A collision-sensing intelligent guardrail system and method based on BeiDou positioning

Abstract

This invention discloses a collision-sensing intelligent guardrail system and method based on BeiDou positioning, comprising the following steps: S1, collecting sensor data along the guardrail and preprocessing it; S2, transmitting the preprocessed sensor data to an edge server and forwarding it to a cloud server via a cellular network; S3, constructing a dual-channel gated convolutional network and inputting the state vector sequence into the gated convolutional network; S4, performing multi-layer nested encoding and decoding operations on the collision feature vector and constructing a collision determination function to determine whether a collision anomaly is triggered; S5, generating a collision warning data packet when the collision level label is active; S6, transmitting the collision alarm data packet to the road traffic command center for rapid response and handling by road maintenance personnel. This invention achieves high-precision real-time perception and intelligent early warning of road guardrail collision status, significantly improving the automation level and response efficiency of traffic safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing and embedded sensing and control technology, and in particular to a collision-sensing intelligent guardrail system and method based on BeiDou positioning. Background Technology

[0002] Against the backdrop of rapid urbanization and high-speed transportation development, road traffic safety has increasingly attracted widespread attention from all sectors of society. Especially on highways, urban elevated roads, and mountain roads characterized by high speeds and complex terrain, guardrails, as the core protective facility of physical safety boundaries, play a crucial role in preventing vehicles from running off the road, mitigating the impact of accidents, and guiding traffic direction. However, with increasing road traffic density and the frequency of extreme driving behaviors, the probability of guardrail collision accidents is constantly rising, thus placing higher demands on traffic flow, the safety of people and property, and road maintenance. To improve the intelligent monitoring capabilities of guardrail safety status, scholars and engineering practitioners both domestically and internationally have proposed a series of guardrail collision detection and response methods.

[0003] Existing guardrail detection technologies primarily rely on installing various physical sensors on the guardrail itself, such as strain gauges, vibration sensors, pressure sensors, cameras, and millimeter-wave radar. These sensors detect the physical changes in the guardrail in real time when subjected to impact or displacement, and transmit the information to a backend server for analysis via a wireless network. In addition, some systems supplement this with manual video surveillance to ensure rapid response to emergencies. This traditional approach offers advantages in real-time performance and completeness during engineering implementation, enabling the monitoring and assessment of guardrail conditions in specific areas, and has played a significant role in handling some traffic accidents.

[0004] However, guardrail collision detection systems based on the aforementioned traditional methods also face several limitations. Firstly, they involve dense hardware deployment and high construction costs. To ensure detection accuracy, multiple sensor nodes and cameras are often densely deployed, especially on highways stretching tens or even hundreds of kilometers, significantly increasing initial investment costs. Furthermore, these devices have high requirements for power supply, signal transmission, and protection levels, further increasing the economic burden of operation and maintenance. Secondly, the system's communication links are complex, and data transmission is significantly limited. Most existing equipment relies on Wi-Fi or cellular networks to transmit sensor and video data. In complex terrains such as mountainous areas and tunnels where network signals are unstable or bandwidth is limited, data suffers from packet loss, delays, and bit errors, severely impacting the system's real-time performance and reliability.

[0005] Secondly, traditional systems generally employ static rule matching or manual threshold recognition strategies in their data processing, which cannot dynamically adapt to diverse road conditions and sudden impact characteristics. This method, relying on manual experience to set parameters, not only suffers from high false alarm rates and severe missed alarms, but also lacks proactive early warning and trend judgment capabilities, resulting in a delayed response mechanism and making it difficult for traffic control centers to take rapid and accurate intervention measures. This is particularly true in complex situations requiring accurate identification of guardrails subjected to minor deformation, continuous vibration disturbances, or hidden damage, where traditional methods are severely inadequate. Furthermore, some systems still rely on manual video verification, further increasing the workload of personnel and increasing the risk of fatigue-induced judgments, misjudgments, and other problems.

[0006] Therefore, how to provide a collision-sensing intelligent guardrail system and method based on BeiDou positioning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a collision-aware intelligent guardrail system and method based on BeiDou positioning. This invention fully utilizes BeiDou high-precision positioning technology, low-power accelerometer, extended Kalman filter algorithm and dual-channel gated convolutional network structure, and describes in detail the entire process from data acquisition, state estimation, feature extraction to collision determination. It has the advantages of low deployment cost, fast response speed, high recognition accuracy and adaptability to complex environments.

[0008] A collision-sensing intelligent guardrail method based on BeiDou positioning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect sensor data along the guardrail and perform preprocessing;

[0010] S2. Transmit the preprocessed sensor data to the edge server, perform nonlinear time-series fitting and smoothing filtering operations to obtain the state vector sequence, and then forward it to the cloud server through the cellular network.

