Transportation risk optimization method and system

Abstract

The invention relates to the technical field of intelligent traffic and road disaster early warning, in particular to a transportation risk optimization method and system, and the method comprises the steps: S1, carrying out the multi-mode event-infrared-optical fiber data collection, and synchronously generating a unified index data package; s2, the physical information neural network combines energy conservation and causal inference assimilation to output a digital twin state tensor; s3, executing forward noise scheduling and reverse denoising by the conditional diffusion model, generating an icing probability field and extracting a high-frequency attention spectrum; s4, road structure factors are calculated according to the probability field and the attention spectrum, a binary unconstrained optimization model is constructed and mapped to an optical coherence Ising machine to be solved, an intervention scheduling scheme is formed through strategy network soft updating and safety verification, and an execution effect is written back and circularly updated. According to the method, sub-second risk identification and millisecond optimization decision are realized, and the accident rate and the salt spreading cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and road disaster early warning technology, and in particular to a transportation risk optimization method and system. Background Technology

[0002] Optimizing transportation risks directly impacts vehicle safety, logistics efficiency, and the lifespan of public infrastructure. In low-temperature, high-humidity, or complex traffic scenarios, road icing, water accumulation, and congestion significantly increase accident rates and maintenance costs. Traditional monitoring methods, relying heavily on fixed weather stations, road surface friction coefficient sensors, or single thermal models, struggle to cover fine-grained spatial data and fail to reflect the real-time effects of traffic interventions, leading to delayed warnings, excessive salt application, and decreased traffic efficiency.

[0003] Existing solutions typically employ: single-modal observation, relying solely on weather stations or infrared cameras to acquire temperature and humidity, lacking the ability to capture microscopic strain and brightness fluctuations in the road surface, thus failing to identify the initial formation of thin ice; pure data-driven models, using convolutional neural networks to directly predict risk, ignoring physical constraints such as energy conservation, which can easily lead to physically meaningless temperature fluctuations during extrapolation or under extreme weather conditions; and static threshold decision-making, triggering salt application based on fixed surface temperature or humidity thresholds, failing to consider traffic flow dynamics and local structural differences, resulting in resource waste and risk omissions. Due to their limited sensing dimensions, lack of physical constraints in the models, and rigid decision-making rules, these methods all have significant limitations in prediction accuracy, response time, and resource utilization. Summary of the Invention

[0004] To address the numerous problems existing in the prior art, this invention provides a transportation risk optimization method and system. This invention uses multimodal perception to model the road surface thermodynamic state, utilizes an embedded energy-conserving physical neural network and causal graph assimilation to generate a real-time digital twin, then constructs a risk probability field using conditional diffusion inference, and uses an optical coherent Ising machine to solve the binary optimization model and output intervention schemes, forming a second-level closed-loop optimization, which significantly improves the advance warning and salting efficiency.

[0005] A transportation risk optimization method includes the following steps:

[0006] S1. Collect multimodal observation data, and perform pulse coding, noise reduction, demodulation, and temporal-spatial synchronization on the data to generate a unified index data package;

[0007] S2. Input the unified index data package into the physical information neural network, adjust the network parameters according to the road surface energy conservation equation, obtain the road surface thermal state tensor, and construct the traffic meteorological intervention causal graph using the causal relationship modeling algorithm. Obtain the digital twin state tensor through four-dimensional variational assimilation.

[0008] S3. Using the digital twin state tensor as a conditional input to the diffusion generation model, risk assessment results are generated through forward diffusion and reverse denoising, and high-frequency attention spectrum is extracted.

[0009] S4. Based on the risk assessment results and the high-frequency attention spectrum, calculate the road structure factor, construct a binary unconstrained optimization model, map the binary unconstrained optimization model to the optical coherent Ising machine for solving, obtain the spin state results, use the policy network to adjust and safely verify the spin state results to form an intervention scheduling scheme set, execute the intervention scheduling scheme set and write the execution effect and new observation data back to the unified index data packet, and cyclically update steps S2-S4.

