Guardrail collision warning method and system

Abstract

The invention discloses a guardrail collision warning method and system, and the method comprises the steps: carrying out the nonlinear phase alignment of a multi-source heterogeneous signal through an adaptive variational mode decomposition algorithm according to the vibration acceleration, strain tensor and acoustic emission signals collected in real time through a multi-mode sensor array disposed at a guardrail key node, generating a three-dimensional dynamic strain field distribution vector; inputting the three-dimensional dynamic strain field distribution vector into a nonlinear dynamics reconstruction module, and extracting a chaotic feature fingerprint spectrum of the collision event; performing collision intensity grading processing on the chaos feature fingerprint spectrum, and outputting a quantitative evaluation matrix including a collision grade, a damage radius and a residual intensity prediction value; and triggering a multi-mode alarm protocol according to the quantitative evaluation matrix, and synchronously transmitting the multi-mode alarm protocol to a traffic management center and an adjacent vehicle OBU terminal. According to the embodiment of the invention, the state of the guardrail can be monitored in real time, and a collision event can be accurately evaluated and warned.

Description

Technical Field

[0001] The present invention belongs to the technical field of early warning, and particularly relates to a guardrail collision warning method and system. Background Art

[0002] With the acceleration of the urbanization process and the increase in traffic volume, road traffic safety issues have become increasingly prominent. Among many traffic safety hazards, guardrails, as an important part of road traffic facilities, play a key role in protecting pedestrians and maintaining traffic order. However, when a guardrail is collided, it directly affects the effectiveness of its protection function, which may lead to subsequent accidents. Therefore, timely detection and warning of guardrail collision events are crucial for reducing traffic accidents and ensuring driving safety.

[0003] Traditional guardrail collision detection methods mainly rely on means such as visual monitoring, pressure sensors, or mechanical switches. These methods often have problems such as slow response time, limited monitoring range, high false alarm rate, and it is difficult to accurately judge the intensity and influence range of collision events in real time. In a dynamic traffic environment, relying solely on static monitoring means is particularly insufficient and cannot meet the requirements of fast and reliable collision warning. Summary of the Invention

[0004] The purpose of the present invention is to provide a guardrail collision warning method and system to solve the deficiencies in the prior art, which can monitor the state of the guardrail in real time and accurately evaluate and warn collision events.

[0005] An embodiment of the present application provides a guardrail collision warning method, and the method includes: According to the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by a multi-modal sensor array deployed at key nodes of the guardrail, non-linear phase alignment processing is performed on the multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector. Among them, the sensor array uses a piezoelectric-fiber composite sensing unit, and the spatio-temporal correlation weight of each node signal is calculated based on mutual information entropy; The three-dimensional dynamic strain field distribution vector is input into a non-linear dynamics reconstruction module, and a high-dimensional phase space projection is constructed based on the geometric invariance principle of the chaotic attractor. Through the joint calculation of the largest Lyapunov exponent and Kolmogorov entropy, the chaotic feature fingerprint spectrum of the collision event is extracted. Among them, the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectory caused by environmental noise; Perform collision intensity grading processing on the chaotic feature fingerprint spectrum, construct a collision energy propagation model using a spatio-temporal convolutional pulse neural network, and fuse the strain gradient distribution, energy decay rate, and material damage threshold parameters through a dynamic weight allocation mechanism to output a quantitative evaluation matrix including collision level, damage radius, and predicted value of residual strength; Trigger a multimodal alarm protocol according to the quantization evaluation matrix, dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack, and encode the alarm information into a time-varying frequency-hopping pulse sequence by using chaotic encryption modulation technology, and synchronously transmit it to the traffic management center and adjacent vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

[0006] Optionally, the vibration acceleration, strain tensor and acoustic emission signals collected in real time by the multimodal sensor array deployed according to the key nodes of the guardrail are subjected to non-linear phase alignment processing of multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector. Among them, the sensor array uses a piezoelectric-fiber composite sensing unit and calculates the spatio-temporal correlation weights of the signals of each node based on mutual information entropy, including: According to the vibration acceleration, strain tensor and acoustic emission signals collected in real time by the piezoelectric-fiber composite sensing unit, calculate the spatio-temporal correlation weights of the signals of each node through the mutual information entropy algorithm to generate a dynamic correlation matrix of multi-source signals. Among them, the mutual information entropy algorithm introduces a time-delay embedding technology to eliminate the heterodyne interference between sensors; Input the dynamic correlation matrix into the adaptive variational mode decomposition module, dynamically optimize the mode number and bandwidth parameters based on the Kullback-Leibler divergence, perform non-linear phase alignment processing on multi-source heterogeneous signals, and output a set of decomposition modes with a unified time reference; Perform multi-scale covariance analysis on the set of decomposition modes, extract the principal component features of the strain field through the asymmetric tensor decomposition algorithm, and generate a three-dimensional feature tensor containing strain gradient, vibration energy and acoustic emission event density; Input the three-dimensional feature tensor into the physical field fusion network, adopt a dual-channel attention mechanism to perform cross-modal feature fusion on piezoelectric signals and fiber signals, and finally output a spatio-temporally continuous three-dimensional dynamic strain field distribution vector.

[0007] Optionally, input the three-dimensional dynamic strain field distribution vector into the non-linear dynamics reconstruction module, construct a high-dimensional phase space projection based on the geometric invariance principle of the chaotic attractor, and extract the chaotic feature fingerprint map of the collision event through the joint calculation of the largest Lyapunov exponent and Kolmogorov entropy. Among them, the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectory caused by environmental noise, including: According to the three-dimensional dynamic strain field distribution vector, reconstruct the high-dimensional phase space by using the delay coordinate embedding method, dynamically select the embedding dimension and time-delay parameters based on Takens' theorem, and generate an initial phase trajectory set; Input the initial phase trajectory into the hyperbolic embedding algorithm, screen out the phase trajectory segments that satisfy the hyperbolic constraint by calculating the local curvature tensor, eliminate the pseudo-phase trajectories caused by environmental noise, and obtain the core phase trajectory cluster after noise reduction; Perform chaotic invariance analysis on the core phase trajectory cluster, jointly calculate the largest Lyapunov exponent and Kolmogorov entropy, and generate a chaotic dynamics parameter matrix; Input the chaotic dynamics parameter matrix into the geometric topology mapping module, extract the topological invariant features of the phase trajectory based on the persistent homology algorithm, and output a chaotic feature fingerprint map including the dimension of the strange attractor, fractal structure, and recurrence characteristics.

[0008] Optionally, perform collision intensity grading on the chaotic feature fingerprint map, construct a collision energy propagation model using a spatio-temporal convolutional spiking neural network, fuse the strain gradient distribution, energy decay rate, and material damage threshold parameters through a dynamic weight allocation mechanism, and output a quantitative evaluation matrix including collision level, damage radius, and predicted residual strength value, including: According to the chaotic feature fingerprint map, map the dimension of the strange attractor to the collision energy space through topological data binning technology, and construct multi-level classification boundary conditions for collision intensity; Input the multi-level classification boundary conditions into the spatio-temporal convolutional spiking neural network, use the LIF neuron model to simulate the biological pulse triggering mechanism, dynamically model the collision energy propagation path, and output a probability density function including spatio-temporal energy distribution; Optimize the physical constraints of the probability density function, construct a dynamic weight allocation function based on the material damage constitutive equation, fuse the strain gradient distribution, energy decay rate, and material fatigue threshold parameters, and generate a multi-objective coupled damage evolution model; Input the damage evolution model into the Monte Carlo-Finite Element hybrid solver, and finally output a quantitative evaluation matrix including collision level, damage radius, and residual strength through alternating iteration of random sampling and deterministic solution.

[0009] Optionally, trigger a multi-modal alarm protocol according to the quantitative evaluation matrix, dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack, encode the alarm information into a time-varying frequency hopping pulse sequence using chaotic encryption modulation technology, and synchronously transmit it to the traffic management center and adjacent vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response, including: According to the collision level parameter in the quantitative evaluation matrix, trigger a priority-based multi-modal alarm protocol, dynamically allocate communication spectrum resources through the LoRaWAN protocol stack, and generate a channel occupancy status map; Input the alarm information and the channel occupancy status spectrum into the chaotic encryption modulation module, generate a time-varying key stream based on the Lorenz equation, encode the data into chaotic pulse sequences using differential frequency hopping technology, and output an anti-jamming encrypted signal; Construct a joint game model of collision events and channel interference, solve the optimal transmission strategy through Nash equilibrium, dynamically adjust the transmission power and frequency hopping interval, and optimize the end-to-end transmission delay to the sub-second level; Broadcast the anti-jamming encrypted signal through a multi-antenna MIMO array, synchronously transmit it to the traffic management center and neighboring vehicle OBU terminals, and achieve ultra-low bit error rate parsing at the receiving end based on the Turbo decoding algorithm.

[0010] Optionally, perform chaotic invariance analysis on the core phase trajectory cluster, jointly calculate the largest Lyapunov exponent and Kolmogorov entropy, and generate a chaotic dynamics parameter matrix, including: According to the geometric distribution characteristics of the core phase trajectory cluster, use the multifractal spectrum analysis method to calculate the Hölder exponent of each phase trajectory, and generate a fractal dimension weight matrix, where the multifractal spectrum analysis method detects the singular point distribution at the edge of the phase trajectory through the wavelet transform modulus maximum chain; Input the fractal dimension weight matrix into the local linearization LLA algorithm, fit the phase trajectory tangent space based on the weighted least squares method, dynamically correct the noise sensitivity coefficient in the calculation of the Lyapunov exponent, and output a sequence of Lyapunov exponents with the maximum robustness; Perform information geometric manifold modeling on the core phase trajectory cluster, quantify the information flow density of the phase space by calculating the Ricci curvature tensor, and dynamically adjust the estimation window of the Kolmogorov entropy in combination with the Kontsevich entropy formula to generate an entropy value evolution spectrogram; Input the Lyapunov exponent sequence and the entropy value evolution spectrogram into the dynamic mode decomposition DMD module, solve the joint distribution function of chaotic dynamics parameters through the Tikhonov regularization method, and finally output a three-dimensional chaotic dynamics parameter matrix including fractal dimension, exponential sensitivity, and entropy production rate.

[0011] Another embodiment of the present application provides a guardrail collision warning system, the system includes: A processing module, configured to perform nonlinear phase alignment processing on multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by a multi-modal sensor array deployed at key nodes of the guardrail, and generate a three-dimensional dynamic strain field distribution vector, where the sensor array uses a piezoelectric-fiber composite sensing unit, and calculates the spatio-temporal correlation weight of each node signal based on mutual information entropy; An extraction module, configured to input the three-dimensional dynamic strain field distribution vector into a non-linear dynamics reconstruction module, construct a high-dimensional phase space projection based on the geometric invariance principle of a chaotic attractor, and extract the chaotic feature fingerprint spectrum of a collision event through the joint calculation of the largest Lyapunov exponent and Kolmogorov entropy. Wherein, the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectory caused by environmental noise; A construction module, configured to perform collision intensity grading processing on the chaotic feature fingerprint spectrum, construct a collision energy propagation model using a spatio-temporal convolutional pulse neural network, and fuse strain gradient distribution, energy decay rate, and material damage threshold parameters through a dynamic weight allocation mechanism, and output a quantization evaluation matrix including predicted values of collision level, damage radius, and residual strength; An alarm module, configured to trigger a multi-modal alarm protocol according to the quantization evaluation matrix, dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack, encode alarm information into a time-varying frequency-hopping pulse sequence using chaotic encryption modulation technology, and synchronously transmit it to a traffic management center and adjacent vehicle OBU terminals. Wherein, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

[0012] Another embodiment of the present application provides a storage medium, in which a computer program is stored. Wherein, the computer program is set to execute the method described in any one of the above when running.

[0013] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0014] Compared with the prior art, a guardrail collision alarm method provided by the present invention, based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by a multi-modal sensor array deployed at key nodes of the guardrail, performs non-linear phase alignment processing on multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector; inputs the three-dimensional dynamic strain field distribution vector into a non-linear dynamics reconstruction module to extract the chaotic feature fingerprint spectrum of a collision event; performs collision intensity grading processing on the chaotic feature fingerprint spectrum and outputs a quantization evaluation matrix including predicted values of collision level, damage radius, and residual strength; triggers a multi-modal alarm protocol according to the quantization evaluation matrix and synchronously transmits it to a traffic management center and adjacent vehicle OBU terminals, so as to be able to monitor the state of the guardrail in real time and accurately evaluate and alarm collision events. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1Hardware block diagram of a computer terminal for a guardrail collision warning method provided by an embodiment of the present invention; Figure 2 Flow schematic diagram of a guardrail collision warning method provided by an embodiment of the present invention; Figure 3 Structural schematic diagram of a guardrail collision warning system provided by an embodiment of the present invention. Detailed implementation manners

[0016] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] An embodiment of the present invention first provides a guardrail collision warning method, which can be applied to an electronic device, such as a computer terminal, specifically, a general computer, etc.