[0011] S3. Construct a gated convolutional network based on a dual-channel interface, input the state vector sequence into the gated convolutional network, perform feature extraction and fusion, and output a collision feature vector.

[0012] S4. Perform multi-layer nested encoding and decoding operations on the collision feature vector to obtain a dimensionality-reduced feature representation, and construct a collision determination function to determine whether a collision anomaly is triggered, while also marking the collision level label.

[0013] S5. When the collision level label is active, generate a collision warning data packet, which includes the collision occurrence time, guardrail location coordinates and collision severity level.

[0014] S6. Transmit the collision alarm data packet to the road traffic command center for road maintenance personnel to respond and handle quickly.

[0015] Optionally, the sensor data includes spatial coordinate data and acceleration data. The spatial coordinate data is provided by the GNSS positioning module, and the acceleration data is provided by the MPU6050 sensor module. Each module is connected in series via a bus to form a group, and each group contains 20 to 50 modules.

[0016] Optionally, the preprocessing includes timestamp alignment, coordinate calculation, anomaly drift removal, signal denoising, missing value imputation, and data standardization.

[0017] Optionally, the edge server is equipped with solar panels and batteries to supply power to the entire train convoy. It is installed on the edge of the road, has cellular communication capabilities, and is connected to the train convoy at one end and to a cloud server at the other end.

[0018] Optionally, the gated convolutional network is constructed based on a dual-channel mechanism. The first channel receives acceleration data and extracts time-dependent features, while the second channel receives spatial coordinate data and extracts displacement morphology features. The output features of the two channels are fused in the residual connection structure, and a dynamic attention gating mechanism is introduced to reconstruct the time-dependent features and displacement morphology features with weights to form a collision feature vector.

[0019] Optionally, the collision level label includes situations such as vehicle collision with guardrail, guardrail damage caused by human intervention, road subsidence, significant guardrail displacement, road geological deformation, and no collision.

[0020] Optionally, S2 specifically includes:

[0021] S21. Obtain the spatial coordinate data and acceleration data of each acquisition node in the series structure, wherein the spatial coordinate data is three-dimensional position data. The acceleration data are triaxial acceleration data. , Indicates the index of the time of data collection. Indicates the first Geographic coordinates at any given time This represents the acceleration values ​​in the three orthogonal directions at the corresponding moment;

[0022] S22. Align the 3D coordinates and acceleration data of each acquisition node with timestamps, interpolate missing samples, and perform linear interpolation to unify the sampling frequency, thus constructing an observation sequence set. ;

[0023] S23. Transmit the preprocessed observation sequence set to the edge server, and apply nonlinear time series fitting and extended Kalman filtering algorithms to perform nonlinear state estimation and noise suppression on the observation sequence. After filtering, a state vector sequence is obtained. ,in This represents the estimated current location. This represents the speed estimate. This represents the estimated acceleration value;

[0024] S24. The state vector sequence is forwarded to the cloud server via a cellular network. The nonlinear state estimation uses the following Kalman state prediction and update.

[0025] ;

[0026] in, Indicates the first State estimate at time 10:00 Indicates the first The predicted state value at time 10:00. Indicates the first Kalman gain at time step This represents the time-aligned observation vector. This represents the observation matrix.

[0027] Optionally, S3 specifically includes:

[0028] S31. Construct a gated convolutional network based on a dual-channel configuration, and process the state vector sequence... and These are used as inputs for the first and second channels, respectively.

[0029] S32, will The input is fed into the first channel, and a one-dimensional convolution extraction operation is performed to generate a time-dependent feature tensor. The first channel uses a convolutional kernel with a gated activation function to locally model the acceleration sequence in the temporal dimension;

[0030] S33, will The input is fed into the second channel, where a one-dimensional convolution extraction operation is performed to generate a displacement morphological feature tensor. The second channel models the displacement morphological features in the coordinate sequence through a convolutional structure;

[0031] S34. Convert the output tensors of the two channels. and Fusion is performed within the residual connectivity structure to construct a fused feature tensor. The fusion process introduces a dynamic attention gating mechanism, which forms a collision feature vector through weighted reconstruction:

[0032] ;

[0033] in, Indicates the first The collision feature vector output at time step 1. Indicates the first The time-dependent feature tensor at time step Indicates the first The displacement morphological feature tensor at time step. This represents the attention gating weight matrix. This represents the gating bias vector. This represents the hyperbolic tangent activation function. This represents the Sigmoid activation function.