[0010] Preferably, the multimodal observation data consists of an event stream generated by an event camera, an infrared radiation temperature sequence generated by an infrared thermal imaging module, and a road surface temperature and strain sequence generated by a distributed fiber Bragg sensor.

[0011] Preferably, when generating a unified index data package, the event stream, infrared radiation temperature sequence, and road surface temperature sequence are aligned with the strain sequence using a sliding window of fixed time length, and index fields are constructed using a unified time identifier and geographic coordinates.

[0012] Preferably, the physical information neural network includes at least one layer of leakage integration triggering neurons and at least one fully connected layer using a nonlinear activation function, and uses the residual of the road energy conservation equation as a loss function term during training.

[0013] Preferably, the causal relationship modeling algorithm determines the causal structure through a constrained acyclic graph optimization method and updates the edge weight parameters through backpropagation using a differentiable ordinary differential equation solver.

[0014] Preferably, the four-dimensional variational assimilation minimizes both observation bias and background field bias within a preset time window, and employs a quasi-Newton iterative strategy to solve for the corrected digital twin state tensor.

[0015] Preferably, the diffusion generation model uses cosine noise scheduling in the forward diffusion stage, uses a graph convolutional network with residual connections in the reverse denoising stage, and extracts high-frequency components from the network attention matrix through fast Fourier transform to form a side-level high-frequency attention spectrum.

[0016] Preferably, the road structure factor is calculated by weighting the icing probability tensor, the total uncertainty tensor, and the edge-level high-frequency attention spectrum, and is used as a coupling matrix element to construct a binary unconstrained optimization model.

[0017] Preferably, the policy network adjusts the network weights using a soft update method after receiving the spin state results. The security verification is completed by jointly determining the symbolic constraint solution and the confidence of the deep model, and an intervention scheduling scheme set is formed after the verification is passed.

[0018] A transportation risk optimization system for implementing the aforementioned transportation risk optimization method, the system comprising:

[0019] The sensing module is used to collect multimodal observation data, perform pulse coding, noise reduction, demodulation and temporal-space synchronization on the collected data, and generate a unified index data package;

[0020] The assimilation module is used to input the unified index data package into the physical information neural network, adjust the parameters of the physical information neural network according to the road surface energy conservation equation, obtain the road surface thermal state tensor, and construct a traffic meteorological intervention causal graph using a causal relationship modeling algorithm, and obtain the digital twin state tensor through four-dimensional variational assimilation.

[0021] The inference module is used to take the digital twin state tensor as a conditional input to the diffusion generation model, generate risk assessment results through forward diffusion and reverse denoising, and extract high-frequency attention spectrum;

[0022] The optimization module is used to calculate the road structure factor based on the risk assessment results and the high-frequency attention spectrum, construct a binary unconstrained optimization model, map the binary unconstrained optimization model to the optical coherent Ising machine to solve the spin state results, and form an intervention scheduling scheme set through policy network adjustment and safety verification; the update module is used to execute the intervention scheduling scheme set, write the execution effect and new observation data into the unified index data package, and trigger the cyclic update of the assimilation module, the inference module and the optimization module.

[0023] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0024] By employing multimodal event-infrared-fiber fusion sensing technology, sub-second and sub-meter level road condition capture effects were achieved; by embedding energy conservation residual technology into physical information neural networks, stable temperature prediction effects were achieved in extrapolation scenarios; by combining causal graph differentiability inference with four-dimensional variational assimilation technology, real-time assessment of intervention effects and adaptive digital twin effects were achieved; by employing conditional diffusion generation and high-frequency attention spectrum extraction technology, high-resolution reconstruction of risk probability fields and sensitive edge identification effects were achieved; and by employing optical coherence Ising machine-assisted policy network soft update technology, millisecond-level scheduling optimization and online security verification effects were achieved. Attached Figure Description

[0025] Figure 1 This is a schematic flowchart of the method of the present invention;

[0026] Figure 2 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0027] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0028] like Figure 1 As shown, a transportation risk optimization method includes the following steps:

[0029] S1. Collect multimodal observation data, and perform pulse coding, noise reduction, demodulation, and temporal-spatial synchronization on the data to generate a unified index data package;