[0018] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 Hardware block diagram of a computer terminal for a guardrail collision warning method provided by an embodiment of the present invention. As Figure 1 shown, the computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0019] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any guardrail collision warning method.

[0020] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0021] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any guardrail collision warning method.

[0022] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in [[ ]] is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0023] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0024] See Figure 2 , embodiments of the present invention provide a guardrail collision warning method, which may include the following steps: S201, according to the vibration acceleration, strain tensor and acoustic emission signals collected in real time by the multi-modal sensor array deployed at the key nodes of the guardrail, perform non-linear phase alignment processing on the multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector, wherein the sensor array uses a piezoelectric-optical fiber composite sensing unit, and calculates the spatio-temporal correlation weights of the signals of each node based on mutual information entropy; This step uses a piezoelectric-optical fiber composite sensor array to collect the vibration, strain and acoustic emission signals of the guardrail, performs alignment processing on the multi-source signals through an adaptive variational mode decomposition (VMD) algorithm, and eliminates the phase difference between the signals. Mutual information entropy is used to calculate the spatio-temporal correlation between different sensor nodes to ensure the accuracy of data fusion. Finally, a three-dimensional dynamic strain field distribution vector is generated to reflect the stress state of the guardrail, solves the problem of signal heterogeneity of multi-source sensors, improves the data fusion accuracy, provides a high-fidelity input for subsequent collision detection, and avoids false judgment or missed detection.

[0025] Specifically, according to the vibration acceleration, strain tensor and acoustic emission signals collected in real time by the piezoelectric-optical fiber composite sensing unit, calculate the spatio-temporal correlation weights of the signals of each node through the mutual information entropy algorithm to generate a dynamic correlation matrix of multi-source signals, wherein the mutual information entropy algorithm introduces a time delay embedding technique to eliminate the heterodyne interference between sensors; this step uses mutual information entropy to calculate the spatio-temporal correlation of different sensor signals, combines the time delay embedding technique to eliminate heterodyne interference, generates a dynamic correlation matrix, ensures the accuracy of multi-source data fusion, improves the robustness of sensor data fusion, avoids false judgment caused by signal asynchronization, and lays a foundation for subsequent signal processing.

[0026] Input the dynamic correlation matrix into the adaptive variational mode decomposition module, dynamically optimize the mode number and bandwidth parameters based on the Kullback-Leibler divergence, perform non-linear phase alignment processing on multi-source heterogeneous signals, and output a set of decomposed modes with a unified time reference; this step adopts the adaptive VMD algorithm to dynamically adjust the mode number and bandwidth, eliminate the signal phase difference, ensure the alignment of multi-source signals under the unified time reference, improve data consistency, solve the problem of inconsistent phases of multi-source signals, improve the accuracy of subsequent analysis, and avoid misjudgment caused by signal misalignment.

[0027] Perform multi-scale covariance analysis on the set of decomposed modes, extract the principal component features of the strain field through the asymmetric tensor decomposition algorithm, and generate a three-dimensional feature tensor containing strain gradient, vibration energy, and acoustic emission event density; this step uses multi-scale covariance analysis to extract the signal principal components, combines the asymmetric tensor decomposition algorithm, generates a three-dimensional feature tensor, comprehensively reflects the strain gradient, vibration energy, and acoustic emission event density, provides a more comprehensive characterization of the force state of the guardrail, and provides high-dimensional feature input for subsequent collision detection.

[0028] Input the three-dimensional feature tensor into the physical field fusion network, adopt a two-channel attention mechanism to perform cross-modal feature fusion on piezoelectric signals and fiber optic signals, and finally output a spatio-temporally continuous three-dimensional dynamic strain field distribution vector. This step uses the two-channel attention mechanism to adaptively weight and fuse piezoelectric and fiber optic signals, generate a spatio-temporally continuous strain field distribution vector, improve the data representation ability, enhance the intelligent level of data fusion, and improve the accuracy and robustness of collision detection.

[0029] Exemplarily, a specific implementation method is as follows: Step 1: According to the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the piezoelectric-fiber composite sensing unit, calculate the spatio-temporal correlation weights of the signals at each node through the mutual information entropy algorithm, and generate a dynamic correlation matrix of multi-source signals, where the mutual information entropy algorithm introduces the time delay embedding technology to eliminate the heterodyne interference between sensors.

[0030] Specific implementation method: The piezoelectric - fiber composite sensing unit consists of a piezoelectric ceramic sheet and a distributed fiber Bragg grating (FBG) sensor. The piezoelectric sensor captures high - frequency vibration acceleration signals at a sampling rate of 10 kHz, while the fiber optic sensor measures low - frequency strain tensors at a sampling rate of 1 kHz. The acoustic emission signals are collected by the broadband response (0.1 - 1 MHz) of the piezoelectric unit. Due to the sampling rate differences between different sensors, time - delay embedding technology is required to align the time series. For example, for the strain data of the fiber optic sensor, the sampling rate is increased to 10 kHz through cubic spline interpolation, and a time - delay compensation algorithm based on mutual information entropy is introduced: First, calculate the mutual information entropy between the piezoelectric and fiber optic signals. When the delay time is τ, the mutual information value reaches the maximum. At this time, the fiber optic signal sequence is shifted backward by τ sampling points to achieve phase alignment. In specific operations, assume that the mutual information entropy between the piezoelectric vibration signal and the fiber optic strain signal before alignment is 0.35. Through searching with a sliding time window (step size 1 μs), it is found that when the fiber optic signal is delayed by 3.2 ms, the mutual information entropy increases to 0.78, and at this time, the time synchronization of the two signals is completed.

[0031] The calculation of the mutual information entropy weight uses the sliding window joint probability distribution estimation. The guardrail is divided into 20 key nodes along the length direction, and 3 piezoelectric sensors and 2 groups of orthogonal fiber gratings are deployed at each node. For the vibration signal of the i - th node and the strain signal of the j - th node, the joint probability distribution p(x_i, y_j) and the marginal distributions p(x_i), p(y_j) are calculated by the kernel density estimation method, and the mutual information entropy value is calculated using the KL divergence formula. For example, in a certain collision event, the mutual information entropy between the vibration signal of node 5 and the strain signal of node 8 is 0.65, while the mutual information entropy between node 5 and node 12 is only 0.12, indicating that the collision energy mainly propagates along the direction of node 5→8. When finally generating the dynamic correlation matrix, the mutual information entropy values of each node pair are normalized as weight coefficients to form a 20×20 symmetric matrix, and the main diagonal elements are the entropy - weight average values of the multimodal signals (vibration, strain, acoustic emission) of the nodes themselves.

[0032] Time - delay embedding technology: Using the Takens embedding theorem, a delay vector X(t)=[x(t), x(t + τ), x(t + 2τ),..., x(t+(m - 1)τ)] is constructed for the low - frequency strain signal of the fiber optic sensor, where the embedding dimension m = 3 is determined by the false nearest neighbor method, and the delay time τ = 5 ms is determined by the first zero - crossing point of the autocorrelation function.

[0033] Calculation of mutual information entropy: The probability density is estimated using the Parzen window kernel function, the kernel bandwidth h = 0.1σ (σ is the signal standard deviation), and the sliding window length is set to 200 ms to cover at least two collision pulse periods.

[0034] Step 2: Input the dynamic correlation matrix into the adaptive variational mode decomposition module, dynamically optimize the mode number and bandwidth parameters based on the Kullback-Leibler divergence, perform non-linear phase alignment processing on the multi-source heterogeneous signals, and output a set of decomposed modes with a unified time reference.

[0035] The core of the adaptive variational mode decomposition (AVMD) module lies in dynamically determining the mode number K and the bandwidth parameter α. Traditional VMD requires presetting the K value, while this method uses the KL divergence as the objective function for optimization: initialize K = 3, after performing VMD decomposition for each candidate K value, calculate the sum of the KL divergences between each mode and the original signal. When K increases from 3 to 5, the total KL divergence decreases from 1.24 to 0.87; when continuing to increase K to 6, the KL divergence only decreases to 0.85 and mode aliasing occurs. Therefore, K = 5 is selected as the optimal mode number. The adjustment of the bandwidth parameter α is achieved through spectral coherence analysis: for the central frequency f_k of each mode, calculate its frequency band overlap with the adjacent mode f_{k±1}, when the overlap exceeds 20%, increase α from the initial value of 2000 to 2500 to narrow the bandwidth.

[0036] Non-linear phase alignment is achieved by iteratively adjusting the modal phases. Taking the vibration signal in a certain collision event as an example, there is a phase offset caused by sensor resonance at 5 kHz in the original signal. Five modes are obtained by AVMD decomposition (the central frequencies are 120 Hz, 850 Hz, 3.2 kHz, 5 kHz, and 8 kHz respectively), and the phase of the 5 kHz mode lags by 15° due to sensor resonance. The instantaneous phase φ(t) of this mode is extracted through the Hilbert transform, and a linear phase compensation θ(f) = 15°×(f / 5 kHz) is applied in the frequency domain, finally aligning this mode with other modes in the time-frequency plane. After all modes are phase-corrected, they are reconstructed into time-synchronized signal components through inverse transformation.

[0037] KL divergence optimization: Set the search range of K to 3 - 8, check the modal orthogonality each time K is increased, and terminate the search if the correlation coefficient between two modes exceeds 0.3.

[0038] Bandwidth adjustment: For the spectrum S_k(f) of mode k, calculate the energy proportion within the interval [f_k -Δf, f_k +Δf] (Δf = α / 2), when the proportion is lower than 90%, perform iterative update according to α←α×1.2.

[0039] Step 3: Perform multi-scale covariance analysis on the set of decomposed modes, extract the principal component features of the strain field through the asymmetric tensor decomposition algorithm, and generate a three-dimensional feature tensor containing strain gradient, vibration energy, and acoustic emission event density.

[0040] Multi-scale covariance analysis adopts the method of combining wavelet transform with a sliding window. First, perform 5-layer discrete wavelet decomposition (using the db4 wavelet basis) on each modal signal to obtain the detail coefficients D1 - D5 and the approximation coefficient A5. On the time scale, set three analysis windows: short-term (50 ms, corresponding to D1 - D2), medium-term (200 ms, D3 - D4), and long-term (800 ms, A5). Calculate the covariance matrix for each window: For example, within the short-term window, the covariance between the vibration mode of node 5 and the strain mode of node 8 is 0.72, indicating a high correlation between the two in the microsecond-level dynamic response. Asymmetric tensor decomposition is constructed for three-dimensional data (node × modal × scale). Assume the original tensor is X ∈ R^{20×5×3}, and use the CP decomposition model to approximate it as the product of three factor matrices A (node space), B (modal space), and C (scale space). Solve it iteratively by the alternating least squares method, set the rank R = 4, and the fitting error drops below 5% after 10 iterations. The principal components extracted after decomposition include: the forward propagation mode in the node space (weight vector [0.32, 0.28, ..., 0.05]), the high-frequency energy dominant mode in the modal space (weights [0.41, 0.39, 0.12, 0.08]), and the short-term impact feature in the scale space (weights [0.67, 0.23, 0.10]).

[0041] The construction of the three-dimensional feature tensor specifically includes: 1. Strain gradient: Calculate the strain difference between adjacent nodes along the longitudinal direction of the guardrail. For example, the strain difference Δε = ε_{i + 1}-ε_i between node i and i + 1, and after normalization, map it to the [0, 1] interval.

[0042] 2. Vibration energy: Calculate the RMS value of each modal signal within a 1-second window length. For example, the RMS of the 5 kHz modal is 0.32g.

[0043] 3. Acoustic emission event density: Count the number of acoustic emission pulses exceeding the threshold (such as 50 dB) per unit time. For example, 18 events are detected within 200 ms after the collision, and the density is 90 events / s.

