A highway vehicle speed dynamic detection system

Abstract

The application relates to the technical field of traffic control and discloses a highway vehicle speed dynamic detection system. The system comprises a distributed optical fiber vibration sensing array laid along a lane, an edge computing node and a management and control platform. The edge computing node collects vibration waveforms, constructs graph structure data based on vehicle road surface coupling dynamics space, completes lane matching and feature decoupling through a graph attention network, performs continuous wavelet transform to generate a time-frequency graph and extract feature components, constructs a loss function containing a space consistency constraint term and a time smoothness constraint term based on a distributed consensus algorithm, performs collaborative verification, is synchronized to the platform, and completes overspeed judgment and trajectory tracing. The single-point detection architecture limitation is broken, continuous vibration space-time evolution features are extracted, the speed inversion has space continuity and time stability, and the lane-changing speed measurement loophole is eliminated. The noise is separated through the combination of a dynamics mechanism and deep learning, signal aliasing is inhibited, and data reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of traffic control technology and discloses a dynamic vehicle speed detection system for highways. Background Technology

[0002] Current highway speed detection primarily employs fixed-point radar, video checkpoints, or section speed measurement schemes. Video checkpoints and fixed-point radar typically install cameras or radar probes at specific road sections, calculating instantaneous speed at a single point by capturing vehicle images or analyzing the frequency shift of reflected waves. Section speed measurement sets up capture devices at the beginning and end of a road section, calculating the average speed based on the time difference between the two ends of the section. These conventional schemes rely on visual imaging or single-point electromagnetic detection principles, with detection equipment deployed in a decentralized and isolated manner. Each device independently completes data acquisition and processing, lacking data exchange and collaborative calculation mechanisms among them.

[0003] Existing fixed-point detection equipment is based on a single-point or cross-section detection architecture, which can only extract the instantaneous state of a vehicle when it passes through a fixed cross-section. It cannot obtain the continuous vibration spatiotemporal evolution characteristics of a vehicle during its journey between adjacent detection cross-sections. This makes it impossible to perform collaborative inversion and verification of continuous vehicle speeds across the entire road segment, resulting in regulatory loopholes where vehicles can evade speed detection by changing lanes midway to leave the detection cross-section or by obscuring their license plates. Summary of the Invention

[0004] The purpose of this invention is to provide a solution to the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A dynamic vehicle speed detection system for highways includes a distributed optical fiber vibration sensor array laid under the highway lanes, IoT edge computing nodes deployed along the route, and an intelligent traffic management platform. The distributed fiber optic vibration sensor array is connected to the IoT edge computing node, and the IoT edge computing node is connected to the intelligent traffic management platform; The IoT edge computing node collects the full waveform data of vibration from the distributed fiber optic vibration sensing array in real time, and completes lane assignment matching and vehicle feature decoupling of the vibration signal based on the traffic engineering pavement dynamics model to eliminate invalid signals. A time-frequency domain deep learning vehicle speed inversion algorithm is used to extract features from the decoupled single-vehicle vibration time series signal, and the vehicle speed inversion is completed by combining the spatial distribution parameters of the distributed optical fiber vibration sensing array. The adjacent IoT edge computing nodes perform collaborative verification of vibration data and vehicle speed inversion results through the IoT architecture, and synchronize the verified vehicle speed data to the intelligent traffic management platform to complete vehicle speeding determination and driving trajectory tracing.

[0006] Preferably, the process of completing lane assignment matching and vehicle feature decoupling of vibration signals based on the traffic engineering pavement dynamics model includes: mapping the full waveform data of vibration to a preset vehicle-pavement coupled dynamics space, and constructing a graph structure data with fiber optic sensing sampling points as nodes and dynamics transmission paths as edge weights; The graph structure data is input into a pre-trained graph attention network, and the spatial correlation between the nodes is calculated through the multi-head attention mechanism in the graph attention network. Lane assignment matching is then achieved based on the spatial correlation. The graph attention network outputs a high-dimensional latent variable containing the characteristics of vehicle tire vibration force. The high-dimensional latent variable is then orthogonally projected and separated from the environmental noise characteristics to complete the decoupling of vehicle features.

[0007] Preferably, the step of using a time-frequency domain deep learning vehicle speed inversion algorithm to extract features from the decoupled single-vehicle vibration time-series signal includes: performing continuous wavelet transform on the single-vehicle vibration time-series signal to generate a two-dimensional time-frequency map, and inputting the two-dimensional time-frequency map into a feature extraction network containing parallel multi-scale convolutional kernels; In the feature extraction network, a channel attention mechanism is introduced to recalibrate the channel weights of the time-frequency feature map output by the parallel multi-scale convolution kernel, and to filter out the time-frequency feature components that are strongly correlated with vehicle speed. The selected time-frequency feature components are flattened into a one-dimensional feature vector, which is used to characterize the dynamic load time-frequency evolution characteristics of the vehicle in the current vibration acquisition cycle.

