoder learns the spatiotemporal characteristics of the main cable displacement under normal load conditions, and its multi-head self-attention mechanism is calculated using the following formula: ； in, These are query, key, and value matrices, respectively. Let be the dimension of the key vector. , , , It is a learnable parameter matrix; The discriminator is used to distinguish between the true normal displacement sequence and the reconstructed sequence. Its loss function is adversarial loss, and the formula is: ; Where D is the discriminator and G is the generator. This is a normal sample displacement. As latent variables; The energy-saving management module is used to adaptively adjust the GNSS sampling frequency and LoRa communication frequency based on environmental conditions and main cable response characteristics. The incremental learning module is used to trigger the incremental learning mechanism when an abnormal signal is detected, and to optimize the baseline response model online.

8. The apparatus according to claim 7, characterized in that, The edge intelligent computing module employs a hybrid deep learning model: a convolutional neural network-long short-term memory network model, wherein: The convolutional neural network part is used to extract spatial features from the displacement data, and its convolution operation is represented as follows: ; in, For the first The input feature map of the layer, and The first The kernel weights and biases of the layers. This represents the convolution operation. For activation functions; The Long Short-Term Memory (LSTM) network is used to capture the time dependencies in the displacement data, and its cell state update formula is as follows: ； in, Input for the current time step. This is the hidden state from the previous time step. , These are the activation vectors for the forget gate, input gate, and output gate, respectively. The current cell state, This indicates element-wise multiplication.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the GNSS-based self-driven high-precision sensing method for the main cable of long bridges as described in any one of claims 1-6.

Images