[0044] Wavelet scale selection: The D1 layer corresponds to 0 - 2.5 kHz (sampling rate 5 kHz), the D2 layer corresponds to 1.25 - 2.5 kHz, and so on to the A5 layer corresponding to 0 - 78 Hz.

[0045] Asymmetric tensor decomposition: Use the Tensorly library to implement CP decomposition, set non-negative constraints to enhance physical interpretability, and retain the components with a variance contribution rate exceeding 85%.

[0046] Feature normalization: The strain gradient is normalized using min-max normalization, the vibration energy is logarithmically transformed as log(1 + RMS), and the acoustic emission density is standardized using Z-score normalization.

[0047] Step 4: Input the three-dimensional feature tensor into the physical field fusion network, and use a dual-channel attention mechanism to perform cross-modal feature fusion on the piezoelectric signal and the fiber optic signal, and finally output a spatio-temporally continuous three-dimensional dynamic strain field distribution vector.

[0048] The physical field fusion network adopts a dual-branch structure: the piezoelectric branch processes the vibration acceleration and acoustic emission features (dimension 20×5×2), and the fiber optic branch processes the strain tensor features (20×3×1). Each branch contains 3 layers of convolution (kernel size 3×3×3), followed by batch normalization and ReLU activation. The specific implementation of the dual-channel attention mechanism includes: 1. Channel attention: For the feature map F_p ∈ R^{20×5×2} of the piezoelectric branch, the channel descriptor v_p ∈ R^2 is obtained through global average pooling, and then the attention weight w_p = σ(W_2δ(W_1v_p)) is generated through a fully connected layer, where W_1 ∈ R^{4×2}, W_2 ∈ R^{2×4}, and σ is the sigmoid function. For example, it is calculated that w_p = [0.63, 0.37], enhancing the contribution of the vibration acceleration feature.

[0049] 2. Spatial attention: For the feature map F_f ∈ R^{20×3×1} of the fiber optic branch, a 3×3 convolution is used to generate the attention map A_f ∈ R^{20×3}, highlighting the regions of sudden change in the strain gradient (e.g., A_f = 0.91 at node 8).

[0050] 3. Cross-modal interaction: Information exchange is achieved through a cross-convolution layer. For example, the second convolution layer of the piezoelectric branch receives the features of the first layer of the fiber optic branch as input supplements.

[0051] In the final fusion stage, the feature maps of the two branches are concatenated in the node dimension (20×(5 + 3)), reduced to 20×3 using 1×1 convolution, and then a three-dimensional strain field with spatial continuity is generated through cubic spline interpolation. For example, the maximum strain gradient of 0.85 mm / m is detected between nodes 5 - 8 of the guardrail, the vibration energy is concentrated in the 3.2 kHz frequency band, the peak value of the acoustic emission density reaches 120 events / s, and the numerical values of the output vector in the three physical field dimensions of (x, y, z) after fusion are [0.72, 0.85, 0.91].

[0052] Convolution kernel configuration: The number of output channels of the first layer of convolution in the piezoelectric branch is 16, and that in the fiber optic branch is 8; the number of hidden units in the fully connected layer of the attention module is 4.

[0053] Interpolation method: Cubic convolution interpolation is adopted, and the interpolation point spacing is set to 0.5 meters along the length direction of the guardrail, generating a strain field distribution of 100 continuous spatial points in total.

[0054] Dynamic range control: The dynamic range of the output vector is compressed, and μ-law companding (μ = 255) is used to meet the input requirements of the subsequent module.

[0055] S202, input the three-dimensional dynamic strain field distribution vector into the nonlinear dynamics reconstruction module, construct a high-dimensional phase space projection based on the geometric invariance principle of the chaotic attractor, and extract the chaotic characteristic fingerprint spectrum of the collision event through the joint calculation of the maximum Lyapunov exponent and the Kolmogorov entropy. Among them, the reconstruction module uses the hyperbolic embedding algorithm to eliminate the pseudo-phase trajectories caused by environmental noise; This step utilizes the chaotic dynamics theory to map the strain field vector into a high-dimensional phase space, and quantifies the chaotic characteristics of the system through the maximum Lyapunov exponent and the Kolmogorov entropy. The hyperbolic embedding algorithm filters out environmental noise interference, extracts the real collision characteristic fingerprint spectrum, ensures the robustness of the detection, effectively distinguishes real collisions from environmental noise, improves the detection reliability, and provides an accurate characteristic basis for subsequent collision intensity grading.

[0056] Specifically, according to the three-dimensional dynamic strain field distribution vector, the delay coordinate embedding method can be used to reconstruct the high-dimensional phase space, dynamically select the embedding dimension and time delay parameters based on Takens' theorem, and generate an initial set of phase trajectories; this step uses the delay coordinate embedding method to map the one-dimensional strain signal into a high-dimensional phase space, adaptively determines the optimal embedding dimension (such as 3 - 5 dimensions) and time delay parameters (such as the zero-crossing point of the autocorrelation function) through Takens' theorem, avoiding phase space distortion caused by artificial setting. The reconstructed set of phase trajectories retains the dynamic characteristics of the original signal, provides a basis for subsequent chaotic analysis, solves the problem that low-dimensional observation signals are difficult to characterize complex dynamic systems, and the high-dimensional phase space reconstruction can accurately capture the nonlinear characteristics of guardrail collisions, laying a mathematical foundation for chaotic fingerprint extraction.

[0057] Input the initial phase trajectories into the hyperbolic embedding algorithm, screen out the phase trajectory segments that meet the hyperbolic constraints by calculating the local curvature tensor, eliminate the pseudo-phase trajectories caused by environmental noise, and obtain the core phase trajectory cluster after noise reduction; this step uses the hyperbolic embedding algorithm to calculate the local curvature of the phase trajectories (such as the principal curvature and Gaussian curvature), retains the trajectory segments that meet the hyperbolic condition (such as the saddle point structure), filters out the pseudo-trajectories (such as circular or divergent trajectories) generated by environmental noise such as wind vibration and vehicle vibration, ensures the signal purity of subsequent analysis, significantly improves the signal-to-noise ratio, avoids misjudgment of chaotic characteristics caused by noise interference, enhances the robustness of collision detection, and is especially suitable for the complex environment of highways.

[0058] Perform chaotic invariance analysis on the core phase trajectory cluster, jointly calculate the largest Lyapunov exponent and Kolmogorov entropy, and generate a chaotic dynamics parameter matrix; calculate the largest Lyapunov exponent (characterizing the sensitivity of the system to initial values) through the Wolf algorithm, and combine the KS entropy (reflecting the information loss rate) to quantify the chaos degree of the phase trajectory. The parameter matrix synthesizes the numerical values of both (such as λ1>0 and a sudden increase in the entropy value is determined as a strong collision), distinguishes different types of dynamic behaviors (such as periodic vibration and collision impact), provides quantitative chaotic characteristics of collision events, and the joint criterion of the exponent and entropy is more reliable than a single index, avoiding misclassifying high-energy noise as real collisions.

[0059] Input the chaotic dynamics parameter matrix into the geometric topology mapping module, extract the topological invariant features of the phase trajectory based on the persistent homology algorithm, and output a chaotic feature fingerprint map containing the dimension of the strange attractor, fractal structure, and recurrence characteristics. Use the persistent homology algorithm to analyze the topological structure of the phase trajectory (such as the number of holes, Betti numbers), and combine recurrence quantification analysis (RQA) to extract the fractal dimension of the attractor (such as the correlation dimension of 2.7) and the recurrence rate (such as >80% is determined as a deterministic collision), generating a multi-dimensional fingerprint map. The topological features have strong invariance to noise and parameter changes, and the fingerprint map provides highly discriminative features for collision classification, being more adaptable to non-linear impact scenarios than traditional time-frequency analysis.

[0060] Exemplarily, a specific implementation manner includes: Step 1: According to the three-dimensional dynamic strain field distribution vector, use the delay coordinate embedding method to reconstruct the high-dimensional phase space, dynamically select the embedding dimension and time delay parameters based on the Takens theorem, and generate an initial phase trajectory set; The core idea of the delay coordinate embedding method is to reconstruct a high-dimensional phase space that can reflect the dynamic characteristics of the original system through the combination of time delay and embedding dimension for one-dimensional time series data. For example, assume that the strain field distribution vector collected by the sensor is a three-dimensional strain gradient sequence containing 100 nodes, and the data of each node is sampled at a frequency of 1 kHz to form a time series. To reconstruct the phase space, two key parameters need to be determined first: the time delay (τ) and the embedding dimension (m).

[0061] Time delay selection: Use the mutual information method to determine the optimal time delay. Specifically, calculate the mutual information amount between the original time series and its own sequence after a delay of τ. When the mutual information amount first reaches the local minimum, the corresponding τ is the optimal delay. For example, for a certain acoustic emission signal, by traversing the delay values from τ = 1 to τ = 50, it is found that when τ = 12, the mutual information amount is the lowest, indicating that the redundant information between the sequences is the least at this time, and the independent dynamic information can be retained to the greatest extent.

[0062] Embedding dimension determination: The embedding dimension is dynamically optimized based on the False Nearest Neighbors (FNN) algorithm. The algorithm starts from m=2 and gradually increases the dimension, calculating whether each data point has a "false neighbor" with its nearest neighbor in the high-dimensional space due to insufficient dimension. For example, when m=3, the proportion of false neighbors drops below 5%, and it is considered that m=3 is sufficient to characterize the dynamic characteristics of the system. In actual operation, if the strain field data has strong nonlinear characteristics (such as mutations under impact loads), a higher embedding dimension (such as m=5~7) may be required to fully expand the phase trajectory.

[0063] Phase space reconstruction: The original time series {x(t)} is converted into a point set X(t)=[x(t),x(t+τ),x(t+2τ),..., x(t+(m-1)τ)] in the phase space. For example, for the vibration acceleration signal, if τ=10ms and m=4, each phase point is composed of a four-dimensional vector of acceleration values at the current moment and within the next 30ms. The entire time series is traversed by a sliding window, and finally an initial phase trajectory set containing tens of thousands of phase points is generated.

[0064] Dynamic application of Takens' theorem: Takens theorem guarantees that as long as the embedding dimension m ≥ 2d + 1 (d is the real dimension of the system), the reconstructed phase space can maintain the same differential homeomorphism as the original system. In practice, the real dimension of the system is unknown, so it is necessary to combine self-consistency verification: if the geometric structure of the phase trajectory (such as the shape of the attractor) no longer changes significantly with the increase of m under a certain m value, then the m value is considered to meet the requirements. For example, when the attractor of a certain section of strain field data presents a stable spiral structure when m = 5, and there is no significant difference in structure when m = 6, then m = 5 is selected.

[0065] Step 2: Input the initial phase trajectory into the hyperbolic embedding algorithm, filter out the phase trajectory segments that meet the hyperbolic constraint by calculating the local curvature tensor, remove the pseudo phase trajectory caused by environmental noise, and obtain the core phase trajectory cluster after denoising; The core goal of the hyperbolic embedding algorithm is to separate the real dynamic signal from the environmental noise through geometric constraints. The key lies in calculating the local curvature tensor of the phase trajectory and distinguishing the noise from the real signal based on the hyperbolic criterion (i.e., the sign consistency of the curvature eigenvalue).

[0066] Local curvature tensor calculation: For each phase point X(t), select k nearest neighbor points (e.g. k=20) in its neighborhood and use principal component analysis (PCA) to calculate the local tangent space and normal space. The specific steps are: 1. Construct the covariance matrix of the neighborhood point set and calculate its eigenvalues and eigenvectors.

[0067] 2. The eigenvector corresponding to the largest eigenvalue is the main direction of the tangent space, and the rest are the normal space directions.

[0068] 3. Estimate the curvature tensor through the second derivative to reflect the degree of curvature of the phase trajectory in the tangent space and the normal space. For example, in the region where the strain field changes abruptly, the normal component of the curvature tensor increases significantly, indicating a drastic change in the dynamic state.

[0069] Application of the hyperbolicity criterion: The hyperbolic system requires that the phase trajectory diverges exponentially in the tangent space (positive Lyapunov exponent) and contracts exponentially in the normal space (negative Lyapunov exponent). The criterion is realized by analyzing the signs of the eigenvalues of the curvature tensor: If the curvature eigenvalue of the tangent space is positive and that of the normal space is negative, it is determined that the region satisfies the hyperbolic constraint.