[0008] Preferably, the step of combining the spatial distribution parameters of the distributed optical fiber vibration sensing array to complete the vehicle speed inversion includes: obtaining the coordinate positions of each sampling point in the distributed optical fiber vibration sensing array to construct a spatial position sequence, and performing a tensor outer product operation on the spatial position sequence and the corresponding feature extraction results to construct a spatiotemporal joint feature matrix. A velocity decoder based on a graph convolutional network is used to process the spatiotemporal joint feature matrix. The learnable Laplacian matrix in the graph convolutional network is used to aggregate the information of adjacent nodes in the spatial position sequence, and output a spatiotemporal tensor reflecting the propagation speed of vehicle vibration waves along the distributed optical fiber vibration sensing array. The spatiotemporal tensor is then mapped to the vehicle speed inversion value.

[0009] Preferably, the adjacent IoT edge computing nodes perform collaborative verification of vibration data and vehicle speed inversion results through the IoT architecture, including: dividing the IoT edge computing nodes located in adjacent detection intervals into an edge computing cluster, and in each detection cycle, each IoT edge computing node sends its local vibration feature representation and vehicle speed inversion results to the master node in the edge computing cluster; The master node constructs a local loss function for the vehicle speed inversion result based on a distributed consensus algorithm, and minimizes the local loss function through gradient descent iteratively until the variance of the vehicle speed inversion results between adjacent nodes is less than a preset convergence threshold, and outputs the verified vehicle speed data.

[0010] Preferably, the process of determining vehicle speeding and tracing driving trajectory includes: recombining the verified vehicle speed data according to timestamps and spatial location sequences to generate a multi-dimensional vehicle speed sequence; The morphological similarity distance between the multidimensional vehicle speed sequence and the preset speed limit standard sequence for each lane is calculated using a dynamic time warping algorithm. When the morphological similarity distance is less than the speeding determination threshold, a speeding warning is triggered. Meanwhile, the missing spatial position values ​​in the multidimensional vehicle speed sequence are interpolated and smoothed based on the Kalman filter algorithm, and the interpolated and smoothed spatial position sequence is fitted into a continuous driving trajectory curve to complete the driving trajectory tracing.

[0011] Preferably, the construction of the graph structure data with fiber optic sensing sampling points as nodes and dynamic transmission paths as edge weights includes: for each pair of fiber optic sensing sampling points that have a spatially adjacent relationship, constructing an elastic mechanical transfer function based on the stiffness distribution of the road structure layer; Perform an inverse Laplace transform on the elastic mechanical transfer function to extract the attenuation coefficient and phase delay parameter of the vibration energy between sampling points; The attenuation coefficient and the phase delay parameter are nonlinearly fused to generate the edge weight, so that the information transmission path between connected nodes in the graph structure data matches the real fluctuation propagation mechanism of vehicle tire excitation force in the road structure.

[0012] Preferably, the step of generating a two-dimensional time-frequency graph by performing continuous wavelet transform on the single vehicle vibration time series signal includes: establishing a candidate set containing multiple mother wavelet basis functions, and calculating the mutual information between each mother wavelet basis function and the local extreme point distribution characteristics based on the local extreme point distribution characteristics of the single vehicle vibration time series signal; The mother wavelet basis function with the largest mutual information is selected as the target basis function. A sliding window continuous wavelet transform is performed on the single vehicle vibration time series signal using the target basis function. By adjusting the scaling factor and translation factor of the sliding window, the two-dimensional time-frequency diagram with adaptive resolution matching capability is generated.

[0013] Preferably, the step of aggregating the neighboring node information in the spatial location sequence using the learnable Laplacian matrix in the graph convolutional network includes: constructing an initial adjacency matrix based on the spatial location sequence, and converting the initial adjacency matrix into a difference matrix between the angle matrix and the adjacency matrix; Learnable adaptive adjustment parameters are introduced into the diagonal elements of the difference matrix to generate the learnable Laplacian matrix; In each layer of graph convolution calculation of the velocity decoder, the learnable Laplacian matrix is ​​used to perform spectral domain filtering on the spatiotemporal joint feature matrix to dynamically adjust the node information aggregation intensity at different spatial distances and suppress the aliasing interference of vibration signals from distant nodes.