[0030] The event camera outputs a pulsed event stream on a nanosecond scale, with each event containing four pieces of information: x-axis, y-axis, timestamp, and polarity. The infrared thermal imaging module outputs a two-dimensional temperature grayscale matrix at a fixed frame rate. The distributed fiber Bragg sensor returns the spectrally demodulated road surface temperature and strain sequences at a rate of over a thousand times per second. The three data streams differ significantly in temporal resolution, spatial resolution, and data format. Directly feeding them into subsequent models in parallel would lead to temporal jitter, spatial misalignment, and inconsistent feature dimensions. This invention first achieves hard synchronization at the acquisition end: all sensors share a unified pulsed network timing signal, establishing the same starting timestamp reference at the microsecond level. Based on this, the invention employs a multi-level alignment algorithm based on a sliding time window at the edge end to achieve soft synchronization. Specifically, let the sliding window length be Δt, and all timestamps in the event stream fall within [t...]. i ,t i Events with a time interval of +Δt are archived into the same frame, and the median timestamp recorded in the infrared thermal imaging frame is t. ir The measurement time between the fiber optic measuring point and the closest point to the vehicle is denoted as t. fb If |t ir -t i |<Δt / 2 and|t fb -t i If |<Δt / 2, all three are considered to be at the same moment and written to the same index row. This judgment can be formalized as:

[0031]

[0032] In the formula, t iThe formula represents the sliding window start time, Δt represents the sliding window length, and |·| represents the absolute value operation. This formula ensures closed-set consistency within the time dimension. After this alignment, a unified index data package is generated. Each record contains an event flow density map, the corresponding infrared temperature matrix, road surface temperature and strain numerical vectors, latitude and longitude coordinates, and a unified timestamp. To reduce noise in the event flow, this invention employs a threshold integral pulse encoder. Let the membrane potential of the original event flow pulse sequence be V. m The integration time constant is τ. Pulses are counted within one encoding period τ. When the count value reaches a threshold V... thr A reset pulse is output periodically to clear the membrane potential. This process can be described as follows:

[0033]

[0034] Where n is the number of events within the encoding period. When the pulse density is below the threshold, the encoder automatically suppresses sporadic pulses to reduce false noise.

[0035] Noise reduction of the infrared temperature sequence employs a spatial median filter combined with a time-domain bidirectional exponential smoothing algorithm. The time smoothing coefficient α satisfies 0 < α < 1, and its output temperature... The calculation formula is:

[0036]

[0037] In the formula, T ir (k) represents the initial temperature. The result is shown in the filtering diagram. It has been verified that random fluctuations can be removed without losing the instantaneous temperature peak value when α = 0.3.

[0038] The strain signal after fiber Bragg demodulation typically contains baseline drift caused by source drift. This invention addresses this by fitting a first-order polynomial ∈ baseline (x) = ax + b, and the coefficients a and b are obtained using the least squares method. Then, ∈(x,t)-∈ baseline (x) yields the correction value, thus eliminating baseline drift.

[0039] In the spatial dimension, a mapping is established between the event stream sampling location and the infrared pixel location through linear interpolation, and the nearest neighbor principle is used to match the infrared pixel center point with the fiber optic measurement point. This mapping is stored in latitude and longitude form in a unified index data packet, allowing subsequent models to retrieve arbitrary modal data with fixed coordinates.

[0040] Example: The sensing system of this invention was deployed on a bridge section of a highway in a mountainous area. The inter-frame interval of the event camera was 50 microseconds, the infrared thermal imaging frame rate was 120 frames per second, and the fiber Bragg sampling rate was 2 kHz. The sliding window length Δt was set to 10 milliseconds, and the threshold V... thrSet to 150 events. Continuous on-site data collection for 5 hours generated a unified index data packet with a size of 3.2 gigabytes. Compared to traditional independent data collection schemes, this invention reduces missing alignment records within the same time window by 92%. Real-time alignment processing latency is less than 4 milliseconds, meeting the requirements of subsequent second-level inference links.