[0070] For the noise-dominated region, the signs of the curvature eigenvalues are chaotic (e.g., negative curvature appears in the tangent space), then it is marked as a pseudo-phase trajectory. For example, if a positive curvature peak appears in the normal space of a certain phase trajectory, it indicates the existence of abnormal fluctuations (such as wind vibration interference) and needs to be removed.

[0071] Pseudo-phase trajectory removal strategy: Adopt the sliding window detection method to divide the phase trajectory into segments of length L (e.g., L = 50 points), and count the proportion of each segment that satisfies the hyperbolic constraint. If the proportion is lower than the threshold (e.g., 70%), it is determined as a noise segment and deleted. For example, when a guardrail is slightly scratched by a vehicle, the hyperbolicity ratio of the phase trajectory reaches 85%, while the ratio of the segment caused by environmental vibration (such as strong wind) is only 40%, and the latter is automatically filtered.

[0072] Generation of the core phase trajectory cluster: Stitch the remaining phase trajectory segments in chronological order to form the core phase trajectory cluster after noise reduction. To enhance robustness, kernel density estimation (KDE) can be further applied to smooth the distribution of phase points and eliminate isolated noise points. For example, visualizing the core phase trajectory of a certain collision event, it can be seen that it presents a clear ring attractor structure, while the noise segments show scattered point clouds.

[0073] Step 3: Conduct chaotic invariance analysis on the core phase trajectory cluster, jointly calculate the largest Lyapunov exponent and the Kolmogorov entropy, and generate a chaotic dynamics parameter matrix; The chaotic invariance analysis aims to quantify the dynamic complexity of the system. The largest Lyapunov exponent (LLE) characterizes the sensitive dependence of the system on the initial conditions, and the Kolmogorov entropy (K-entropy) reflects the information loss rate. The two together constitute the core parameters of the chaotic characteristic fingerprint.

[0074] Calculation of the largest Lyapunov exponent: Adopt the Wolf algorithm to directly track the divergence rate of the phase trajectory: 1. Select a reference point X(t0) in the core phase trajectory cluster, find its nearest neighbor point X'(t0), and calculate the initial distance d0.

[0075] 2. Evolve with time, track the change of the distance d(t) between the two points until d(t) exceeds a predetermined threshold (such as 10% of the phase space diameter).

[0076] 3. Record the divergence time Δt and update the exponential estimate according to the formula LLE = (1 / Δt) * ln(d(t) / d0).

[0077] 4. Repeat the above process multiple times and take the average value as the final LLE. For example, the calculation result of LLE for a certain collision event shows that at the moment of collision, LLE suddenly increases to 0.8 bit / s, indicating that the system enters a strong chaotic state.

[0078] Kolmogorov entropy estimation: Calculate the K entropy based on the phase space partitioning method: 1. Divide the phase space into hypercube grids with side length ε.

[0079] 2. Statistically calculate the probability p_i of the phase trajectory visiting different grids.

[0080] 3. Calculate the entropy value K = -lim_{ε→0} lim_{τ→0} (1 / τ)Σp_i ln p_i.

[0081] In practical applications, the Grassberger-Procaccia algorithm is used for approximate estimation: the K entropy is fitted through the correlation integral C(ε) (that is, the proportion of the distance between phase points less than ε).

[0082] Joint analysis and parameter matrix generation: Slide and calculate the LLE and K entropy according to a time window (such as 100 ms) to form a time series curve. For example, in a certain collision event: Before the collision: LLE≈0.1, K entropy≈0.3, and the system is in a weak chaotic state.

[0083] At the moment of collision (t = 0 ms): LLE suddenly rises to 1.5, and the K entropy reaches 2.0, indicating severe chaos.

[0084] During the decay stage after the collision (t = 50 - 200 ms): The LLE / K entropy gradually decreases, reflecting the energy dissipation process.

[0085] Finally, arrange these parameters in the time - space dimension to generate a three - dimensional chaotic dynamics parameter matrix containing LLE, K entropy, and fractal dimension.

[0086] Step 4: Input the chaotic dynamics parameter matrix into the geometric topology mapping module, extract the topological invariant features of the phase trajectory based on the persistent homology algorithm, and output a chaotic feature fingerprint spectrum containing the dimension of the strange attractor, fractal structure, and recurrence characteristics.

[0087] Specific implementation: Persistent Homology is a topological data analysis method that captures the topological invariant features (such as connected components, loops, holes, etc.) of the phase space structure through multi-scale filtering and quantifies their "persistence" (i.e., the duration of existence during parameter changes).

[0088] Construct the Vietoris-Rips complex: 1. With the phase points as the core, gradually increase the radius ε to construct a simplicial complex: When ε = 0, each phase point exists independently.

[0089] As ε increases, edges (1-simplices) are formed between phase points with a distance less than ε.

[0090] Furthermore, three interconnected edges form a face (2-simplex), and so on.

[0091] 2. Record the homology groups of the complex at different ε values (e.g., H0 is the number of connected components, H1 is the number of loops, etc.). For example, for a certain attractor, 10 connected components are formed when ε = 0.1, merged into 1 when ε = 0.3, and 3 loop structures appear simultaneously.

[0092] Generate the persistence bar chart: For each topological feature (such as a loop), record its birth radius ε_birth and death radius ε_death, and plot it as an interval bar of (ε_birth, ε_death). For example, the H1 bar chart of a certain collision event shows a persistent loop (ε_birth = 0.2, ε_death = 0.5), indicating the existence of a stable ring structure in the attractor.

[0093] Extract the topological fingerprint: 1. Dimension of the strange attractor: Determine the scale of the main connected component through the longest persistence interval of the H0 bar chart.

[0094] 2. Fractal structure: The number and length of persistent loops in the H1 bar chart reflect the winding complexity of the attractor. For example, a high-intensity collision generates 5 persistent loops, while a low-intensity collision generates only 1.

[0095] 3. Recurrence characteristics: Analyze the self-similarity of the attractor through the distribution of the H2 bar chart (holes). If multiple holes appear repeatedly at different scales, the system has fractal recurrence characteristics.

[0096] Output of chaotic characteristic fingerprint spectrum: Fuse topological features with chaotic parameters (LLE, K-entropy) to generate a multi-dimensional feature vector. For example: Dimension index: [Radius of the main connected component = 0.35, Number of loops = 3, Number of holes = 2]; Fractal index: [Hausdorff dimension = 2.7, Correlation dimension = 2.3]; Recursion index: [Density of the diagonal structure in the recurrence plot = 0.8].

[0097] Finally, these parameters form a fingerprint spectrum that can uniquely identify the collision type, which is used for subsequent pattern classification and alarm decision-making.

[0098] In another implementation, specifically, perform chaotic invariance analysis on the core phase trajectory cluster, jointly calculate the maximum Lyapunov exponent and Kolmogorov entropy, and generate a chaotic dynamics parameter matrix, which may include: According to the geometric distribution characteristics of the core phase trajectory cluster, use the multifractal spectrum analysis method to calculate the Hölder exponent of each phase trajectory, and generate a fractal dimension weight matrix. Among them, the multifractal spectrum analysis method detects the distribution of singular points on the edge of the phase trajectory through the wavelet transform modulus maximum chain; this step uses the Daubechies wavelet (order N = 6) to perform multi-scale decomposition on the phase trajectory, and locates the singular points (Hölder exponent α ∈ [0.2, 0.8]) through the modulus maximum chain tracking technology. Use the box-counting method to calculate the local dimension spectrum f(α), generate a weight matrix (resolution 0.01), and quantify the contribution degree of fractal features at different scales. Multifractal analysis can identify weak chaotic features (α < 0.5) generated by minor collisions, and the weight matrix provides an adaptive adjustment basis for subsequent index calculations, improving the detection sensitivity to minor collisions.

[0099] In chaotic dynamics analysis, the geometric distribution characteristics of the core phase trajectory cluster reflect the inherent nonlinear dynamic behavior of collision events. To implement multifractal spectrum analysis, it is first necessary to construct a wavelet transform modulus maximum chain to capture the local singular points on the edge of the phase trajectory. Here, the Mexican Hat wavelet is used as the basis function. Due to its symmetry and sensitivity to signal mutation points, it can effectively identify the regions where the energy of the phase trajectory mutates. For example, in the scenario of guardrail collision, when the vehicle impacts the guardrail at different angles, the phase trajectory will show different densities of singular point distributions: low-speed collisions may only generate sparse singular points, while high-speed collisions will form dense modulus maximum chains at the edge of the phase space.

[0100] In practice, the core phase trajectory cluster is first decomposed by multi-scale wavelet. By adjusting the wavelet scale parameters, the local features of the phase trajectory are scanned layer by layer from coarse to fine. For the decomposition results at each scale, the modulus maxima points (i.e., local extreme points of the wavelet coefficients) are extracted, and these points are connected into a chain structure through the principle of spatial continuity to form a modulus maximum chain. The distribution density and topological structure of these chains are directly related to the multi-fractal characteristics of the collision event. For example, when the guardrail material breaks, the acoustic emission signal will trigger a high-density, short-period modulus maximum chain in the phase trajectory.

[0101] Next, based on the distribution of the modulus maxima chain, the Hölder index of each phase trajectory segment is calculated. The Hölder index is used to quantify the local singularity of the signal: the smaller the index value, the more drastic the signal mutation in this area. The specific method is to perform log-linear regression analysis on each modulus maximum chain and estimate the Hölder index by the decay rate of the wavelet coefficient with scale. For example, in the early stage of guardrail collision, due to the diffusion effect of stress wave propagation, the Hölder index of the phase trajectory may show a gradient distribution from low to high; in the material failure stage, the index will drop sharply due to local strain concentration.

[0102] Finally, the Hölder index of all phase trajectory segments is weighted and aggregated according to the spatial position to generate a fractal dimension weight matrix. Each element of the matrix corresponds to a local area in the phase space, and the weight value reflects the dynamic complexity of the area. For example, on the collision energy propagation path, the high-value area of the weight matrix may correspond to the weak point of the guardrail support structure, providing key input for subsequent damage assessment.

[0103] ‌The fractal dimension weight matrix is input into the local linearization LLA algorithm, and the phase trajectory tangent space is fitted based on the weighted least squares method. The noise sensitivity coefficient in the Lyapunov exponent calculation is dynamically corrected, and the Lyapunov exponent sequence with the greatest robustness is output; this step adopts the LLA (local linear approximation) algorithm, and the weighted objective function (regularization parameter λ=0.1) is constructed with the fractal weight as the coefficient, and the Jacobian matrix of the phase trajectory tangent space is solved by SVD decomposition. The noise suppression factor β (range 0.1-1.0) is dynamically adjusted to reduce the calculation variance of the maximum Lyapunov exponent λ1 by more than 60%. The weighted fitting effectively suppresses the influence of measurement noise on the chaotic parameters, and the coefficient of variation of the output exponential sequence is <5%, ensuring the comparability of different collision events.

[0104] The core objective of the Local Linear Approximation (LLA) algorithm is to accurately estimate the maximum Lyapunov exponent by modeling the local geometric structure of the phase trajectory. This exponent reflects the sensitivity of the system to initial conditions and is a key indicator for judging chaotic characteristics. However, traditional methods are prone to estimation biases due to phase trajectory perturbations in a noisy environment. Therefore, introducing a fractal dimension weight matrix as prior information for the weighted least squares method can significantly enhance robustness.

[0105] The specific implementation is divided into three stages: Neighborhood construction and tangent space fitting: For each reference point in the phase space, K neighboring points (e.g., K = 20) within its neighborhood are selected based on the Euclidean distance. The local tangent space is fitted using the weighted least squares method, where the weights are provided by the fractal dimension weight matrix. For example, in regions with a higher fractal dimension (corresponding to complex dynamic behaviors), larger weights are assigned to enhance the model's ability to capture nonlinear features.

[0106] Noise sensitivity coefficient correction: By analyzing the tangent space fitting residuals, the noise sensitivity coefficient is dynamically adjusted. If the residuals in a certain region are significantly higher than the average level, it is determined that there is strong noise interference there, and its contribution to the calculation of the Lyapunov exponent is reduced. For example, in the vibration signal of the guardrail, the pseudo-phase trajectories caused by wind noise usually exhibit low fractal dimensions and high residuals, which can be automatically suppressed by this mechanism.

[0107] Exponent sequence generation: Along the evolution direction of the phase trajectory, the divergence rate of the tangent space is calculated point by point. The maximum Lyapunov exponent sequence is obtained through time averaging. For example, in the process of collision energy diffusion, the exponent value will decay as the propagation distance increases, and this characteristic can be used to distinguish the main collision area from the secondary vibration area.