[0014] Preferably, the construction of the local loss function for the vehicle speed inversion result based on the distributed consensus algorithm includes: extracting the vibration feature representations of adjacent nodes within the overlapping area of ​​the time window, and calculating the cosine similarity of the vibration feature representations of the overlapping area as a spatial consistency constraint term; Calculate the rate of change of the derivative of the vehicle speed inversion result output by each node in the time dimension, and use the mean square error of the rate of change of the derivative as a time smoothness constraint term; The spatial consistency constraint and the temporal smoothness constraint are weighted and summed to generate the local loss function, so as to jointly optimize the spatial continuity and temporal stability of the vehicle speed inversion results between adjacent nodes in the gradient descent iteration.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention collects full-waveform vibration data from fiber optic arrays by deploying edge computing nodes along the route. It constructs a spatiotemporal joint feature matrix by performing tensor outer product operations on the spatial location sequence and feature extraction results. A learnable Laplacian matrix in a graph convolutional network is then used to aggregate information from neighboring nodes to output vehicle speed inversion values. Neighboring edge computing nodes are grouped into clusters. The master node constructs a local loss function based on a distributed consensus algorithm, incorporating spatial consistency and temporal smoothness constraints. Gradient descent iteratively optimizes the vehicle speed inversion results between neighboring nodes. This scheme breaks the limitations of single-point detection architectures, extracting continuous spatiotemporal vibration evolution features between neighboring nodes. This ensures that the vehicle speed inversion across the entire road segment possesses spatial continuity and temporal stability, eliminating regulatory loopholes such as evading speed measurement by changing lanes and leaving the detection section.

[0016] 2. This invention maps vibration waveforms to a vehicle-road coupled dynamics spatial construction graph structure. It extracts attenuation coefficients and phase delay parameters based on the elasticity transfer function to generate edge weights, and combines a graph attention network to calculate spatial correlation, completing lane assignment matching and vehicle feature decoupling. When generating the two-dimensional time-frequency graph, the target mother wavelet basis function is selected based on mutual information, and a channel attention mechanism is used to filter time-frequency feature components. In the overspeed determination and trajectory tracing stages, a dynamic time warping algorithm is used to calculate morphological similarity distance, and a Kalman filter algorithm is used to interpolate and smooth missing spatial location values. This method integrates road dynamics transfer mechanisms with deep learning feature extraction, separating environmental noise interference, suppressing vibration signal aliasing at distant nodes, and improving the analytical capability of single-vehicle vibration time-series signals and the reliability of vehicle speed inversion data. Attached Figure Description

[0017] Figure 1 A flowchart of the overall process for a highway vehicle speed dynamic detection system; Figure 2 Flowchart for decoupling lane assignment matching and vehicle features; Figure 3 Here is a flowchart of the time-frequency domain feature extraction process; Figure 4 Flowchart for vehicle speed inversion; Figure 5 A flowchart for the collaborative verification process between adjacent nodes; Figure 6 This is a flowchart for speeding detection and driving trajectory tracing. Detailed Implementation

[0018] In one embodiment, reference Figure 1The highway vehicle speed dynamic detection system comprises a distributed fiber optic vibration sensor array laid beneath the highway lanes, IoT edge computing nodes deployed along the route, and an intelligent traffic management platform. The distributed fiber optic vibration sensor array communicates with the IoT edge computing nodes, which in turn communicate with the intelligent traffic management platform. The distributed fiber optic vibration sensor array uses tightly packed communication optical fibers, laid along the highway's driving direction between the asphalt subbase and the water-stabilized base layer of each lane. The sampling point spacing of the sensor array is set to 0.5m, and the sampling frequency is set to 10kHz, used to collect full waveform data of road surface vibration generated by vehicle tire excitation. One IoT edge computing node is deployed every 1km along the highway, with each edge computing node managing a 1km length of the distributed fiber optic vibration sensor array. The edge computing nodes use industrial-grade processors, are configured with industrial-grade storage and communication modules, and communicate with the distributed fiber optic vibration sensor array within their corresponding management section via fiber optic access interfaces, with adjacent IoT edge computing nodes via Ethernet interfaces, and establish a wide-area communication connection with the intelligent traffic management platform via 5G industrial communication modules. The intelligent traffic management platform is deployed in the highway traffic management center. It adopts a distributed server cluster architecture and is equipped with data storage units, speeding judgment units, trajectory tracking units, and early warning release units.

[0019] The IoT edge computing node collects real-time full-waveform vibration data from a distributed fiber optic vibration sensor array. This data includes the vibration displacement amplitude, phase, and frequency information for each sampling point at each sampling moment. The data format is a time-series floating-point array with dimensions of [number of sampling points, number of sampling points per sampling period]. The single sampling period is set to 100ms, corresponding to 1000 sampling points per period. The IoT edge computing node uses a traffic engineering pavement dynamics model to perform lane assignment matching and vehicle feature decoupling of the vibration signals to eliminate invalid signals. This model is constructed based on parameters such as the asphalt layer thickness, elastic modulus, water-stabilized base stiffness, and subgrade resilient modulus of the current pavement, and is used to characterize the propagation law of vehicle tire excitation force in the pavement structure. The IoT edge computing node inputs the collected full-waveform vibration data into the road dynamics model of traffic engineering, completes the lane assignment matching of vibration signals, determines the lane number corresponding to each vibration signal, and completes vehicle feature decoupling. It separates the vibration time sequence signal corresponding to a single vehicle from the mixed vibration signals, eliminates invalid signals such as environmental vibration and inherent vibration of the road structure, and outputs the decoupled single vehicle vibration time sequence signal.