[0034] Optionally, S4 specifically includes:

[0035] S41, The collision feature vector The input is a nested encoding network structure, which is then subjected to layer-by-layer nonlinear mapping and compression to obtain a multi-layer latent feature representation sequence. ,in This represents the feature vector output by the first layer of encoding. This represents the feature vector output by the second layer of encoding. This represents the feature vector output by the third layer of encoding;

[0036] S42, will The input decoding structure is used to perform reverse mapping and dimensionality reduction reconstruction to generate a reconstructed feature vector. The dimensionality reduction reconstruction process retains the core collision features and compresses redundant information.

[0037] S43. Based on the reconstructed feature vector, construct a collision determination function to determine whether the current time is a collision anomaly state. The expression of the collision determination function is:

[0038] ;

[0039] in, Indicates the first Reconstructed feature vector at time step Denotes the first linear mapping matrix. Denotes the second linear mapping matrix. This represents the first bias vector. This represents the second bias vector. Represents a non-linear activation function. This represents the Softmax normalization function. Indicates the first The collision determination label at any time includes vehicle collision with guardrail, guardrail damage caused by human intervention, road subsidence, guardrail displacement, geological deformation, and no-collision state.

[0040] A collision-sensing intelligent guardrail system based on BeiDou positioning according to an embodiment of the present invention includes:

[0041] The data processing module is used to collect spatial coordinate data and acceleration data along the guardrail and perform data preprocessing operations.

[0042] The edge computing module is used to transmit the preprocessed sensor data to the edge server, perform nonlinear time-series fitting operations and smoothing filtering operations to generate a guardrail state vector sequence, and forward the state vector sequence to the cloud server through the cellular network.

[0043] The feature extraction module is used to construct a gated convolutional network based on a dual-channel interface. The state vector sequence is input into the gated convolutional network, and features are extracted and fused to output a guardrail collision feature vector.

[0044] The collision determination module is used to perform multi-layer nested encoding and decoding operations on the guardrail collision feature vector to obtain a dimensionality-reduced feature representation, and to construct a collision determination function to determine whether a collision anomaly is triggered and to mark the collision level label.

[0045] The early warning generation module is used to generate a collision early warning data packet when the collision level label is active. The collision early warning data packet includes the collision occurrence time, guardrail location coordinates, and collision severity level.

[0046] The response and handling module is used to transmit collision alarm data packets to the road traffic command center for road maintenance personnel to respond and handle quickly.

[0047] The beneficial effects of this invention are:

[0048] First, this invention integrates a GNSS high-precision positioning chip with an MPU6050 three-axis accelerometer to construct a multi-dimensional sensing data acquisition mechanism for guardrail status. Combined with a bus-type serial deployment structure and a distributed computing architecture of edge servers, it achieves high-density, low-power data perception along the guardrail, significantly reducing the overall construction and maintenance costs of the system and overcoming the deployment difficulties caused by traditional solutions that rely on dense camera deployment and high-power sensors.

[0049] Secondly, this invention introduces an extended Kalman filter algorithm on the edge processing side to perform nonlinear state estimation and noise suppression on the original sensor data, effectively improving the stability and reliability of the data. At the same time, based on a dual-channel gated convolutional network structure, it performs parallel modeling and deep fusion of acceleration features and spatial location features respectively, and then extracts core collision semantic information through a nested encoder-decoder mechanism. Combined with a collision determination function, it realizes the classification and recognition of various collision states, which greatly improves the accuracy and intelligence level of collision detection and overcomes the response lag and misjudgment problems of traditional static rules and manual thresholding methods.

[0050] Finally, by combining cellular networks and IoT communication structures, this invention can achieve alarm response within seconds after a collision event, generate structured collision data packets and push them to the traffic command center in real time, enabling rapid location, intelligent classification and remote response to guardrail collision events, significantly improving the efficiency of road safety event handling and decision support capabilities, while providing high-value data support for subsequent road health monitoring and intelligent transportation big data analysis. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a collision-sensing intelligent guardrail method based on BeiDou positioning proposed in this invention;

[0053] Figure 2 This is a schematic diagram illustrating the data processing and forwarding of a collision-sensing intelligent guardrail method based on BeiDou positioning proposed in this invention.

[0054] Figure 3 This is a schematic diagram of the structure of a collision-sensing intelligent guardrail method based on BeiDou positioning proposed in this invention.

[0055] Figure 4 This is a modular structure diagram of a collision-sensing intelligent guardrail system based on BeiDou positioning proposed in this invention.

[0056] Figure 5 This is a global workflow diagram of a collision-sensing intelligent guardrail system based on BeiDou positioning proposed in this invention.