[0108] In practical applications, the LLA algorithm needs to optimize parameters in combination with the physical properties of the guardrail material. For example, aluminum alloy guardrails have a relatively high elastic modulus, and the divergence rate of the tangent space of their phase trajectories changes relatively gently, so a larger neighborhood radius needs to be set; while composite material guardrails require an adaptive neighborhood selection strategy due to their anisotropy.

[0109] Perform information geometric manifold modeling on the core phase trajectory cluster, quantify the information flow density in the phase space by calculating the Ricci curvature tensor, dynamically adjust the estimation window of the Kolmogorov entropy in combination with the Kontsevich entropy formula, and generate an entropy value evolution spectrogram; this step maps the phase trajectory to a Riemannian manifold, and uses the Ricci curvature (calculation step ε = 0.05) to characterize the information transmission efficiency. Based on the Kontsevich formula, adaptively select the entropy estimation window (the length is adjustable from 50 to 500 points), generate a two-dimensional time-entropy value spectrogram (resolution 1 bit / s / Hz), reveal the non-equilibrium dissipation process of the collision energy, the information geometric method breaks through the stationarity assumption of traditional entropy calculation, and the spectrogram can clearly show the entropy mutation at the moment of collision (amplitude > 3 bit / ms), improving the event positioning accuracy to ±10 ms.

[0110] Information geometric manifold modeling regards the phase space as a differential manifold, and its geometric structure is determined by the statistical characteristics of the dynamic system. The Ricci curvature tensor is used in this model to characterize the convergence or divergence characteristics of information propagation, and further quantify the entropy production rate of the chaotic system. To achieve this goal, first, a discretized manifold representation of the phase trajectory needs to be constructed.

[0111] The specific steps are as follows: Manifold discretization: Map the core phase trajectory cluster into a graph structure, where each phase trajectory point is used as a graph node, and the edge weight between nodes is determined by the geodesic distance (or approximated by the Euclidean distance) between phase points. For example, in the guardrail collision event, the nodes corresponding to the high-frequency vibration region will form a densely connected subgraph.

[0112] Ricci curvature calculation: Adopt the discretized definition of Ollivier-Ricci curvature, and estimate the curvature by comparing the difference between the geodesic distance and the Euclidean distance in the node neighborhood. For example, in the energy concentration region (corresponding to a high strain gradient), the curvature value is usually negative, indicating that the information flow shows a divergent characteristic here.

[0113] Dynamic entropy value estimation: Combine the Kontsevich entropy formula and dynamically adjust the estimation window of the Kolmogorov entropy according to the curvature distribution. When a negative curvature region is detected, narrow the window to capture the rapid entropy increase process; in the positive curvature region, expand the window to improve statistical stability. For example, at the initial stage of the collision, multiple negative curvature hotspots appear in the phase space, and the system will switch to a high time-resolution mode to track the entropy mutation in real time.

[0114] The finally generated entropy value evolution spectrogram is a spatio-temporal matrix, whose rows correspond to time windows, columns correspond to phase space regions, and the element values are normalized entropy values. This spectrogram can intuitively display the propagation path of the collision energy. For example, at the root of the guardrail column, the entropy value will continuously increase due to stress concentration until the material damage threshold triggers an alarm.

[0115] Input the Lyapunov exponent sequence and the entropy value evolution spectrogram into the Dynamic Mode Decomposition (DMD) module, solve the joint distribution function of chaotic dynamics parameters through the Tikhonov regularization method, and finally output a three-dimensional chaotic dynamics parameter matrix including fractal dimension, exponential sensitivity, and entropy production rate. In this step, DMD (Dynamic Mode Decomposition) is used to extract the dominant modes (truncated rank r = 5), solve the ill-conditioned equation through Tikhonov regularization (parameter γ = 0.01), and establish the joint probability distribution of λ1 - entropy - fractal dimension (kernel density estimation bandwidth h = 0.3). The output matrix contains 15 characteristic dimensions, which can explain more than 90% of the system variance. The three-dimensional parameter matrix comprehensively characterizes the transient dynamics of the collision, providing an evidence chain from multiple perspectives for quantitative evaluation and being more reliable than a single index judgment.

[0116] Dynamic Mode Decomposition (DMD) is a data-driven dimensionality reduction technique that can extract the dominant modes and their evolution laws from high-dimensional dynamic data. In this step, DMD is used to fuse the Lyapunov exponent sequence (characterizing system stability) and the entropy value evolution spectrogram (characterizing disorder) to generate a unified dynamics parameter matrix. A specific implementation includes: Data matrix construction: Arrange the Lyapunov exponent sequence in a row vector aligned by time, and expand the entropy value spectrogram as a column vector according to the spatial region, and combine them into a high-dimensional observation matrix. For example, for a system with 100 time points and 50 spatial regions, the dimension of the observation matrix is 100×50.

[0117] Tikhonov regularization solution: To avoid overfitting, Tikhonov regularization is introduced in the matrix decomposition process of DMD. The regularization parameter is adaptively selected by the L-curve method to balance the model complexity and fitting accuracy. For example, when the environmental noise is strong, the system will automatically increase the regularization coefficient to suppress the spurious modes.

[0118] Mode screening and parameter fusion: Sort the DMD modes obtained by decomposition according to the energy proportion, and retain the top K dominant modes (such as K = 5). Each mode contains the spatial distribution pattern and its growth / decay rate. Feature fusion is achieved by performing the Hadamard product of the fractal dimension weight matrix and the modal spatial distribution. For example, a certain mode may have a high amplitude in the guardrail crossbeam area and a low value corresponding to the fractal dimension at the same time, indicating a simplified vibration mode at that place.

[0119] Three-dimensional parameter matrix generation: The three dimensions of the finally output matrix correspond to the fractal dimension, exponential sensitivity, and entropy production rate respectively. Each element value comprehensively reflects the dynamic characteristics of a specific spatio-temporal position. For example, the combination of a high fractal dimension, high sensitivity, and high entropy production rate of an element in the matrix may indicate a critical area where material failure is about to occur.

[0120] In actual deployment, this parameter matrix can be directly input into the decision-making system. For example, when it is detected that the entropy production rate in a certain area exceeds the threshold and is accompanied by a sharp increase in exponential sensitivity, the system will immediately trigger enhanced monitoring and early warning for that location.

[0121] S203. Perform collision intensity grading on the chaotic feature fingerprint map, construct a collision energy propagation model using a spatio-temporal convolutional spiking neural network, and fuse the strain gradient distribution, energy decay rate, and material damage threshold parameters through a dynamic weight allocation mechanism to output a quantitative evaluation matrix containing the collision level, damage radius, and predicted value of the residual strength. This step uses a spiking neural network (SNN) to simulate the collision energy propagation process. Combining with a dynamic weight allocation mechanism, it comprehensively evaluates the strain gradient, energy decay, and material damage threshold, outputs a quantitative matrix, accurately predicts the collision level, damage range, and the residual strength of the guardrail, realizes the precise grading of collision intensity, provides a scientific basis for subsequent alarm decisions, and improves the intelligent level of guardrail safety early warning.

[0122] Specifically, according to the chaotic feature fingerprint map, the dimension of the strange attractor can be mapped to the collision energy space through topological data binning technology to construct multi-level classification boundary conditions for collision intensity; use topological data binning (TDA) to divide the attractor dimension (such as 2.1 - 3.5 dimensions) into energy level intervals (such as 0 - 5 kJ for slight, 5 - 20 kJ for medium), and set dynamic classification boundaries in combination with material mechanics experimental data (such as the plastic deformation threshold of an aluminum alloy guardrail corresponds to 20 kJ) to avoid misclassification caused by fixed thresholds, realize the adaptive grading of collision intensity. The energy space mapping is more stable than directly using the original signal, and is especially suitable for the general evaluation of multi-type guardrails (steel / composite materials).

[0123] Input the multi-level classification boundary conditions into the spatio-temporal convolutional spiking neural network, use the LIF neuron model to simulate the biological pulse triggering mechanism, dynamically model the collision energy propagation path, and output a probability density function containing spatio-temporal energy distribution; the spatio-temporal convolutional spiking neural network (SC-SNN) simulates the synaptic transmission characteristics through LIF neurons. After inputting the boundary conditions, the convolutional kernel captures the attenuation law of energy in the guardrail length direction (space) and millisecond-level time periods (time) (such as the exponential decay coefficient α = 0.3), and outputs a probability density function to characterize the energy hot spot distribution. The spiking neural network is better at dealing with millisecond-level transient impacts than the traditional RNN. The probability density function quantifies the uncertainty of energy propagation and provides a probability basis for damage prediction.

[0124] Optimize the physical constraints of the probability density function, construct a dynamic weight allocation function based on the material damage constitutive equation, fuse the strain gradient distribution, energy decay rate, and material fatigue threshold parameters to generate a multi-objective coupled damage evolution model; use the probability density function as the initial input, introduce the material constitutive equation (such as the Johnson-Cook model) to constrain the energy-strain relationship, dynamically adjust the weights (such as the strain gradient weight of 0.6 and the fatigue threshold weight of 0.3), solve the damage evolution equation by the Lagrange multiplier method, and output the predicted remaining life value (such as 70% of the residual strength). Physical constraints avoid the overfitting problem of pure data-driven methods. The multi-objective coupled model simultaneously considers instantaneous impact and cumulative fatigue damage, improving the comprehensiveness of prediction.

[0125] Input the damage evolution model into the Monte Carlo-Finite Element hybrid solver, and finally output a quantitative evaluation matrix including the collision level, damage radius, and residual strength through the alternating iteration of random sampling and deterministic solution. The hybrid solver performs Monte Carlo random sampling on the energy distribution (such as 1000 times). After each sampling, the finite element solver (such as ANSYS LS-DYNA) is called to calculate the local stress concentration. After iterative convergence, the confidence intervals of the collision level (such as Level 3), damage radius (such as ±1.5 m), and residual strength (such as 82 MPa ± 5%) are statistically analyzed. Combining the advantages of probability and deterministic methods, the quantitative evaluation matrix supports risk decisions (such as whether to close the lane) and is more reliable than a single deterministic result.

[0126] Exemplarily, a specific implementation method is as follows: Step 1: According to the chaotic characteristic fingerprint spectrum, map the dimension of the strange attractor to the collision energy space through the topological data binning technique to construct multi-level classification boundary conditions for the collision intensity. The core of this step is to convert the chaotic dynamics parameters into engineering-interpretable collision intensity levels. First, it is necessary to establish the mapping relationship between chaotic characteristics and physical energy. In specific implementation, the topological data binning technique (Topological Data Binning) is used to analyze the dimension of the strange attractor in the chaotic characteristic fingerprint spectrum. For example, the persistent homology algorithm is used to extract the persistent periods of the circular structure (H1 feature) and the hole structure (H2 feature) of the attractor, and the topological features of different dimensions are associated and mapped with the energy calibration data obtained from laboratory collision tests.

[0127] During actual operation, the system will pre-store a database of attractor topological features corresponding to typical collision scenarios (such as low-speed rubbing of a vehicle at 5 km / h, medium-speed collision at 20 km / h, and high-speed impact at 50 km / h). When the dimension data of the real-time collected strange attractor is input, the Wasserstein distance (i.e., the optimal transport cost) between it and the database samples is calculated to determine the closest collision energy interval. For example, when it is detected that the continuous period of the H1 dimension of the attractor is in the range of 0.8 - 1.2 seconds and a triple bifurcation structure appears in the H2 dimension, it can be correspondingly mapped to the medium collision intensity interval with an energy value of 15 - 25 kJ.

[0128] The construction of the classification boundary conditions uses the Support Vector Domain Description (SVDD) algorithm to establish a non-linear decision boundary in the energy-topological feature space. For example, in a three-dimensional feature space, the H1 continuous period, Hurst exponent, and maximum Lyapunov exponent are used as input features. After mapping the data to a high-dimensional space through the kernel trick, the smallest hypersphere containing the same-class data points is found. When the real-time data point is outside the hypersphere, it is determined as a higher-level collision event. The system sets three levels of collision intensity classification: Level 1 (energy < 10 kJ), Level 2 (10 - 30 kJ), Level 3 (> 30 kJ), and each level corresponds to different hypersphere radius and center position parameters.