[0020] The IoT edge computing node employs a time-frequency domain deep learning vehicle speed inversion algorithm to extract features from the decoupled single-vehicle vibration time-series signal. This algorithm includes a time-frequency feature extraction branch and a vehicle speed inversion branch. The IoT edge computing node performs a time-frequency domain transformation on the decoupled single-vehicle vibration time-series signal to generate a corresponding time-frequency feature map. A deep learning network then extracts features from this map, obtaining a high-dimensional feature vector related to vehicle speed. The IoT edge computing node then combines the spatial distribution parameters of a distributed fiber optic vibration sensor array to complete the vehicle speed inversion. These parameters include the absolute coordinates of each sampling point, the spacing between sampling points, and the orientation of the sensor array. The extracted high-dimensional feature vector is fused with the spatial distribution parameters to output the inverted vehicle speed value for a single vehicle within the current detection range.

[0021] Adjacent IoT edge computing nodes collaboratively verify vibration data and vehicle speed inversion results through an IoT architecture. IoT edge computing nodes deployed adjacent to each other along the highway establish point-to-point communication connections via wired Ethernet, forming a collaborative IoT edge computing architecture. After each sampling period, adjacent IoT edge computing nodes exchange locally collected vibration characteristic data and locally calculated vehicle speed inversion results. Based on preset collaborative verification rules, they verify the spatial continuity and temporal consistency of the vehicle speed inversion results, eliminate abnormal vehicle speed inversion results, and output the verified vehicle speed data.

[0022] The IoT edge computing nodes synchronize the verified vehicle speed data to the intelligent traffic management platform. This speed data includes the vehicle's lane number, sampling timestamp, spatial coordinates of the sampling point, and the retrieved speed value. Upon receiving the speed data, the intelligent traffic management platform compares it with the speed limit for the corresponding lane to determine if the vehicle is speeding. Simultaneously, based on the vehicle's speed data and corresponding spatial location information across different detection zones throughout the road segment, it generates a continuous driving trajectory for the vehicle, enabling trajectory tracing.

[0023] In one embodiment, lane assignment matching and vehicle feature decoupling of vibration signals are achieved based on a traffic engineering pavement dynamics model. The specific implementation is as follows: The full-waveform vibration data is mapped to a predefined vehicle-pavement coupled dynamics space. This space is a multi-dimensional feature space constructed based on pavement structure dynamic parameters, with dimensions including vibration amplitude, vibration frequency, phase shift, pavement structure layer stiffness, and vibration energy attenuation coefficient. After mapping the full-waveform vibration data collected at each sampling point to this space, a multi-dimensional dynamic feature vector corresponding to each sampling point is obtained. This constructs a graph structure data with fiber optic sensing sampling points as nodes and dynamic transmission paths as edge weights.

[0024] For each pair of fiber optic sensing sampling points with spatial adjacency, an elastic transfer function based on the stiffness distribution of the road surface structure layers is constructed. The expression for the elastic transfer function is: in, The frequency of vibration. The elastic modulus of the pavement structural layer. Let be the moment of inertia of the pavement structure layer. The density of the road surface material, The cross-sectional area of ​​the pavement structure layer. This is the attenuation coefficient of the vibration wave in the road surface structure. The phase coefficient of the vibration wave. The spatial distance between two adjacent sampling points. It is the imaginary unit.

[0025] For elastic mechanical transfer function Performing the inverse Laplace transform yields the vibration transfer response function in the time domain. ,from Extract the attenuation coefficient of vibration energy between sampling points. With phase delay parameter The attenuation coefficient With phase delay parameter Perform nonlinear fusion calculations to generate edge weights The formula for calculating edge weights is: in, , The preset weighting coefficients, This is the theoretical propagation delay time of the vibration wave between the two sampling points. , This represents the theoretical propagation speed of the vibration wave in the road surface structure.

[0026] The constructed graph structure data is input into a pre-trained graph attention network. The graph attention network consists of two graph attention layers and one fully connected output layer. Each graph attention layer employs an eight-head multi-head attention mechanism. The spatial correlation between nodes is calculated using the multi-head attention mechanism within the graph attention network. For each target node, the multi-head attention mechanism calculates the attention coefficient between the target node and all its neighboring nodes. The formula for calculating the attention coefficient is as follows: in, For nodes With nodes Attention coefficient between them , They are nodes The node feature vector of the node. The weight matrix is ​​a learnable linear transformation. For learnable attention vectors, This is a vector concatenation operation. It is a non-linear activation function.

[0027] The attention coefficients of all neighboring nodes are subjected to softmax normalization to obtain the normalized attention weights. , That is, nodes and nodes The spatial correlation between nodes is determined. Based on the magnitude of the spatial correlation, nodes with a spatial correlation greater than a preset correlation threshold are grouped into the same lane group to achieve lane affiliation matching. Each lane group corresponds to a driving lane on the highway.