[0041] This invention achieves modality consistency during the acquisition phase, ensuring that subsequent neural networks and causal graphs are no longer burdened by time and space compensation. It eliminates feature drift caused by asynchronicity at its root, providing deterministic input for subsequent thermal state estimation and risk assessment. This design reduces downstream computational complexity while maintaining acquisition accuracy, improving the overall system's response rate and stability.

[0042] Preferably, the multimodal observation data consists of an event stream generated by an event camera, an infrared radiation temperature sequence generated by an infrared thermal imaging module, and a road surface temperature and strain sequence generated by a distributed fiber Bragg sensor.

[0043] The multimodal observation data is designed to simultaneously capture microscopic thermal changes, macroscopic radiation distribution, and structural strain behavior of the road surface, providing complementary and consistent input for subsequent digital twins. Event cameras record brightness change pulses with microsecond-level temporal resolution, enabling timely responses to rapid phenomena such as thin ice and water film flickering on the bridge surface. Infrared thermal imaging modules continuously output a radiation temperature matrix in a two-dimensional array, reflecting large-scale thermal distribution and aiding in the identification of hot and cold spots. Distributed fiber Bragg sensors are continuously laid along the road structure, calculating temperature and strain in real time through changes in fiber cavity length, achieving in-situ monitoring with sub-meter spatial resolution. All three types of sensors natively output digital signals, facilitating synchronous processing at the edge.

[0044] Each pulse in the event stream contains an x-coordinate, y-coordinate, timestamp, and polarity. To extract road surface activity intensity from the event stream, this invention defines the event density within a unit time window as:

[0045]

[0046] Where, N e D represents the number of events recorded within a time length Δt. e The dimension of D is "events per second". When thin ice cracks on the road surface or vehicles kick up water vapor, the local brightness changes rapidly, causing D to... e The value increases significantly in the corresponding pixel region and can be used as a risk precursor signal input to the thermal state estimation network.

[0047] The temperature matrix output by the infrared thermal imaging module is converted into a physical temperature field T after radiation correction. ir(x,y). This invention interpolates the temperature field into a grid consistent with the pixels of the event camera and reduces random noise through one-dimensional time filtering. The short-term fluctuations in the filtered temperature sequence are mainly caused by differences in road thermal inertia and disturbances from passing vehicles, providing important observational evidence for the radiation and convection terms in the energy conservation equation.

[0048] Distributed fiber optic Bragg sensors measure temperature and strain based on the Bragg wavelength shift principle. When the road surface undergoes freezing contraction or frost heave cracking, strain fluctuations will appear before the visible ice layer. This invention uses equidistant fiber optic cables sampling at 2,000 times per second to obtain the road surface temperature sequence T. fb The system uses (z,t) and the strain sequence ε(z,t), where z represents the position along the fiber direction. The system performs a Hilbert transform on the original spectral signal, extracts the phase, performs least-squares linear fitting to eliminate light source drift, and then uniformly transforms it to the camera coordinate system to achieve cross-modal space alignment.

[0049] The three data streams must maintain consistency in both time and space. This invention employs a sliding window synchronization mechanism: with a sliding window length of 10 milliseconds, the system first resamples the event stream into fixed frames based on the start time; then, it searches for the record closest to the infrared frame and the fiber optic measurement point time using the center time of the sliding window. Data within the same frame is considered as long as the time difference is less than half the sliding window length. Spatial alignment is achieved through a projection matrix and a nearest neighbor strategy: after mapping event pixels to physical coordinates, the corresponding infrared pixel center is found, and the nearest fiber optic measurement point is matched. These three elements are combined into a multimodal observation vector. The aligned observation vector is then written into a unified index data packet according to its timestamp and latitude / longitude.

[0050] Example: An event camera with a 60-degree field of view and an infrared module were deployed on a mountain highway bridge at an altitude of 1400 meters. The fiber optic sensor chain was 500 meters long with a spacing of 0.5 meters. The system ran continuously for 8 hours, generating 15 million unified index records. Compared to a single-mode solution using only the infrared module, this invention detected thin ice precursors four seconds earlier in icy fog conditions, reducing the false alarm rate to 3%. Experimental data showed that after the event density threshold was triggered, the rate of infrared temperature decrease increased significantly, and the fiber optic strain exhibited a negative jump. The three-mode synergy provided complete and observable evidence for the rapid icing process.