[0129] Step 2: Input the multi-level classification boundary conditions into the spatio-temporal convolutional spiking neural network, use the LIF neuron model to simulate the biological pulse triggering mechanism, dynamically model the collision energy propagation path, and output the probability density function containing the spatio-temporal energy distribution; The spatio-temporal convolutional spiking neural network (STC-SNN) constructed in this step is a new hybrid architecture that integrates spatial convolution and temporal pulse mechanisms. The network input layer receives the multi-level collision feature vectors generated by the classification boundary conditions. The first layer uses a three-dimensional convolutional kernel (size 3×3×5) to extract features from the spatial position (X, Y coordinates) of the guardrail structure and the time window (5 sampling points). For example, when a sudden increase in strain is detected within a 20 ms time window at the 12th node position of the guardrail, the convolutional layer will capture the propagation delay pattern of this event at adjacent nodes (11th, 13th).

[0130] The core processing layer adopts the Leaky Integrate-and-Fire (LIF) neuron model, and the membrane potential dynamics follow the exponential decay mechanism. In specific implementation, the membrane potential update formula for each neuron is as follows: when an input pulse arrives, the membrane potential accumulates according to an exponential function, and when it reaches the threshold (such as 30 mV), an output pulse is triggered and the potential is reset. For example, when simulating energy propagation, the neuron cluster in the 5th layer will, after receiving a pulse sequence from the collision starting point, reconstruct the propagation speed (such as 150 m / s) and attenuation characteristics of the energy wavefront in the guardrail structure through the spatio-temporal integration effect of the membrane potential.

[0131] The network training adopts an improved spatio-temporal backpropagation (STBP) algorithm, combined with the local spike-timing-dependent plasticity (STDP) rule. The loss function is defined as the Wasserstein distance between the predicted energy distribution and the actual measurement results of a laser Doppler vibrometer. The training dataset contains 200 sets of real vehicle collision test data with different collision angles (15° - 75°), speeds (10 - 60 km / h), and contact areas (5 - 50 cm²). The trained network can accurately predict the energy propagation path. For example, in a 30 km / h oblique collision scenario, the peak region of the output energy density will show an elliptical expansion pattern in the 45° direction.

[0132] Step 3: Physically constrain and optimize the probability density function, construct a dynamic weight allocation function based on the material damage constitutive equation, fuse the strain gradient distribution, energy decay rate, and material fatigue threshold parameters to generate a multi-objective coupled damage evolution model; In this stage, the data-driven prediction results are combined with physical mechanisms. First, establish the damage constitutive equation of Q345 steel, and its dynamic yield strength can be expressed as a piecewise function of the strain rate: when the strain rate is lower than 10 s -1 the Johnson-Cook model is used, and when it is higher than this value, the Zerilli-Armstrong model is used. For example, at the moment of collision (strain rate is about 50 s -1 ), the material yield strength will increase from 345 MPa in the static state to 420 MPa.

[0133] The design of the dynamic weight distribution function adopts a multi-objective optimization framework. Three optimization objectives are defined: (1) minimizing the mean square error between the strain gradient distribution and the digital image correlation (DIC) measurement results; (2) maximizing the matching degree between the energy decay rate and the energy spectrum of the acoustic emission signal; (3) ensuring that the damage accumulation does not exceed the material fatigue threshold (set to 0.95 according to the Miner linear damage criterion). The Pareto front is solved by the NSGA-II multi-objective genetic algorithm to obtain the weight coefficients of each parameter. For example, at the guardrail post part, the weight of the strain gradient may reach 0.6, while at the crossbeam part, the weight of the energy decay rate is dominant (0.55).

[0134] The damage evolution model integrates the continuum damage mechanics (CDM) and the discrete element method (DEM). At the microscale, the Voronoi mesh is used to divide the material grain structure, and the damage initiation criterion at the grain boundary is defined: when the local equivalent strain exceeds 0.15, the connection stiffness between adjacent grain units begins to linearly decay. At the macroscale, the cohesive zone model (CZM) is used to describe the crack propagation behavior, and the mode-I fracture toughness KIC = 120 MPa·m^0.5 is set. For example, the simulation shows that when the collision energy reaches 28 kJ, the damage zone will extend along the 45° direction from the collision point, forming a crack path with a length of about 80 cm.

[0135] Step 4: Input the damage evolution model into the Monte Carlo - finite element hybrid solver, and through the alternating iteration of random sampling and deterministic solution, finally output a quantitative evaluation matrix including the collision level, damage radius, and residual strength.

[0136] The core of the hybrid solver is to alternately perform probability sampling and deterministic calculation. The Monte Carlo module first performs Latin hypercube sampling on the material parameters (such as elastic modulus, Poisson's ratio), considering the influence of manufacturing tolerances of ±5%. For example, 1000 samplings are performed on the yield strength of Q345 steel to generate a parameter set with a normal distribution of N(345, 15^2) MPa. At the same time, the initial defect distribution of the guardrail structure is randomly generated, including welding residual stress (20 - 50 MPa) and microcrack size (0.1 - 0.5 mm).

[0137] The explicit dynamic analysis is adopted in the finite element solution stage, and the time step is set to 1e - 6 seconds according to the Courant condition. Each Monte Carlo sample corresponds to an Abaqus explicit solution task to calculate the maximum equivalent plastic strain (PEEQ) and the damage variable (DAMAGE). For example, in a certain sample, the position where the maximum PEEQ reaches 0.25 at 0.1 second after the collision is determined as the plastic hinge formation region.

[0138] The alternating iteration process is optimized through convergence judgment. After each round of iteration, the coefficient of variation of the damage radius (the diameter of the circumscribed circle of the area where PEEQ > 0.2) is statistically analyzed. When the coefficient of variation is < 2% for three consecutive rounds, the calculation is terminated. The final output matrix includes: collision level (Level 1 - 3), damage radius (mean ± standard deviation, e.g., Level 2 corresponds to 1.2 ± 0.3 m), and residual strength coefficient (the ratio of the current bearing capacity to the initial value). These parameters are encapsulated in JSON format, including metadata such as timestamp, GPS coordinates, and sensor node ID, providing a quantitative basis for subsequent early warning decisions.

[0139] S204, trigger a multi-modal alarm protocol according to the quantization evaluation matrix, dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack, and encode the alarm information into a time-varying frequency-hopping pulse sequence using chaotic encryption modulation technology, and synchronously transmit it to the traffic management center and neighboring vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

[0140] This step uses the LoRaWAN protocol to dynamically allocate spectrum resources and combines chaotic encryption technology to ensure the secure transmission of alarm information. The game model optimizes the transmission strategy, reduces the impact of channel interference, achieves sub-second alarm response, ensures that information is quickly delivered to the management center and surrounding vehicles, improves the real-time performance and security of alarm information transmission, reduces the secondary risk of traffic accidents, and enhances the road safety protection ability.

[0141] Specifically, according to the collision level parameter in the quantization evaluation matrix, a priority-based multi-modal alarm protocol can be triggered, dynamically allocate communication spectrum resources through the LoRaWAN protocol stack, and generate a channel occupancy status map; this step dynamically activates different priority alarm modes (such as starting an emergency broadcast above Level 3) by analyzing the collision level in the quantization evaluation matrix (such as Level 1 - 5). Utilize the ADR (Adaptive Data Rate) function of the LoRaWAN protocol stack to real-time scan the RSSI (Received Signal Strength) and SNR (Signal-to-Noise Ratio) of the 868 MHz / 915 MHz frequency band, generate a two-dimensional channel occupancy heat map (resolution up to 100 kHz / 1 ms), provide a decision basis for subsequent spectrum allocation, realize the intelligent matching of collision severity and communication resource allocation, and the channel status map can avoid congested frequency bands, improve the transmission success rate of alarm information, and is especially suitable for the scenario of concurrent alarms of multiple guardrail nodes.

[0142] The alarm information and the channel occupancy status spectrum are input into the chaotic encryption modulation module. A time-varying key stream is generated based on the Lorenz equation. The differential frequency hopping technology is used to encode the data into chaotic pulse sequences, and an anti-jamming encrypted signal is output. In this step, the Lorenz chaotic system (parameters σ = 10, β = 8 / 3, ρ = 28) is used to generate an unpredictable key sequence, and the optimal frequency hopping mode is selected in combination with the channel spectrum (for example, the carrier frequency is switched every 50 ms, and the frequency hopping interval is ≥ 1 MHz). The alarm data (latitude and longitude, collision level) is encoded into chaotic phase pulses through DQPSK (Differential Quadrature Phase Shift Keying) modulation. The output signal has noise-like characteristics (peak-to-average ratio is lower than 2 dB). Chaotic encryption makes the signal difficult to be intercepted or forged, and the differential frequency hopping technology can resist multipath fading and narrowband interference, ensuring communication security in the complex electromagnetic environment of highways.

[0143] Construct a joint game model of collision events and channel interference, solve the optimal transmission strategy through Nash equilibrium, dynamically adjust the transmission power and frequency hopping interval, and optimize the end-to-end transmission delay to the sub-second level. In this step, a non-cooperative game model is established: player A (collision event) aims to minimize the delay, and player B (channel interference) attempts to maximize the bit error rate. The Q-learning algorithm (learning rate α = 0.01) is used to iteratively solve the Nash equilibrium point, dynamically optimize the transmission power (adjustable from 14 - 20 dBm) and frequency hopping interval (adaptive from 50 - 200 ms), and converge to the optimal strategy within 500 ms. Game optimization enables the system to automatically increase the power or shorten the frequency hopping period when the channel is congested. The measured end-to-end delay can be controlled within 800 ms, meeting the response requirements of the AEB (Automatic Emergency Braking) system.

[0144] The anti-jamming encrypted signal is broadcast through a multi-antenna MIMO array and synchronously transmitted to the traffic management center and adjacent vehicle OBU terminals, and ultra-low bit error rate parsing at the receiving end is realized based on the Turbo decoding algorithm. In this step, a 4×4 MIMO antenna array (beamforming gain ≥ 12 dBi) is used for space diversity transmission, and the Alamouti space-time coding is combined to enhance the diversity gain. The receiving end uses a parallel concatenated Turbo code (iterated 8 times, constraint length K = 3) for decoding. The measured bit error rate is lower than 10 -6 (when SNR ≥ 6 dB), ensuring that the coordinate information error is < 0.5 meters. The MIMO multi-antenna technology extends the communication coverage radius to 300 meters, and Turbo decoding ensures the accurate restoration of key information, avoiding incorrect path planning due to bit errors.

[0145] A specific implementation method is as follows: Step 1: According to the collision level parameter in the quantization evaluation matrix, trigger a priority-based multi-modal alarm protocol, dynamically allocate communication spectrum resources through the LoRaWAN protocol stack, and generate a channel occupancy status spectrum; The core of this step is to achieve intelligent scheduling of alarm information transmission resources. First, the system defines the alarm priority according to the collision levels (such as Level 1-3) in the quantization evaluation matrix. For example: Level 1 (low risk): Trigger local LED flashing and short-range Bluetooth broadcasting, and set the priority weight to 0.3; Level 2 (medium risk): Start LoRaWAN regional broadcasting, with a priority of 0.6; Level 3 (high risk): Activate the cellular network (such as LTE-M) and vehicle-to-vehicle (V2X) communication with neighboring vehicles, with a priority of 1.0.

[0146] In the LoRaWAN spectrum allocation, an Adaptive SF Allocation algorithm is adopted. The system scans 16 channels in the 868 MHz band in real time, and identifies idle channels through energy detection (ED) and preamble sniffing techniques. For example, when the RSSI (Received Signal Strength Indicator) of channel 5 is detected to be lower than -110 dBm, it is marked as an available channel. The dynamic allocation strategy uses weighted round-robin scheduling (WRR), allocating spreading factor SF7 (fast transmission rate but short distance) for high-priority alarms, and SF12 (slow rate but wide coverage) for low-priority ones. The channel occupancy status map is presented in the form of a two-dimensional heat map, with the horizontal axis representing frequency (125 kHz step) and the vertical axis representing time slots (one unit every 30 ms). Red indicates high occupancy areas, and green indicates available areas. For example, within 200 ms after a collision, the system may select channel 3 (center frequency 868.3 MHz) and channel 11 (869.5 MHz) as the primary and backup channels.

[0147] Step 2: Input the alarm information and the channel occupancy status map into the chaotic encryption modulation module, generate a time-varying key stream based on the Lorenz equation, and encode the data into chaotic pulse sequences using differential frequency hopping technology to output anti-jamming encrypted signals; The implementation of chaotic encryption is divided into two stages: key generation and signal modulation. First, the Lorenz equation is used as the chaotic source, and its parameters are set as follows: Initial values: x0 = 10.0, y0 = 28.0, z0 = 8 / 3 (x0, y0, z0 are classical chaotic parameters); Step size Δt = 0.01 seconds, and the first 500 transient processes are truncated after 1000 iterations.