[0028] refer to Figure 2 The fully connected output layer of the graph attention network outputs a high-dimensional latent variable containing the characteristics of vehicle tire vibration force. The dimension of the high-dimensional latent variable is set to 128. The high-dimensional latent variable is separated from the environmental noise features by orthogonal projection, and the orthogonal basis matrix of the road environmental noise is extracted in advance through unsupervised learning. High-dimensional hidden variables Projected onto orthogonal basis matrix The orthogonal complement space is used to obtain the separated vehicle feature vectors. The calculation formula is: By performing this orthogonal projection separation operation, the environmental noise feature components in the high-dimensional latent variables are removed, the vehicle features are decoupled, and the decoupled single-vehicle vibration time series signal is output.

[0029] In one embodiment, reference Figure 3 A time-frequency domain deep learning vehicle speed inversion algorithm is used to extract features from the decoupled single-vehicle vibration time-series signal. The specific implementation is as follows: A continuous wavelet transform is performed on the single-vehicle vibration time-series signal to generate a two-dimensional time-frequency graph. A candidate set containing various mother wavelet basis functions is established, including Morlet wavelet, Mexican Hat wavelet, Haar wavelet, and Daubechies wavelet series. Based on the local extremum distribution characteristics of the single-vehicle vibration time-series signal, the mutual information between each mother wavelet basis function and the local extremum distribution characteristics is calculated. The formula for calculating the mutual information is: in, This is a sequence of local extrema points for a single vehicle vibration time-series signal. The sampling sequence of the mother wavelet basis functions. for and The joint probability distribution, , They are respectively and The marginal probability distribution.

[0030] The mother wavelet basis function with the highest mutual information is selected as the target basis function. A sliding window continuous wavelet transform is then performed on the single-vehicle vibration time series signal using the target basis function. The calculation formula for the continuous wavelet transform is as follows: in, These are the wavelet transform coefficients. This is a single-vehicle vibration timing signal. Let be the target mother wavelet basis function. for The complex conjugate function, As a scale factor, This is the translation factor. The scaling factor of the sliding window is adjusted. With translation factor It generates a two-dimensional time-frequency graph with adaptive resolution matching capability. The horizontal axis of the two-dimensional time-frequency graph is the time dimension, the vertical axis is the frequency dimension, and the pixel value is the amplitude of the wavelet transform coefficient at the corresponding time-frequency position.

[0031] The two-dimensional time-frequency map is input into a feature extraction network containing parallel multi-scale convolutional kernels. The feature extraction network contains three parallel convolutional branches, each of which uses convolutional kernels of 33, 55, and 77 respectively. Each convolutional branch contains a convolutional layer, a batch normalization layer, and a ReLU activation layer in sequence. The output feature maps of the three branches are concatenated in the channel dimension to obtain a multi-scale fused time-frequency feature map.

[0032] In the feature extraction network, a channel attention mechanism is introduced to recalibrate the channel weights of the time-frequency feature map output by the parallel multi-scale convolutional kernels. The channel attention mechanism adopts the SE-Net structure. First, a global average pooling operation is performed on the multi-scale fused time-frequency feature map to obtain a global feature vector of the channel dimension. Then, a nonlinear transformation is performed on the global feature vector through two fully connected layers to output the weight coefficients corresponding to each channel. The weight coefficients are multiplied with the corresponding channels of the original multi-scale fused time-frequency feature map to complete the channel weight recalibration and filter out the time-frequency feature components that are strongly correlated with vehicle speed.

[0033] The selected time-frequency feature components are input into the global average pooling layer and flattened into a one-dimensional feature vector. The dimension of the one-dimensional feature vector is set to 256. The one-dimensional feature vector is used to characterize the dynamic load time-frequency evolution characteristics of the vehicle in the current vibration acquisition cycle.

[0034] In one embodiment, reference Figure 4 Vehicle speed inversion is achieved by combining the spatial distribution parameters of a distributed fiber optic vibration sensor array. The specific implementation method is as follows: The coordinate positions of each sampling point in the distributed fiber optic vibration sensor array are obtained to construct a spatial position sequence, which is: ,in The number of sampling points, For the first A two-dimensional absolute coordinate vector of each sampling point, with dimension 21, containing the mileage marker and lateral lane coordinates of the sampling point. This spatial location sequence... The corresponding one-dimensional feature vector of the feature extraction result Perform tensor outer product operations to construct the spatiotemporal joint feature matrix. The calculation formula is: in, For tensor outer product operators, spatiotemporal joint characteristic matrix The dimension is .

[0035] A speed decoder based on graph convolutional networks is used to analyze the spatiotemporal joint feature matrix. The processing speed decoder consists of three graph convolutional layers and one linear regression output layer. It aggregates neighboring node information in the spatial location sequence using a learnable Laplacian matrix in the graph convolutional network, based on the spatial location sequence... Construct the initial adjacency matrix Initial adjacency matrix for A symmetric matrix, where If and only if the sampling point With sampling points If the spatial distance between them is less than the preset adjacency threshold, otherwise The initial adjacency matrix Transform into angle matrix Adjacency Matrix The difference matrix, i.e., the initial form of the Laplace matrix. , where the angle matrix for diagonal matrix, .