[0051] Through this design, the present invention can output structured and synchronous observation vectors at the perception layer, providing spatiotemporally consistent input for subsequent physical information neural networks. Event streams capture microsecond-level brightness jumps, compensating for insufficient infrared frame rates; the infrared temperature matrix provides a planar thermal field distribution, correcting for event noise; and fiber optic measuring points record internal temperature gradients and strain evolution, providing boundary conditions for the energy conservation equation. The complementary fusion of these three sources of information in a unified index data package improves the observability of thermal state estimation, thereby enhancing the reliability and lead time of risk assessment. This multimodal fusion strategy is the foundation for achieving self-driven digital twins and real-time risk optimization, ensuring that subsequent models can still obtain stable and accurate input data under extreme weather and complex traffic scenarios.

[0052] Preferably, when generating a unified index data package, the event stream, infrared radiation temperature sequence, and road surface temperature sequence are aligned with the strain sequence using a sliding window of fixed time length, and index fields are constructed using a unified time identifier and geographic coordinates.

[0053] Fixed-time sliding window is an alignment mechanism for real-time streaming data. Its core idea is to define a closed interval of length Δt on the time axis, mapping observation sources with different sampling frequencies and data structures to the same time anchor point, and then spatially merging them using the same coordinate system. The pulse sequence output by event cameras can achieve microsecond-level temporal accuracy, the frame period of infrared thermal imaging modules is mostly on the order of milliseconds, and the sampling interval of distributed fiber Bragg sensors is typically between sub-milliseconds and milliseconds. If the three data streams are directly superimposed, time jitter may lead to feature overlap errors; when making second-level predictions for freezing disasters, this error accumulates into state estimation drift. This invention provides a unified "time bucket" for different data sources through a sliding window mechanism, thereby ensuring the consistency of twin inputs.

[0054] The principle of sliding windows can be expressed by the following formula:

[0055]

[0056] Where T e Represents the set of timestamps for the event stream; t ir Indicates the center time of the infrared frame closest to the event time; t fb This represents the time of the fiber optic measurement point closest to the event time; Δt is the sliding window length. If the maximum time difference is less than half of the window, the three records are grouped into the same index row. This condition ensures that the alignment error does not exceed Δt / 2, thus meeting the stability interval requirements of the downstream dynamic model. In actual deployment, Δt is set according to road grade and traffic flow density; it can be 10 milliseconds for highways and can be relaxed to 20 milliseconds for national and provincial trunk roads.

[0057] Spatial alignment uses a unified latitude and longitude coordinate system. The center of each pixel in the event camera is calibrated using intrinsic and extrinsic parameters to obtain a 3D ray, and its intersection with the road surface elevation determines the geographic coordinates. Infrared thermal imaging pixels and event pixels share the same extrinsic parameter matrix. The absolute coordinates of fiber optic measurement points are determined during the laying phase. To address mapping errors, this invention employs nearest neighbor search: for any event pixel coordinate (φ... e ,λ e Search for the minimum Euclidean distance in the infrared grid and fiber optic point set. When the distance is less than the threshold ε, s Data from the same location is considered valid at that time. The threshold is typically set to 0.2 meters to eliminate inconsistencies caused by calibration deviations.

[0058] The unified index data package's field design balances fast retrieval with subsequent tensor requirements. Fields include: time stamp, latitude and longitude coordinates, event density value, infrared temperature matrix index, fiber optic temperature value, fiber optic strain value, sensor health marker, and data quality level. To support second-level truth recovery, the time stamp uses ISO-8601 format with microsecond-level precision, and latitude and longitude are represented using WGS-84 double-precision floating-point. Event density values ​​are recorded as single-byte unsigned integers, ranging from 0 to 255, corresponding to 0 to 5000 events per second, with linear scaling ensuring compression efficiency. The infrared matrix index is a list of row and column numbers used for location within the floating-point temperature block file; fiber optic temperature and strain values ​​are directly stored as floating-point numbers, and the data quality level is assessed by the missing rate within a sliding window, categorized into four levels from 0 to 3.