[0057] Figure 6 This is a schematic diagram of the functional modules of a collision-sensing intelligent guardrail system based on BeiDou positioning proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figure 1-3 A collision-sensing intelligent guardrail method based on BeiDou positioning includes the following steps:

[0060] S1. Collect sensor data along the guardrail and perform preprocessing;

[0061] S2. Transmit the preprocessed sensor data to the edge server, perform nonlinear time-series fitting and smoothing filtering operations to obtain the state vector sequence, and then forward it to the cloud server through the cellular network.

[0062] S3. Construct a gated convolutional network based on a dual-channel interface, input the state vector sequence into the gated convolutional network, perform feature extraction and fusion, and output a collision feature vector.

[0063] S4. Perform multi-layer nested encoding and decoding operations on the collision feature vector to obtain a dimensionality-reduced feature representation, and construct a collision determination function to determine whether a collision anomaly is triggered, while also marking the collision level label.

[0064] S5. When the collision level label is active, generate a collision warning data packet, which includes the collision occurrence time, guardrail location coordinates and collision severity level.

[0065] S6. Transmit the collision alarm data packet to the road traffic command center for road maintenance personnel to respond and handle quickly.

[0066] This invention provides a collision-sensing intelligent guardrail method based on BeiDou positioning. It constructs a complete guardrail collision monitoring, feature recognition, early warning generation and response handling process through six consecutive steps, and has the ability to monitor the entire process in a closed loop, which significantly improves the efficiency of automated processing and real-time response to road safety incidents.

[0067] In this embodiment, the sensor data includes spatial coordinate data and acceleration data. The spatial coordinate data is provided by the GNSS positioning module, and the acceleration data is provided by the MPU6050 sensor module. Each module is connected in series via a bus to form a group, and each group contains 20 to 50 modules.

[0068] This invention integrates the GNSS positioning module and the MPU6050 acceleration sensing module on the guardrail node and uses a bus serial structure to form a group, which not only ensures the integrity and multidimensionality of the collected data, but also effectively reduces the deployment cost and equipment energy consumption, and improves the system's deployment adaptability and maintenance convenience.

[0069] In this embodiment, the preprocessing includes timestamp alignment, coordinate calculation, anomaly drift removal, signal denoising, missing value imputation, and data standardization.

[0070] This invention employs preprocessing operations such as timestamp alignment, coordinate calculation, anomaly removal, denoising, and standardization, which effectively improves the spatiotemporal consistency and stability of the collected data, providing a high-quality input foundation for subsequent state estimation and feature modeling.

[0071] In this embodiment, the edge server is equipped with solar panels and batteries to supply power to the entire train convoy. It is installed on the edge of the road, has cellular communication capabilities, and is connected to the train convoy at one end and to the cloud server at the other end.

[0072] This invention utilizes an edge server combined with solar power and cellular communication to achieve a low-power, long-lasting, and high-coverage edge computing architecture, effectively improving the system's real-time processing capabilities and remote data upload efficiency, and reducing reliance on on-site manual intervention.

[0073] In this embodiment, the gated convolutional network is constructed based on a dual-channel mechanism. The first channel receives acceleration data and extracts time-dependent features, while the second channel receives spatial coordinate data and extracts displacement morphology features. The output features of the two channels are fused in the residual connection structure, and a dynamic attention gating mechanism is introduced to reconstruct the time-dependent features and displacement morphology features by weighting, forming a collision feature vector.

[0074] This invention designs a dual-channel gated convolutional network structure based on acceleration and coordinate data. By using residual connections and dynamic attention gating mechanisms, it achieves deep fusion of time-dependent features and displacement morphological features, effectively improving the system's accuracy and generalization ability in recognizing collision types and state changes.

[0075] In this embodiment, the collision level label includes vehicle collision with guardrail, guardrail damage caused by human intervention, road subsidence, guardrail significant displacement, road geological deformation, and no collision.

[0076] This invention clearly classifies and outputs six collision level labels, covering different types of structural anomalies such as vehicle collisions, human-caused damage, and road subsidence. It has high-resolution and multi-dimensional state recognition capabilities, which facilitates accurate identification and targeted handling by traffic management departments.

[0077] In this embodiment, S2 specifically includes:

[0078] S21. Obtain the spatial coordinate data and acceleration data of each acquisition node in the series structure, wherein the spatial coordinate data is three-dimensional position data. The acceleration data are triaxial acceleration data. , Indicates the index of the time of data collection. Indicates the first Geographic coordinates at any given time This represents the acceleration values ​​in the three orthogonal directions at the corresponding moment;

[0079] S22. Align the 3D coordinates and acceleration data of each acquisition node with timestamps, interpolate missing samples, and perform linear interpolation to unify the sampling frequency, thus constructing an observation sequence set. ;

[0080] S23. Transmit the preprocessed observation sequence set to the edge server, and apply nonlinear time series fitting and extended Kalman filtering algorithms to perform nonlinear state estimation and noise suppression on the observation sequence. After filtering, a state vector sequence is obtained. ,in This represents the estimated current location. This represents the speed estimate. This represents the estimated acceleration value;

[0081] S24. The state vector sequence is forwarded to the cloud server via a cellular network. The nonlinear state estimation uses the following Kalman state prediction and update.