[0036] In the difference matrix Introduce learnable adaptive adjustment parameters into the diagonal elements. Generate a learnable Laplacian matrix The expression for the Laplace matrix can be learned as follows: in, It is a diagonal matrix, with diagonal elements These are learnable, adaptively adjustable parameters that are iteratively optimized through backpropagation during model training.

[0037] In the graph convolution computation of each layer of the speed decoder, a learnable Laplacian matrix is ​​utilized. spatiotemporal joint characteristic matrix For spectral domain filtering, the calculation expression for the graph convolutional layer is: in, For the first The input feature matrix of the layer graph convolutional layer, For the first The output feature matrix of the layer graph convolutional layer, It is the identity matrix. It is a ReLU nonlinear activation function. For the first The learnable weight matrix of the layered graph convolutional layer. Through this spectral domain filtering operation, the aggregation intensity of node information at different spatial distances is dynamically adjusted to suppress the aliasing interference of vibration signals from distant nodes and complete the aggregation of information from adjacent nodes.

[0038] The output of the final graph convolutional layer of the velocity decoder is the spatiotemporal tensor reflecting the propagation speed of vehicle vibration waves along the distributed fiber optic vibration sensing array. spacetime tensor The dimension is Spacetime tensor Mapped to vehicle speed inversion value , for spacetime tensor The elements in the sample are weighted and averaged, and the vehicle speed inversion value is obtained by combining the sampling point spacing and the sampling time interval. The calculation formula is: in, The time difference between vibration signal propagation between adjacent sampling points. The spatial distance between adjacent sampling points. For spacetime tensor The Middle The value of each element.

[0039] In one embodiment, adjacent IoT edge computing nodes perform collaborative verification of vibration data and vehicle speed inversion results through an IoT architecture. The specific implementation is as follows: IoT edge computing nodes located in adjacent detection zones are divided into edge computing clusters. Each edge computing cluster contains 3-5 consecutively deployed IoT edge computing nodes. In each edge computing cluster, a master node is elected according to a preset master node election rule, and the remaining nodes are slave nodes. The master node election rule is as follows: During cluster initialization, each node broadcasts its own hardware performance parameters and network link quality parameters, selecting the node with the best hardware performance and network link quality as the master node. The master node's term is 24 hours, and the election process is re-executed after the term ends.

[0040] During each detection cycle, each slave node sends its local vibration feature representation and vehicle speed inversion result to the master node in the edge computing cluster. The vibration feature representation is a one-dimensional feature vector output by the feature extraction stage, and the vehicle speed inversion value is the output result of the vehicle speed inversion stage. The data transmission process adopts the TCP / IP protocol, and the transmission cycle is consistent with the vibration data sampling cycle.

[0041] refer to Figure 5 The master node constructs a local loss function for the vehicle speed inversion results based on a distributed consensus algorithm. It iteratively minimizes this local loss function using gradient descent until the variance of the vehicle speed inversion results between adjacent nodes is less than a preset convergence threshold, at which point it outputs the verified vehicle speed data. Vibration feature representations of adjacent nodes within the overlapping region of the time window are extracted. This overlapping region is the overlapping detection area within the jurisdiction of adjacent nodes, with a length set to 50m. The cosine similarity of the vibration feature representations in the overlapping region is calculated as a spatial consistency constraint. The calculation formula is: in, For nodes Vibration feature representation vector within the overlapping region, Adjacent nodes Vibration feature representation vector within the overlapping region, For vector dot product operation, Let L2 be the norm of the vector.

[0042] Calculate the rate of change of the derivative of the vehicle speed inversion results output by each node in the time dimension, and use the mean square error of the rate of change of the derivative as a time smoothness constraint. The calculation formula is: in, The number of sampling periods within the time window. For the first Vehicle speed inversion values ​​within each sampling period The vehicle speed inversion value is at the th The time derivative of each sampling period.

[0043] Spatial consistency constraint With time smoothness constraint Perform weighted summation to generate a local loss function. The calculation formula is: in, , These are the preset weighting coefficients for the spatial consistency constraint and the temporal smoothness constraint, respectively.

[0044] The master node distributes the constructed local loss function to all slave nodes in the cluster. Each node uses the gradient descent iterative algorithm to minimize the local loss function. During the iteration, the local vehicle speed inversion results are updated synchronously. After each iteration, each node uploads the updated vehicle speed inversion results to the master node. The master node calculates the variance of the vehicle speed inversion results between adjacent nodes. When the variance is less than the preset convergence threshold, the iteration stops and the verified vehicle speed data is output.

[0045] In one embodiment, reference Figure 6 The system completes vehicle speeding detection and trajectory tracing, specifically as follows: The verified vehicle speed data is reassembled according to timestamps and spatial location sequences to generate a multi-dimensional vehicle speed sequence. Multidimensional vehicle speed sequence The dimension is ,in The sampling period is the number of sampling periods, and each dimension corresponds to the timestamp, spatial location coordinates, and vehicle speed inversion value, respectively.