[0059] The effect of window alignment is particularly prominent during the rapid development phase of bridge deck icing. Field measurements show that when the temperature drops to -3 degrees Celsius and the relative humidity is above 85%, the event density at the wheel leading edge increases continuously for two window cycles, the infrared temperature decreases continuously by 0.6 degrees Celsius at the same coordinate point, and the fiber optic strain exhibits an inflection point from tension to compression. Without window alignment, the peak values ​​of these three indicators are misaligned by approximately 50 milliseconds, which the digital twin would incorrectly identify as short-term noise. With the alignment achieved in this invention, the three indicators are updated synchronously within the same row at the same time, allowing the digital twin network to promptly capture the release of latent heat of phase change and issue an icing risk warning 4.2 seconds in advance.

[0060] To further verify the effectiveness, the present invention and the traditional single-frequency synchronization scheme were simultaneously deployed on a 500-meter-long test road surface. The traditional scheme uses infrared frames as a reference to perform linear interpolation between the event stream and fiber optic data. During a two-hour test, 720 high-risk moments were recorded. The present invention missed 6 detections, while the traditional scheme missed 51. The present invention had 20 false alarms, while the traditional scheme had 48. Statistical analysis shows that the sliding window mechanism significantly improves detection sensitivity and reduces false alarms.

[0061] The sliding window length and threshold can adapt to traffic conditions. During peak traffic periods, the system shortens Δt based on real-time traffic density to improve its ability to capture high-speed temperature fluctuations; during low-traffic periods at night, the system automatically widens Δt to reduce data redundancy. Threshold ε s The spatial error is dynamically adjusted based on the calibration residuals of the optical fiber and the camera to keep it within a safe range.

[0062] In summary, by using a fixed-time sliding window combined with spatial nearest neighbor alignment, this invention solves the problems of temporal misalignment and spatial error caused by different sampling frequencies of multi-source sensors, ensuring that all observation data are seamlessly integrated under the same coordinate system and time axis. This provides spatiotemporally consistent input for subsequent physical information neural networks and causal graphs, significantly improving the observability of digital twins and the accuracy of risk prediction, and providing a stable data foundation for transportation risk optimization.

[0063] S2. Input the unified index data package into the physical information neural network, adjust the network parameters according to the road surface energy conservation equation, obtain the road surface thermal state tensor, and construct the traffic meteorological intervention causal graph using the causal relationship modeling algorithm. Obtain the digital twin state tensor through four-dimensional variational assimilation.

[0064] The physical information neural network is a model structure that directly embeds the conservation laws of road thermodynamics into the neural network training process. Its core idea is to introduce residuals based on the energy conservation equation as an additional loss term, beyond the error between the network's predicted and observed values. For a highway pavement, the energy balance can be written as:

[0065]

[0066] In the formula, ρ represents the density of the road surface material, c represents the specific heat capacity, and T represents the specific heat capacity. s Q represents the surface temperature of the road surface. SW Q represents the shortwave radiation flux. LW Q represents the longwave radiation flux. SEN Q represents the sensible heat flux. LAT Q represents latent heat flux. GRD L represents conduction flux. f Indicates the latent heat of ice melt. This represents the rate of mass change during the ice-water phase transition per unit time. The terms on the right-hand side of the equation are calculated using radiation measurements, meteorological elements, and fiber optic temperature gradients provided by the unified index data package, and can be directly used in residual calculations during training. The network employs a leaky integration trigger neuron layer to encode low-latency event flow features, then superimposes a nonlinear fully connected layer to map to the thermal state tensor space. The loss function consists of a weighted average of the mean square error term and the physical residual term. The physical residual weights are adaptively adjusted based on the energy closure error within the sliding window, ensuring the network maintains energy conservation constraints while satisfying observation fitting, thus avoiding physically meaningless temperature anomalies in extrapolation scenarios.