[0082] ;

[0083] in, Indicates the first State estimate at time 10:00 Indicates the first The predicted state value at time 10:00. Indicates the first Kalman gain at time step This represents the time-aligned observation vector. This represents the observation matrix.

[0084] This invention constructs a nonlinear state estimation process and uses the extended Kalman filter algorithm to jointly model acceleration and displacement data, generating a state vector sequence. This effectively suppresses environmental noise interference and improves the stability of collision data and the accuracy of structure identification.

[0085] In this embodiment, S3 specifically includes:

[0086] S31. Construct a gated convolutional network based on a dual-channel configuration, and process the state vector sequence... and These are used as inputs for the first and second channels, respectively.

[0087] S32, will The input is fed into the first channel, and a one-dimensional convolution extraction operation is performed to generate a time-dependent feature tensor. The first channel uses a convolutional kernel with a gated activation function to locally model the acceleration sequence in the temporal dimension;

[0088] S33, will The input is fed into the second channel, where a one-dimensional convolution extraction operation is performed to generate a displacement morphological feature tensor. The second channel models the displacement morphological features in the coordinate sequence through a convolutional structure;

[0089] S34. Convert the output tensors of the two channels. and Fusion is performed within the residual connectivity structure to construct a fused feature tensor. The fusion process introduces a dynamic attention gating mechanism, which forms a collision feature vector through weighted reconstruction:

[0090] ;

[0091] in, Indicates the first The collision feature vector output at time step 1. Indicates the first The time-dependent feature tensor at time step Indicates the first The displacement morphological feature tensor at time step. This represents the attention gating weight matrix. This represents the gating bias vector. This represents the hyperbolic tangent activation function. This represents the Sigmoid activation function.

[0092] This invention designs a convolutional neural network structure under a gating mechanism, models the state vector by channel and introduces residual fusion and attention weighting mechanisms to extract collision features from multiple scales, thereby enhancing the model's ability to express the coupling relationship between multiple source data.

[0093] In this embodiment, S4 specifically includes:

[0094] S41, The collision feature vector The input is a nested encoding network structure, which is then subjected to layer-by-layer nonlinear mapping and compression to obtain a multi-layer latent feature representation sequence. ,in This represents the feature vector output by the first layer of encoding. This represents the feature vector output by the second layer of encoding. This represents the feature vector output by the third layer of encoding;

[0095] S42, will The input decoding structure is used to perform reverse mapping and dimensionality reduction reconstruction to generate a reconstructed feature vector. The dimensionality reduction reconstruction process retains the core collision features and compresses redundant information.

[0096] S43. Based on the reconstructed feature vector, construct a collision determination function to determine whether the current time is a collision anomaly state. The expression of the collision determination function is:

[0097] ;

[0098] in, Indicates the first Reconstructed feature vector at time step Denotes the first linear mapping matrix. Denotes the second linear mapping matrix. This represents the first bias vector. This represents the second bias vector. Represents a non-linear activation function. This represents the Softmax normalization function. Indicates the first The collision determination label at any time includes vehicle collision with guardrail, guardrail damage caused by human intervention, road subsidence, guardrail displacement, geological deformation, and no-collision state.

[0099] This invention constructs a nested encoding and decoding structure to perform multi-level dimensionality reduction extraction of collision feature vectors, and constructs a collision determination function through nonlinear activation and the Softmax function, thereby achieving high-precision judgment and classification labeling of multiple abnormal states and improving the robustness of the system's anomaly detection.

[0100] refer to Figure 4-6 A collision-sensing intelligent guardrail system based on BeiDou positioning includes:

[0101] The data processing module is used to collect spatial coordinate data and acceleration data along the guardrail and perform data preprocessing operations.

[0102] The edge computing module is used to transmit the preprocessed sensor data to the edge server, perform nonlinear time-series fitting operations and smoothing filtering operations to generate a guardrail state vector sequence, and forward the state vector sequence to the cloud server through the cellular network.

[0103] The feature extraction module is used to construct a gated convolutional network based on a dual-channel interface. The state vector sequence is input into the gated convolutional network, and features are extracted and fused to output a guardrail collision feature vector.