[0046] Multidimensional vehicle speed sequences are calculated using a dynamic time warping algorithm. With the preset speed limit sequence for each lane Morphological similarity distance between them, speed limit standard sequence To be compatible with multi-dimensional vehicle speed sequences Given sequences of the same length, where each element represents the speed limit for the corresponding lane. The resulting sequence is of size [size missing]. Distance matrix ,in , is the first in the multidimensional vehicle speed sequence The element and the speed limit standard sequence number The absolute difference between each element; solved using a dynamic programming algorithm from the distance matrix. of Location to The path with the minimum cumulative distance at each location; the minimum cumulative distance is the morphological similarity distance. .

[0047] Preset overspeed detection threshold When the morphological similarity distance Less than the overspeed threshold When a vehicle is found to be speeding, a speeding warning is triggered, and the speed data and driving trajectory information of the speeding vehicle are synchronized to the highway law enforcement system to complete the speeding evidence collection.

[0048] Simultaneously, based on the Kalman filter algorithm, the missing spatial location values ​​in the multidimensional vehicle speed sequence are interpolated and smoothed. The state equation and observation equation of the Kalman filter are constructed. The state equation is as follows: The observation equation is: in, for The system state vector at any given time contains information about the vehicle's position, velocity, and acceleration. Here is the state transition matrix. To control the input matrix, To control the input amount, For process noise, for The observed values ​​at any given time, namely the spatial location coordinates and vehicle speed inversion values ​​uploaded by the edge computing nodes. For the observation matrix, To observe noise.

[0049] By employing the prediction and update steps of Kalman filtering, missing spatial location values ​​in the multidimensional vehicle speed sequence are interpolated and filled in. Simultaneously, the spatial location sequence is smoothed to eliminate the influence of observation noise. The smoothed spatial location sequence is then fitted into a continuous driving trajectory curve using a cubic spline interpolation algorithm, thus completing the driving trajectory tracing.

Claims

1. A dynamic vehicle speed detection system for highways, characterized in that, This includes distributed fiber optic vibration sensor arrays laid under the highway lanes, IoT edge computing nodes deployed along the route, and intelligent traffic management platforms. The distributed fiber optic vibration sensor array is connected to the IoT edge computing node, and the IoT edge computing node is connected to the intelligent traffic management platform; The IoT edge computing node collects the full waveform data of vibration from the distributed fiber optic vibration sensing array in real time, and completes lane assignment matching and vehicle feature decoupling of the vibration signal based on the traffic engineering pavement dynamics model to eliminate invalid signals. A time-frequency domain deep learning vehicle speed inversion algorithm is used to extract features from the decoupled single-vehicle vibration time series signal, and the vehicle speed inversion is completed by combining the spatial distribution parameters of the distributed optical fiber vibration sensing array. The adjacent IoT edge computing nodes perform collaborative verification of vibration data and vehicle speed inversion results through the IoT architecture, and synchronize the verified vehicle speed data to the intelligent traffic management platform to complete vehicle speeding determination and driving trajectory tracing.

2. The highway vehicle speed dynamic detection system according to claim 1, characterized in that, The method of completing lane assignment matching and vehicle feature decoupling of vibration signals based on traffic engineering pavement dynamics model includes: mapping the full waveform data of vibration to a preset vehicle-pavement coupled dynamics space, and constructing a graph structure data with fiber optic sensing sampling points as nodes and dynamic transmission paths as edge weights; The graph structure data is input into a pre-trained graph attention network, and the spatial correlation between the nodes is calculated through the multi-head attention mechanism in the graph attention network. Lane assignment matching is then achieved based on the spatial correlation. The graph attention network outputs a high-dimensional latent variable containing the characteristics of vehicle tire vibration force. The high-dimensional latent variable is then orthogonally projected and separated from the environmental noise characteristics to complete the decoupling of vehicle features.

3. The highway vehicle speed dynamic detection system according to claim 1, characterized in that, The method of using a time-frequency domain deep learning vehicle speed inversion algorithm to extract features from the decoupled single-vehicle vibration time series signal includes: performing continuous wavelet transform on the single-vehicle vibration time series signal to generate a two-dimensional time-frequency map, and inputting the two-dimensional time-frequency map into a feature extraction network containing parallel multi-scale convolutional kernels. In the feature extraction network, a channel attention mechanism is introduced to recalibrate the channel weights of the time-frequency feature map output by the parallel multi-scale convolution kernel, and to filter out the time-frequency feature components that are strongly correlated with vehicle speed. The selected time-frequency feature components are flattened into a one-dimensional feature vector, which is used to characterize the dynamic load time-frequency evolution characteristics of the vehicle in the current vibration acquisition cycle.