[0148] The generated chaotic sequence is converted into a binary key stream through quantization sampling. For example, take the 501-600th iteration values of the x variable. When x nOutput 1 when > 25, otherwise output 0, and generate a 100-bit key. This key stream is XORed with the alarm information (such as collision level, GPS coordinates) to achieve encryption. For example, the original data byte 0x3A (00111010) is XORed with the key 0xB5 (10110101) to obtain 0x8F (10001111).

[0149] The specific process of differential frequency hopping (DFH) modulation is as follows: 1. Divide the encrypted data stream into groups of 4 bits (such as 1000 1101), and each group is mapped to a frequency hopping pattern; 2. Select available frequency points according to the channel occupancy map and design the frequency hopping sequence. For example, in the available channels 3, 7, 11, set the frequency hopping interval to 10 ms, and the sequence is 3 → 11 → 7 → 3 →...; 3. Adopt Gaussian frequency shift keying (GFSK) modulation, and the carrier frequency changes according to the frequency hopping sequence within each symbol period. For example, the symbol "1" is represented by a +20 kHz frequency offset, and "0" is represented by a -20 kHz frequency offset.

[0150] The finally output chaotic pulse sequence has time-varying characteristics. For example, 5 frequency hops are completed within 50 ms, and the frequency change trajectory is 868.3 MHz → 869.5 MHz → 867.1 MHz → 868.3 MHz → 867.9 MHz.

[0151] Step 3: Construct a joint game model of collision events and channel interference, solve the optimal transmission strategy through Nash equilibrium, dynamically adjust the transmit power and frequency hopping interval, and optimize the end-to-end transmission delay to the sub-second level; The game model regards the collision warning system and the channel interference source as two game participants: Participant A (warning system): The strategy space is the transmit power (10 - 30 dBm) and the frequency hopping interval (5 - 50 ms); Participant B (interference source): The strategies are the interference power (0 - 20 dBm) and the interference frequency point selection.

[0152] The revenue function is designed as: Revenue of the warning system = Transmission success rate × Priority weight - Delay penalty factor × Transmission time; Revenue of the interference source = Interference success rate - Energy consumption cost.

[0153] Adopt the Fictitious Play algorithm to solve the Nash equilibrium: 1. Each participant predicts the behavior of the other party according to the historical strategy distribution. For example, the warning system counts the distribution of interference frequency points in the past 100 times and predicts that the interference may be concentrated in channel 7 in the next time slot; 2. Update the policy through Q - learning reinforcement learning. For example, when the power is selected as 20 dBm and the hopping frequency interval is 20 ms, if the transmission is successful, the Q - value increases by 0.1, and if it fails, it decreases by 0.2; 3. Set the convergence condition as the difference in policy updates being less than 1% for 10 consecutive times.

[0154] Dynamic adjustment example: When the bit error rate (BER) of channel 7 is detected to exceed 1e - 3, the system automatically shortens the hopping frequency interval from 20 ms to 10 ms, and at the same time increases the transmission power from 17 dBm to 25 dBm. Experiments show that this policy can reduce the average delay from 1.2 seconds to 0.8 seconds, meeting the sub - second response requirement.

[0155] Step 4: Broadcast the anti - interference encrypted signal through a multi - antenna MIMO array, synchronously transmit it to the traffic management center and neighboring vehicle OBU terminals, and implement ultra - low bit error rate parsing at the receiving end based on the Turbo decoding algorithm.

[0156] The MIMO transmission adopts a 4×4 antenna configuration, and the specific implementation includes: 1. Space - Time Block Coding: Encode every two encrypted symbols (such as 0x8F and 0x3A) into a 4 - antenna transmission matrix. For example, in the Alamouti coding scheme, antennas 1 and 2 send symbols S1 and S2 in time slot 1; 2. Beamforming: Calculate the steering vector through the channel state information (CSI). For example, using the minimum mean square error (MMSE) algorithm, generate a beam with a main lobe width of 10° in the direction of the traffic management center; 3. Diversity combining: The receiving end uses maximum ratio combining (MRC) and weights and sums according to the signal - to - noise ratio (SNR) of each antenna. For example, if the SNRs of antennas 1 - 4 are 15 dB, 18 dB, 12 dB, and 20 dB respectively, the weight ratio is 0.15:0.18:0.12:0.20.

[0157] At the decoding end, the working process of the Turbo decoder is as follows: 1. Two component decoders (SOVA algorithm) work alternately. The first decoder processes the systematic bits and parity check bits 1 and outputs the extrinsic information to the second decoder; 2. The second decoder uses the extrinsic information and parity check bits 2 for soft decision, and outputs the final hard decision result after 6 iterations; 3. Adopt CRC - 16 check. If the check fails, request re - transmission. Measured data shows that the bit error rate can be as low as 1e - 6 at a signal - to - noise ratio of 10 dB, which is two orders of magnitude better than traditional convolutional codes.

[0158] The synchronous transmission mechanism is implemented through the IEEE 1588 Precision Time Protocol (PTP). The master clock (traffic management center) periodically sends synchronization messages, and the OBU terminal calculates the network delay (such as ±50 μs) to compensate for the clock offset. For example, within 300 ms after a collision occurs, the warning message can reach the management center within a range of 800 m and 50 adjacent vehicle terminals simultaneously, and the end-to-end delay jitter is less than 20 ms.

[0159] It can be seen that, based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multi-modal sensor array deployed at the key nodes of the guardrail, the adaptive variational mode decomposition algorithm is used to perform non-linear phase alignment processing on the multi-source heterogeneous signals to generate a three-dimensional dynamic strain field distribution vector; the three-dimensional dynamic strain field distribution vector is input into the non-linear dynamics reconstruction module to extract the chaotic characteristic fingerprint spectrum of the collision event; the chaotic characteristic fingerprint spectrum is subjected to collision intensity grading processing, and a quantitative evaluation matrix including the collision level, damage radius, and predicted value of the residual strength is output; according to the quantitative evaluation matrix, a multi-modal warning protocol is triggered and synchronously transmitted to the traffic management center and the OBU terminals of adjacent vehicles, so that the state of the guardrail can be monitored in real time, and the collision event can be accurately evaluated and warned.

[0160] Another embodiment of the present invention provides a guardrail collision warning system. Refer to Figure 3 , the system may include: A processing module 301, configured to perform non-linear phase alignment processing on multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multi-modal sensor array deployed at the key nodes of the guardrail to generate a three-dimensional dynamic strain field distribution vector, wherein the sensor array uses a piezoelectric-fiber composite sensing unit and calculates the spatio-temporal correlation weight of each node signal based on mutual information entropy; An extraction module 302, configured to input the three-dimensional dynamic strain field distribution vector into the non-linear dynamics reconstruction module, construct a high-dimensional phase space projection based on the geometric invariance principle of the chaotic attractor, and extract the chaotic characteristic fingerprint spectrum of the collision event through the joint calculation of the largest Lyapunov exponent and the Kolmogorov entropy, wherein the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectory caused by environmental noise; A construction module 303, configured to perform collision intensity grading processing on the chaotic characteristic fingerprint spectrum, construct a collision energy propagation model using a spatio-temporal convolutional pulse neural network, and fuse the strain gradient distribution, energy decay rate, and material damage threshold parameters through a dynamic weight allocation mechanism to output a quantitative evaluation matrix including the collision level, damage radius, and predicted value of the residual strength; An alarm module 304 is used to trigger a multimodal alarm protocol according to the quantization evaluation matrix, dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack, encode alarm information into a time-varying frequency-hopping pulse sequence by using a chaotic encryption modulation technique, and synchronously transmit it to a traffic management center and adjacent vehicle OBU terminals. Wherein, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

[0161] It can be seen that, based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multimodal sensor array deployed at the key nodes of the guardrail, the multi-source heterogeneous signals are subjected to non-linear phase alignment processing through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector; the three-dimensional dynamic strain field distribution vector is input into a non-linear dynamics reconstruction module to extract the chaotic feature fingerprint spectrum of the collision event; the chaotic feature fingerprint spectrum is subjected to collision intensity grading processing to output a quantization evaluation matrix including the collision level, damage radius, and predicted value of the residual strength; according to the quantization evaluation matrix, a multimodal alarm protocol is triggered and synchronously transmitted to a traffic management center and adjacent vehicle OBU terminals, so that the state of the guardrail can be monitored in real time, and the collision event can be accurately evaluated and alarmed.

[0162] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and the computer program is set to execute the steps in any one of the above method embodiments when running.

[0163] Specifically, in this embodiment, the above storage medium can be set to store a computer program for executing the following steps: S201, based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multimodal sensor array deployed at the key nodes of the guardrail, the multi-source heterogeneous signals are subjected to non-linear phase alignment processing through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector, wherein the sensor array uses a piezoelectric-fiber composite sensing unit and calculates the spatio-temporal correlation weight of each node signal based on mutual information entropy; S202, input the three-dimensional dynamic strain field distribution vector into a non-linear dynamics reconstruction module, construct a high-dimensional phase space projection based on the geometric invariance principle of a chaotic attractor, and extract the chaotic feature fingerprint spectrum of the collision event through the joint calculation of the largest Lyapunov exponent and Kolmogorov entropy, wherein the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectory caused by environmental noise; S203. Perform a collision intensity grading process on the chaotic characteristic fingerprint spectrum, construct a collision energy propagation model using a spatio-temporal convolutional pulse neural network, and fuse the strain gradient distribution, energy attenuation rate, and material damage threshold parameters through a dynamic weight allocation mechanism to output a quantitative evaluation matrix including the collision level, damage radius, and predicted residual strength value. S204. Trigger a multi-modal alarm protocol based on the quantitative evaluation matrix, dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack, encode the alarm information into a time-varying frequency-hopping pulse sequence using chaotic encryption modulation technology, and synchronously transmit it to the traffic management center and adjacent vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

[0164] It can be seen that based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multi-modal sensor array deployed at the key nodes of the guardrail, the multi-source heterogeneous signals are subjected to non-linear phase alignment processing through the adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector; the three-dimensional dynamic strain field distribution vector is input into the non-linear dynamics reconstruction module to extract the chaotic characteristic fingerprint spectrum of the collision event; the chaotic characteristic fingerprint spectrum is subjected to a collision intensity grading process to output a quantitative evaluation matrix including the collision level, damage radius, and predicted residual strength value; a multi-modal alarm protocol is triggered based on the quantitative evaluation matrix and synchronously transmitted to the traffic management center and adjacent vehicle OBU terminals, so as to be able to monitor the status of the guardrail in real time and accurately evaluate and alarm collision events.

[0165] An embodiment of the present invention further provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0166] Specifically, the above electronic device may further include a transmission device and an input / output device. Among them, the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0167] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201. Based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multi-modal sensor array deployed at the key nodes of the guardrail, perform non-linear phase alignment processing on the multi-source heterogeneous signals through the adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector, where the sensor array uses a piezoelectric-optical fiber composite sensing unit and calculates the spatio-temporal correlation weight of each node signal based on mutual information entropy. S202. Input the three-dimensional dynamic strain field distribution vector into the nonlinear dynamics reconstruction module. Based on the geometric invariance principle of chaotic attractors, construct a high-dimensional phase space projection. Through the joint calculation of the maximum Lyapunov exponent and Kolmogorov entropy, extract the chaotic characteristic fingerprint spectrum of the collision event. Among them, the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectories caused by environmental noise. S203. Perform collision intensity grading processing on the chaotic characteristic fingerprint spectrum. Use a spatio-temporal convolutional pulse neural network to construct a collision energy propagation model. Through a dynamic weight allocation mechanism, fuse the strain gradient distribution, energy decay rate, and material damage threshold parameters, and output a quantitative evaluation matrix including collision level, damage radius, and predicted residual strength values. S204. Trigger a multi-modal alarm protocol according to the quantitative evaluation matrix. Dynamically allocate communication spectrum resources based on the LoRaWAN protocol stack. Use chaotic encryption modulation technology to encode the alarm information into a time-varying frequency hopping pulse sequence and synchronously transmit it to the traffic management center and adjacent vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

[0168] It can be seen that according to the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the multi-modal sensor array deployed at the key nodes of the guardrail, perform non-linear phase alignment processing on the multi-source heterogeneous signals through the adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector; input the three-dimensional dynamic strain field distribution vector into the nonlinear dynamics reconstruction module to extract the chaotic characteristic fingerprint spectrum of the collision event; perform collision intensity grading processing on the chaotic characteristic fingerprint spectrum and output a quantitative evaluation matrix including collision level, damage radius, and predicted residual strength values; trigger a multi-modal alarm protocol according to the quantitative evaluation matrix and synchronously transmit it to the traffic management center and adjacent vehicle OBU terminals, so as to be able to monitor the state of the guardrail in real time and accurately evaluate and alarm collision events.