[0104] The collision determination module is used to perform multi-layer nested encoding and decoding operations on the guardrail collision feature vector to obtain a dimensionality-reduced feature representation, and to construct a collision determination function to determine whether a collision anomaly is triggered and to mark the collision level label.

[0105] The early warning generation module is used to generate a collision early warning data packet when the collision level label is active. The collision early warning data packet includes the collision occurrence time, guardrail location coordinates, and collision severity level.

[0106] The response and handling module is used to transmit collision alarm data packets to the road traffic command center for road maintenance personnel to respond and handle quickly.

[0107] This invention proposes a complete hardware and algorithm architecture at the system structure level, covering various functional units such as data acquisition, edge processing, feature extraction, collision determination, early warning generation and response handling, and constructs a guardrail collision monitoring solution with intelligent perception, fast response, low power consumption and high reliability.

[0108] Example 1:

[0109] To verify the feasibility of this invention in practice, it was applied to a typical accident-prone section of a highway. This section traverses various terrain features and presents complex road conditions, including high traffic density, limited nighttime visibility, and small curve radii. Traditional guardrails often struggle to provide timely impact information after a collision, delaying maintenance responses. Furthermore, accident identification relies heavily on manual video playback, resulting in low efficiency and a high rate of misjudgments. To improve the accuracy and response speed of guardrail collision events, the collision-sensing intelligent guardrail method based on BeiDou positioning proposed in this invention was deployed on this section, and continuous periodic testing and comparative verification were conducted.

[0110] In the application, GNSS positioning modules and MPU6050 accelerometer modules were first installed every 8 meters along the bidirectional guardrail of the test section. Each group of sensor nodes was connected via a bus to form a group, and each group was connected to an edge server. The edge server was powered by solar panels and interconnected with the cloud via a 4G communication network. The guardrail sensing system collected spatial coordinate data and three-axis acceleration data in real time, and used an extended Kalman filter algorithm to perform nonlinear modeling and smoothing filtering on the multidimensional state to generate a state vector sequence. Subsequently, the state vector was input into a dual-channel gated convolutional network to extract time-dependent features and displacement morphological features, and then a nested encoding and decoding module was used to extract deep semantic features, finally outputting a collision level label.

[0111] The test simulated various typical impact scenarios, including impacts to the guardrail from different directions and angles, at different speeds, and by different vehicle types. It also included minor side scrapes, human-caused damage, and slow deformation caused by geological subsidence. Compared to traditional guardrail monitoring methods that rely solely on camera recording and manual inspection, this invention's system achieved second-level alarms in all test events, with an average positioning error of less than 1.3 meters, a collision classification accuracy rate of 96.2%, and a consistency rate of 95.7% with manual verification results.

[0112] Even under high-frequency traffic interference, the system of this invention maintained stable data transmission and a low false alarm rate, demonstrating excellent noise suppression and preventing false alarms triggered by vehicles passing close by. For scenarios involving continuous weak disturbances or low-speed displacement deformation, the system, based on deep feature reconstruction and attention-based weighted recognition, successfully identified four road subsidence incidents and two human-induced disturbances, all reported more than four hours before manual inspections, effectively improving the early warning capability for anomalies.

[0113] To verify the system's ability to manage structured data, it can automatically generate a warning data package containing fields such as collision time, coordinates, peak acceleration, speed change, and judgment label after detecting a collision anomaly. This data is then displayed and archived for analysis in real time by the traffic control backend system, greatly enhancing the value of data utilization and providing a basis for road big data analysis and maintenance scheduling.

[0114] This invention significantly improves the sensitivity, response speed, and processing accuracy of guardrail collision events in test scenarios. Compared with the traditional mode, it reduces the workload of manual investigation by more than 65% and shortens the response delay time by about 80%, and has good value for promotion and application.

[0115] Table 1. Comparison of core performance indicators of the present invention in the application of guardrail collision detection.

[0116] Test number Impact type Actual trigger time (seconds) System alarm delay (seconds) Positioning error (meters) Judgment Label Recognition accuracy (%) Peak acceleration (g) System false alarms T01 frontal minor impact 3.1 1.2 1.1 Vehicle collision guardrail 97.5 2.8 none T02 Moderate frontal impact 2.6 1.1 1.0 Vehicle collision guardrail 99.1 4.2 none T03 Side-scratching at an angle 4.3 1.6 1.4 Vehicle collision guardrail 94.7 1.9 none T04 The guardrail was shaken by human intervention. 5.0 1.3 1.2 Guardrail damaged by human intervention 92.8 0.8 none T05 Slow geological subsidence - - 1.5 Road subsidence 95.2 0.5 none T06 Warning friction without collision - - - No collision 100 0.2 none T07 High-speed severe impact 2.2 0.9 1.0 Vehicle collision guardrail 98.3 6.1 none T08 guardrail prying and damage 4.8 1.5 1.3 Guardrail damaged by human intervention 91.6 0.9 none

[0117] Table 1 shows the response effect, positioning accuracy and recognition accuracy of the present invention to various abnormal guardrail states in real test scenarios, further verifying its high robustness and practicality in complex traffic environments.