4. The highway vehicle speed dynamic detection system according to claim 1, characterized in that, The process of combining the spatial distribution parameters of the distributed optical fiber vibration sensing array to complete the vehicle speed inversion includes: obtaining the coordinate positions of each sampling point in the distributed optical fiber vibration sensing array to construct a spatial position sequence, and performing a tensor outer product operation on the spatial position sequence and the corresponding feature extraction results to construct a spatiotemporal joint feature matrix. A velocity decoder based on a graph convolutional network is used to process the spatiotemporal joint feature matrix. The learnable Laplacian matrix in the graph convolutional network is used to aggregate the information of adjacent nodes in the spatial position sequence, and output a spatiotemporal tensor reflecting the propagation speed of vehicle vibration waves along the distributed optical fiber vibration sensing array. The spatiotemporal tensor is then mapped to the vehicle speed inversion value.

5. The highway vehicle speed dynamic detection system according to claim 1, characterized in that, The adjacent IoT edge computing nodes perform collaborative verification of vibration data and vehicle speed inversion results through the IoT architecture, including: dividing the IoT edge computing nodes located in adjacent detection intervals into an edge computing cluster; and in each detection cycle, each IoT edge computing node sends its local vibration feature representation and vehicle speed inversion results to the master node in the edge computing cluster. The master node constructs a local loss function for the vehicle speed inversion result based on a distributed consensus algorithm, and minimizes the local loss function through gradient descent iteratively until the variance of the vehicle speed inversion results between adjacent nodes is less than a preset convergence threshold, and outputs the verified vehicle speed data.

6. The highway vehicle speed dynamic detection system according to claim 1, characterized in that, The process of determining vehicle speeding and tracing driving trajectory includes: recombining the verified vehicle speed data according to timestamps and spatial location sequences to generate a multi-dimensional vehicle speed sequence; The morphological similarity distance between the multidimensional vehicle speed sequence and the preset speed limit standard sequence for each lane is calculated using a dynamic time warping algorithm. When the morphological similarity distance is less than the speeding determination threshold, a speeding warning is triggered. Meanwhile, the missing spatial position values ​​in the multidimensional vehicle speed sequence are interpolated and smoothed based on the Kalman filter algorithm, and the interpolated and smoothed spatial position sequence is fitted into a continuous driving trajectory curve to complete the driving trajectory tracing.

7. The highway vehicle speed dynamic detection system according to claim 2, characterized in that, The construction of the graph structure data with fiber optic sensing sampling points as nodes and dynamic transmission paths as edge weights includes: for each pair of fiber optic sensing sampling points that have a spatial adjacency relationship, constructing an elastic mechanical transfer function based on the stiffness distribution of the road structure layer. Perform an inverse Laplace transform on the elastic mechanical transfer function to extract the attenuation coefficient and phase delay parameter of the vibration energy between sampling points; The attenuation coefficient and the phase delay parameter are nonlinearly fused to generate the edge weight, so that the information transmission path between connected nodes in the graph structure data matches the real fluctuation propagation mechanism of vehicle tire excitation force in the road structure.

8. The highway vehicle speed dynamic detection system according to claim 3, characterized in that, The step of generating a two-dimensional time-frequency graph by performing continuous wavelet transform on the single vehicle vibration time series signal includes: establishing a candidate set containing multiple mother wavelet basis functions, and calculating the mutual information between each mother wavelet basis function and the local extreme point distribution characteristics based on the local extreme point distribution characteristics of the single vehicle vibration time series signal; The mother wavelet basis function with the largest mutual information is selected as the target basis function. A sliding window continuous wavelet transform is performed on the single vehicle vibration time series signal using the target basis function. By adjusting the scaling factor and translation factor of the sliding window, the two-dimensional time-frequency diagram with adaptive resolution matching capability is generated.

9. The highway vehicle speed dynamic detection system according to claim 4, characterized in that, The step of aggregating the neighboring node information in the spatial location sequence using the learnable Laplacian matrix in the graph convolutional network includes: constructing an initial adjacency matrix based on the spatial location sequence, and converting the initial adjacency matrix into a difference matrix between the angle matrix and the adjacency matrix; Learnable adaptive adjustment parameters are introduced into the diagonal elements of the difference matrix to generate the learnable Laplacian matrix; In each layer of graph convolution calculation of the speed decoder, the learnable Laplacian matrix is ​​used to perform spectral domain filtering on the spatiotemporal joint feature matrix to dynamically adjust the node information aggregation intensity at different spatial distances.

10. The highway vehicle speed dynamic detection system according to claim 5, characterized in that, The local loss function for the vehicle speed inversion result is constructed based on the distributed consensus algorithm, including: extracting the vibration feature representation of adjacent nodes in the overlapping area of ​​the time window, and calculating the cosine similarity of the vibration feature representation of the overlapping area as a spatial consistency constraint term; Calculate the rate of change of the derivative of the vehicle speed inversion result output by each node in the time dimension, and use the mean square error of the rate of change of the derivative as a time smoothness constraint term; The spatial consistency constraint and the temporal smoothness constraint are weighted and summed to generate the local loss function, so as to jointly optimize the spatial continuity and temporal stability of the vehicle speed inversion results between adjacent nodes in the gradient descent iteration.

Images