[0067] A causal modeling algorithm characterizes the structural dependence between intervention and icing response in scenarios involving the combined effects of traffic and meteorology. First, variables such as road surface thermal state tensor, traffic flow density, oncoming vehicle speed, amount of de-icing salt applied to the road surface, relative humidity, and wind speed are used as nodes in the causal graph. Then, an acyclic graph structure optimization algorithm is employed to continuously learn the adjacency matrix, with its objective function including a trace exponential regularization term to suppress loops. A differentiable ordinary differential equation solver feeds the increments of intervention nodes back to the road surface thermal state nodes, and edge weight gradients propagate backwards simultaneously with observation errors at the network level, achieving overall convergence of structure and parameters. The resulting traffic-meteorological intervention causal graph can generate counterfactual samples within the digital twin, such as the change in thermal state after "adding 5 grams of de-icing salt per square meter to nodes with an icing probability higher than 80%", providing a quantifiable intervention effect for subsequent optimization.

[0068] Four-dimensional variational assimilation uses a sliding time window to dynamically correct the network output field. The assimilation objective function is:

[0069] J = (ss b )B-1(ss b )+(H(s)-y)R -1 (H(s)-y)

[0070] Where s represents the state vector to be optimized, s b Let H represent the previous background field, y represent the observation operator, and B and R represent the background error covariance matrix and observation error covariance matrix, respectively. The observation operator maps the grid state to the observation point location through an interpolation kernel; the background error covariance is estimated statistically from historical residuals; and the observation error covariance is given by the sensor calibration value. The quasi-Newton iteration converges in about ten steps within each window, enabling second-level updates on a single graphics processing unit. The digital twin state tensor remains consistent with the observation after assimilation and possesses complete spatial continuity and energy conservation properties, providing a high-confidence prior for subsequent diffusion generation models.

[0071] Example: A half-hour data segment was selected on a bridge section. The first 30 seconds were used for pre-training of a physical information neural network, and the remaining time period was used for rolling application of four-dimensional variational assimilation. Compared with a conventional convolutional network without energy conservation constraints, the mean square error of road surface temperature decreased from 1 degree Celsius to 0.3 degrees Celsius; the detection of icing precursors was 180 seconds earlier, exceeding the baseline model's 80-second lead time. Furthermore, traffic density and salt application amount were used as exogenous intervention variables input into the causal modeling algorithm. An experimental vehicle fleet repeatedly drove on the same road section, intervening with different salt application doses, collecting a total of 60 sets of control data. The mean square error of the potential response inferred using the causal graph compared to the measured thermal state change was less than 0.1 degrees Celsius, proving that the causal graph can accurately identify the intervention effect.

[0072] By cascading a physical information neural network, a causal relationship modeling algorithm, and a four-dimensional variational assimilation module, this invention obtains a robust estimate of the road surface thermal state without relying on high-frequency external field forecasts and updates the digital twin in real time. Energy conservation constraints suppress network extrapolation instability; the causal structure characterizes the impact path of traffic behavior on the road surface heat budget; and variational assimilation integrates real-time observations with model priors, ultimately producing a spatiotemporally continuous and physically consistent digital twin state tensor, laying a reliable foundation for risk assessment and decision optimization.

[0073] Preferably, the physical information neural network includes at least one layer of leakage integration triggering neurons and at least one fully connected layer using a nonlinear activation function, and uses the residual of the road energy conservation equation as a loss function term during training.

[0074] The design goal of the Physical Information Neural Network (PIN) is to ensure that the data-driven model strictly adheres to the laws of thermodynamic conservation when predicting road surface temperature evolution, thereby maintaining physical interpretability and extrapolation stability under extreme weather conditions and sparse observations. The network consists of two layers: the first is a leakage integration triggering neuron layer, used to process pulse sequences transmitted by event cameras on a microsecond scale; the second is a fully connected layer using a nonlinear activation function, used to convert the neuron layer outputs into a continuous temperature field representation and learn nonlinear heat transfer relationships.