[0118] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A collision-sensing intelligent guardrail method based on BeiDou positioning, characterized in that, Includes the following steps: S1. Collect sensor data along the guardrail and perform preprocessing; S2. Transmit the preprocessed sensor data to the edge server, perform nonlinear time-series fitting and smoothing filtering operations to obtain the state vector sequence, and then forward it to the cloud server through the cellular network. S3. Construct a gated convolutional network based on a dual-channel interface, input the state vector sequence into the gated convolutional network, perform feature extraction and fusion, and output a collision feature vector. The gated convolutional network is constructed based on a dual-channel mechanism. The first channel receives acceleration data and extracts time-dependent features, while the second channel receives spatial coordinate data and extracts displacement morphology features. The output features of the two channels are fused in the residual connection structure, and a dynamic attention gating mechanism is introduced to reconstruct the time-dependent features and displacement morphology features by weighting them to form a collision feature vector. S3 specifically includes: S31. Construct a gated convolutional network based on a dual-channel configuration, and process the state vector sequence... and These are used as inputs for the first and second channels, respectively. S32, will The input is fed into the first channel, and a one-dimensional convolution extraction operation is performed to generate a time-dependent feature tensor. The first channel uses a convolutional kernel with a gated activation function to locally model the acceleration sequence in the temporal dimension; S33, will The input is fed into the second channel, where a one-dimensional convolution extraction operation is performed to generate a displacement morphological feature tensor. The second channel models the displacement morphological features in the coordinate sequence through a convolutional structure; S34. Convert the output tensors of the two channels. and Fusion is performed within the residual connectivity structure to construct a fused feature tensor. The fusion process introduces a dynamic attention gating mechanism, which forms a collision feature vector through weighted reconstruction: ； in, Indicates the first The collision feature vector output at time step 1. Indicates the first The time-dependent feature tensor at time step Indicates the first The displacement morphological feature tensor at time step. This represents the attention gating weight matrix. This represents the gating bias vector. This represents the hyperbolic tangent activation function. This represents the Sigmoid activation function; S4. Perform multi-layer nested encoding and decoding operations on the collision feature vector to obtain a dimensionality-reduced feature representation, and construct a collision determination function to determine whether a collision anomaly is triggered, while also marking the collision level label. S5. When the collision level label is active, generate a collision warning data packet, which includes the collision occurrence time, guardrail location coordinates and collision severity level. S6. Transmit the collision alarm data packet to the road traffic command center for road maintenance personnel to respond and handle quickly.

2. The collision-sensing intelligent guardrail method based on BeiDou positioning according to claim 1, characterized in that, The sensor data includes spatial coordinate data and acceleration data. The spatial coordinate data is provided by the GNSS positioning module, and the acceleration data is provided by the MPU6050 sensor module. Each module is connected in series via a bus to form a group, and each group contains 20 to 50 modules.

3. The collision-sensing intelligent guardrail method based on BeiDou positioning according to claim 1, characterized in that, The preprocessing includes timestamp alignment, coordinate calculation, anomaly drift removal, signal denoising, missing value imputation, and data standardization.

4. The collision-sensing intelligent guardrail method based on BeiDou positioning according to claim 1, characterized in that, The edge server is equipped with solar panels and batteries to supply power to the entire train convoy. It is installed on the edge of the road, has cellular communication capabilities, and is connected to the train convoy on one end and to the cloud server on the other.

5. A collision-sensing intelligent guardrail method based on BeiDou positioning according to claim 1, characterized in that, The collision level labels include vehicle collision with guardrail, guardrail damage caused by human intervention, road subsidence, significant guardrail displacement, road geological deformation, and no collision.