[0169] The above has detailed the structure, characteristics, and effects of the present invention according to the illustrated embodiments. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope defined by the drawings. Any changes made according to the concept of the present invention, or modified into equivalent embodiments with equivalent changes, still within the spirit covered by the specification and drawings, shall be within the protection scope of the present invention.

Claims

1. A guardrail collision warning method, characterized in that, The method includes: Based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by a multimodal sensor array deployed at key nodes of the guardrail, the multi-source heterogeneous signals are processed for non-linear phase alignment through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector. Among them, the sensor array uses a piezoelectric-fiber composite sensing unit, and the spatio-temporal correlation weights of the signals at each node are calculated based on mutual information entropy; The three-dimensional dynamic strain field distribution vector is input into a non-linear dynamics reconstruction module. Based on the geometric invariance principle of chaotic attractors, a high-dimensional phase space projection is constructed. Through the joint calculation of the largest Lyapunov exponent and Kolmogorov entropy, the chaotic characteristic fingerprint spectrum of the collision event is extracted. Among them, the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectories caused by environmental noise; The chaotic characteristic fingerprint spectrum is processed for collision intensity grading. A spatio-temporal convolutional pulse neural network is used to construct a collision energy propagation model. Through a dynamic weight allocation mechanism, the strain gradient distribution, energy attenuation rate, and material damage threshold parameters are fused to output a quantitative evaluation matrix including the collision level, damage radius, and predicted value of the residual strength; According to the quantitative evaluation matrix, a multimodal alarm protocol is triggered. Based on the LoRaWAN protocol stack, communication spectrum resources are dynamically allocated. The alarm information is encoded into a time-varying frequency-hopping pulse sequence using chaotic encryption modulation technology and synchronously transmitted to the traffic management center and neighboring vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

2. The method according to claim 1, wherein Based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by a multimodal sensor array deployed at key nodes of the guardrail, the multi-source heterogeneous signals are processed for non-linear phase alignment through an adaptive variational mode decomposition algorithm to generate a three-dimensional dynamic strain field distribution vector. Among them, the sensor array uses a piezoelectric-fiber composite sensing unit, and the spatio-temporal correlation weights of the signals at each node are calculated based on mutual information entropy, including: Based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the piezoelectric-fiber composite sensing unit, the spatio-temporal correlation weights of the signals at each node are calculated through the mutual information entropy algorithm to generate a dynamic correlation matrix of multi-source signals. Among them, the mutual information entropy algorithm introduces a time-delay embedding technique to eliminate the heterodyne interference between sensors; The dynamic correlation matrix is input into an adaptive variational mode decomposition module. Based on the Kullback-Leibler divergence, the mode number and bandwidth parameters are dynamically optimized, and the multi-source heterogeneous signals are processed for non-linear phase alignment to output a set of decomposition modes with a unified time reference; Multiscale covariance analysis is performed on the set of decomposition modes. The principal component features of the strain field are extracted through an asymmetric tensor decomposition algorithm to generate a three-dimensional feature tensor including strain gradient, vibration energy, and acoustic emission event density; The three-dimensional feature tensor is input into a physical field fusion network. A dual-channel attention mechanism is used to perform cross-modal feature fusion on piezoelectric signals and fiber signals, and finally a spatio-temporally continuous three-dimensional dynamic strain field distribution vector is output.

3. The method according to claim 2, wherein The three-dimensional dynamic strain field distribution vector is input into a nonlinear dynamics reconstruction module, a high-dimensional phase space projection is constructed based on the geometric invariance principle of chaotic attractors, and a chaotic characteristic fingerprint of the collision event is extracted by jointly calculating the maximum Lyapunov exponent and the Kolmogorov entropy, wherein the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectory caused by environmental noise, including: According to the three-dimensional dynamic strain field distribution vector, the delayed coordinate embedding method is used to reconstruct the high-dimensional phase space, and the embedding dimension and time delay parameter are dynamically selected based on Takens theorem to generate the initial phase trajectory set. ‌The initial phase trajectory is input into a hyperbolic embedding algorithm, and phase trajectory segments satisfying the hyperbolic constraint are screened out by calculating the local curvature tensor, and pseudo phase trajectories caused by environmental noise are eliminated to obtain a core phase trajectory cluster after denoising; ‌Performing a chaos invariance analysis on the core phase trajectory cluster, jointly calculating the maximum Lyapunov exponent and the Kolmogorov entropy, and generating a chaos dynamics parameter matrix; The chaotic dynamics parameter matrix is input into the geometric topological mapping module, and the topological invariant features of the phase trajectory are extracted based on the persistent homology algorithm, and a chaotic feature fingerprint map including strange attractor dimension, fractal structure and recursive characteristics is output.

4. The method according to claim 3, wherein The chaotic characteristic fingerprint spectrum is subjected to collision intensity classification processing, and a collision energy propagation model is constructed using a spatiotemporal convolutional pulse neural network. The strain gradient distribution, energy attenuation rate and material damage threshold parameters are integrated through a dynamic weight allocation mechanism to output a quantitative evaluation matrix including collision grade, damage radius and residual strength prediction value, including: According to the chaos characteristic fingerprint map, the strange attractor dimension is mapped to the collision energy space through the topological data binning technology to construct the multi-level classification boundary conditions of the collision intensity; ‌The multi-level classification boundary conditions are input into the spatiotemporal convolutional spike neural network, the LIF neuron model is used to simulate the biological pulse triggering mechanism, the collision energy propagation path is dynamically modeled, and the probability density function containing the spatiotemporal energy distribution is output; ‌Performing physical constraint optimization on the probability density function, constructing a dynamic weight distribution function based on the material damage constitutive equation, integrating strain gradient distribution, energy decay rate and material fatigue threshold parameters, and generating a multi-objective coupled damage evolution model; The damage evolution model is input into a Monte Carlo-finite element hybrid solver, and through alternating iterations of random sampling and deterministic solution, a quantitative evaluation matrix including collision level, damage radius and residual strength is finally output.

5. The method according to claim 4, wherein The multimodal alarm protocol is triggered according to the quantitative evaluation matrix, communication spectrum resources are dynamically allocated based on the LoRaWAN protocol stack, and the alarm information is encoded into a time-varying frequency hopping pulse sequence using chaotic encryption modulation technology, and synchronously transmitted to the traffic management center and the OBU terminal of the adjacent vehicle, wherein the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve a sub-second alarm response, including: Trigger a priority-based multi-modal alarm protocol according to the collision level parameters in the quantization evaluation matrix, dynamically allocate communication spectrum resources through the LoRaWAN protocol stack, and generate a channel occupancy status map; Input the alarm information and the channel occupancy status map into the chaotic encryption modulation module, generate a time-varying key stream based on the Lorenz equation, encode the data into chaotic pulse sequences using differential frequency hopping technology, and output an anti-jamming encrypted signal; Construct a joint game model of collision events and channel interference, solve the optimal transmission strategy through Nash equilibrium, dynamically adjust the transmission power and frequency hopping interval, and optimize the end-to-end transmission delay to the sub-second level; Broadcast the anti-jamming encrypted signal through a multi-antenna MIMO array, synchronously transmit it to the traffic management center and adjacent vehicle OBU terminals, and achieve ultra-low bit error rate parsing at the receiving end based on the Turbo decoding algorithm.

6. The method according to claim 3, characterized in that, Perform chaotic invariance analysis on the core phase trajectory cluster, jointly calculate the maximum Lyapunov exponent and Kolmogorov entropy, and generate a chaotic dynamics parameter matrix, including: According to the geometric distribution characteristics of the core phase trajectory cluster, use the multi-fractal spectrum analysis method to calculate the Hölder exponent of each phase trajectory, and generate a fractal dimension weight matrix. Among them, the multi-fractal spectrum analysis method detects the singular point distribution on the edge of the phase trajectory through the wavelet transform modulus maximum chain; Input the fractal dimension weight matrix into the local linearization LLA algorithm, fit the tangent space of the phase trajectory based on the weighted least squares method, dynamically correct the noise sensitivity coefficient in the calculation of the Lyapunov exponent, and output a sequence of Lyapunov exponents with the maximum robustness; Perform information geometric manifold modeling on the core phase trajectory cluster, quantify the information flow density of the phase space by calculating the Ricci curvature tensor, and dynamically adjust the estimation window of the Kolmogorov entropy in combination with the Kontsevich entropy formula to generate an entropy value evolution spectrogram; Input the Lyapunov exponent sequence and the entropy value evolution spectrogram into the dynamic mode decomposition DMD module, solve the joint distribution function of chaotic dynamics parameters through the Tikhonov regularization method, and finally output a three-dimensional chaotic dynamics parameter matrix including fractal dimension, exponential sensitivity, and entropy production rate.

7. A guardrail collision warning system, characterized in that, The system includes: A processing module for non-linearly phase-aligning multi-source heterogeneous signals through an adaptive variational mode decomposition algorithm based on the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by a multi-modal sensor array deployed at key nodes of the guardrail, and generating a three-dimensional dynamic strain field distribution vector. Among them, the sensor array uses a piezoelectric-fiber composite sensing unit, and calculates the spatio-temporal correlation weight of each node signal based on mutual information entropy; An extraction module for inputting the three-dimensional dynamic strain field distribution vector into a non-linear dynamics reconstruction module, constructing a high-dimensional phase space projection based on the geometric invariance principle of the chaotic attractor, and extracting the chaotic feature fingerprint map of the collision event through the joint calculation of the maximum Lyapunov exponent and Kolmogorov entropy. Among them, the reconstruction module uses a hyperbolic embedding algorithm to eliminate the pseudo-phase trajectories caused by environmental noise; A building block for performing collision intensity grading on the chaotic feature fingerprint spectrum, constructing a collision energy propagation model using a spatio-temporal convolutional pulse neural network, and fusing strain gradient distribution, energy decay rate, and material damage threshold parameters through a dynamic weight allocation mechanism to output a quantitative evaluation matrix containing collision level, damage radius, and predicted residual strength values; An alarm module for triggering a multi-modal alarm protocol according to the quantitative evaluation matrix, dynamically allocating communication spectrum resources based on the LoRaWAN protocol stack, encoding alarm information into a time-varying frequency-hopping pulse sequence using chaotic encryption modulation technology, and synchronously transmitting it to the traffic management center and neighboring vehicle OBU terminals. Among them, the protocol optimizes the transmission delay by constructing a joint game model of collision events and channel interference to achieve sub-second alarm response.

8. The system according to claim 7, characterized in that, The processing module is specifically used for: According to the vibration acceleration, strain tensor, and acoustic emission signals collected in real time by the piezoelectric-fiber composite sensing unit, calculating the spatio-temporal correlation weights of the signals at each node through the mutual information entropy algorithm to generate a dynamic correlation matrix of multi-source signals. Among them, the mutual information entropy algorithm introduces time-delay embedding technology to eliminate the heterodyne interference between sensors; Input the dynamic correlation matrix into the adaptive variational mode decomposition module, dynamically optimize the mode number and bandwidth parameters based on the Kullback-Leibler divergence, and perform non-linear phase alignment processing on multi-source heterogeneous signals to output a set of decomposed modes with a unified time reference; Perform multi-scale covariance analysis on the set of decomposed modes, extract the principal component features of the strain field through the asymmetric tensor decomposition algorithm, and generate a three-dimensional feature tensor containing strain gradient, vibration energy, and acoustic emission event density; Input the three-dimensional feature tensor into the physical field fusion network, adopt a two-channel attention mechanism to perform cross-modal feature fusion on piezoelectric signals and fiber signals, and finally output a spatio-temporally continuous three-dimensional dynamic strain field distribution vector.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, where the computer program is set to execute the method according to any one of claims 1-6 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is set to run the computer program to execute the method according to any one of claims 1-6.

Images