6. The collision-sensing intelligent guardrail method based on BeiDou positioning according to claim 1, characterized in that, S2 specifically includes: S21. Obtain the spatial coordinate data and acceleration data of each acquisition node in the series structure, wherein the spatial coordinate data is three-dimensional position data. The acceleration data are triaxial acceleration data. , Indicates the index of the time of data collection. Indicates the first Geographic coordinates at any given time This represents the acceleration values ​​in the three orthogonal directions at the corresponding moment; S22. Align the 3D coordinates and acceleration data of each acquisition node with timestamps, interpolate missing samples, and perform linear interpolation to unify the sampling frequency, thus constructing an observation sequence set. ; S23. Transmit the preprocessed observation sequence set to the edge server, and apply nonlinear time series fitting and extended Kalman filtering algorithms to perform nonlinear state estimation and noise suppression on the observation sequence. After filtering, a state vector sequence is obtained. ,in This represents the estimated current location. This represents the speed estimate. This represents the estimated acceleration value; S24. The state vector sequence is forwarded to the cloud server via a cellular network. The nonlinear state estimation uses the following Kalman state prediction and update. ； in, Indicates the first State estimate at time 10:00 Indicates the first The predicted state value at time 10:

00. Indicates the first Kalman gain at time step This represents the time-aligned observation vector. This represents the observation matrix.

7. A collision-sensing intelligent guardrail method based on BeiDou positioning according to claim 1, characterized in that, S4 specifically includes: S41, The collision feature vector The input is a nested encoding network structure, which is then subjected to layer-by-layer nonlinear mapping and compression to obtain a multi-layer latent feature representation sequence. ,in This represents the feature vector output by the first layer of encoding. This represents the feature vector output by the second layer of encoding. This represents the feature vector output by the third layer of encoding; S42, will The input decoding structure is used to perform reverse mapping and dimensionality reduction reconstruction to generate a reconstructed feature vector. The dimensionality reduction reconstruction process retains the core collision features and compresses redundant information. S43. Based on the reconstructed feature vector, construct a collision determination function to determine whether the current time is a collision anomaly state. The expression of the collision determination function is: ； in, Indicates the first Reconstructed feature vector at time step Denotes the first linear mapping matrix. Denotes the second linear mapping matrix. This represents the first bias vector. This represents the second bias vector. Represents a non-linear activation function. This represents the Softmax normalization function. Indicates the first The collision determination label at any time includes vehicle collision with guardrail, guardrail damage caused by human intervention, road subsidence, guardrail displacement, geological deformation, and no-collision state.

8. A collision-sensing intelligent guardrail system based on BeiDou positioning, comprising the collision-sensing intelligent guardrail method based on BeiDou positioning as described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to collect spatial coordinate data and acceleration data along the guardrail and perform data preprocessing operations. The edge computing module is used to transmit the preprocessed sensor data to the edge server, perform nonlinear time-series fitting operations and smoothing filtering operations to generate a guardrail state vector sequence, and forward the state vector sequence to the cloud server through the cellular network. The feature extraction module is used to construct a dual-channel gated convolutional network. The state vector sequence is input into the gated convolutional network, and features are extracted and fused to output a guardrail collision feature vector, including: Construct a gated convolutional network based on a dual-channel architecture to encode the state vector sequence... and The gated convolutional network is constructed based on a dual-channel mechanism, with the first channel receiving acceleration data and extracting time-dependent features, and the second channel receiving spatial coordinate data and extracting displacement morphology features. The output features of the two channels are fused in the residual connection structure, and a dynamic attention gating mechanism is introduced to reconstruct the time-dependent features and displacement morphology features by weighting them to form a collision feature vector. Will The input is fed into the first channel, and a one-dimensional convolution extraction operation is performed to generate a time-dependent feature tensor. The first channel uses a convolutional kernel with a gated activation function to locally model the acceleration sequence in the temporal dimension; Will The input is fed into the second channel, where a one-dimensional convolution extraction operation is performed to generate a displacement morphological feature tensor. The second channel models the displacement morphological features in the coordinate sequence through a convolutional structure; The output tensors of the two channels and Fusion is performed within the residual connectivity structure to construct a fused feature tensor. The fusion process introduces a dynamic attention gating mechanism, which forms a collision feature vector through weighted reconstruction: ； in, Indicates the first The collision feature vector output at time step 1. Indicates the first The time-dependent feature tensor at time step Indicates the first The displacement morphological feature tensor at time step. This represents the attention gating weight matrix. This represents the gating bias vector. This represents the hyperbolic tangent activation function. This represents the Sigmoid activation function; The collision determination module is used to perform multi-layer nested encoding and decoding operations on the guardrail collision feature vector to obtain a dimensionality-reduced feature representation, and to construct a collision determination function to determine whether a collision anomaly is triggered and to mark the collision level label. The early warning generation module is used to generate a collision early warning data packet when the collision level label is active. The collision early warning data packet includes the collision occurrence time, guardrail location coordinates, and collision severity level. The response and handling module is used to transmit collision alarm data packets to the road traffic command center for road maintenance personnel to respond and handle quickly.

Images