[0075] Leakage integration triggers the neuron layer to follow the classical circuit integral model. Let the membrane potential be V. m The time constant is τ m The input pulse current is I e The update rule for a discrete time step Δt can be written as:

[0076]

[0077] In the formula R m For film resistance. When V m Exceeding threshold V thr The neurons output pulses and reset the potential to the baseline value. Since the number of pulses is linearly related to the change in incident brightness, the leak-integrated triggering neuron layer can map the event density to the pulse frequency, thereby preserving temporal resolution without framing operations and providing sparse and efficient input for subsequent layers.

[0078] The fully connected layer employs a modified linear unit activation function. Compared to traditional convolutional layers, fully connected mapping is used here because the road surface thermal state tensor is typically coupled to multiple meteorological variables and road structure parameters through energy equations during physical calculations. These variables do not act locally in the tensor dimension but interact globally. The fully connected layer can more directly learn the combined effects of shortwave radiation, longwave radiation, sensible heat, and latent heat on temperature changes.

[0079] To incorporate thermodynamic constraints into the training process, this invention adds a road surface energy conservation residual term to the loss function. Assume the network predicts the surface temperature as follows: If the observed temperature is T, then the mean square error is denoted as:

[0080]

[0081] Where N is the batch size. The energy conservation residual is based on the following closed-form energy expression:

[0082]

[0083] Where ρ is the density of the road surface material, and c is the specific heat capacity. Q is calculated using the difference approximation. SW Q is the shortwave radiation flux. LW Q is the longwave radiation flux. SEN For sensible heat flux, Q LAT For latent heat flux, Q GRD For conduction flux, L f Latent heat from melting ice, The rate of change of phase transition mass per unit time. The residual loss term is defined as:

[0084]

[0085] The comprehensive loss function is written as:

[0086] L = L mse +λ phys L phys

[0087] Where λ phys The weights are adaptively adjusted during training based on the sliding window energy closure error. This design ensures that while the network approximates the observations, it is forced to learn an internal representation that minimizes the energy balance residual, thereby avoiding temperature jumps that violate physical laws during extrapolation.

[0088] Causal relationship modeling and physical networks work synergistically. The edge weights of the causal graph describe the incremental impact of traffic interventions and meteorological factors on terms of the energy equation. For example, salting changes the melting rate in the latent heat term; opening wind deflectors reduces sensible heat flux; and increased traffic density increases the heat transfer coefficient due to mechanical disturbances. By feeding the derivatives of the edge weights back to the network parameters, the network can quickly converge to a new thermal equilibrium state after each intervention, achieving an immediate response to the effects of the intervention.

[0089] Four-dimensional variational assimilation further integrates the network output with new observations. Within the assimilation window, temperature and strain observations and predicted fields are mapped to the same spatial resolution through observation operators, and the background error covariance matrix is ​​obtained from historical residual statistics. Using a quasi-Newton method, the digital twin state tensor can be converged to the observation-consistent optimum in 10 iterations. This assimilation result serves as the background field input for the next window, enabling continuous updates.

[0090] Example: An event camera, an infrared thermal imaging module, and a fiber optic sensor chain are deployed on a bridge surface, with sampling frequencies of 20000 Hz, 120 Hz, and 2000 Hz, respectively. The physical information neural network uses two layers of leakage integration trigger neurons, each with 128 neurons, followed by two fully connected layers, each with 256 neurons. Weight λ phys Initialized to 100. In a 2-hour winter experiment, the energy residual variance decreased from 30 to 2 before and after network training, and the temperature mean square error decreased from 1 to 0.3. Compared with the network without physical constraints, the temperature extrapolation error was reduced by 70% after 1 hour, indicating that the energy conservation residual term effectively suppressed drift.

[0091] By introducing a leak-integrated trigger neuron layer, event-level information directly drives thermal state changes, improving sensitivity to the initial thin ice formation stage. A fully connected layer, combined with energy-conserving residuals, avoids high-frequency spurious temperatures that are physically inconsistent. A causal graph provides intervention mapping, enabling the system to quickly adapt after salting or speed limiting. Four-dimensional variational assimilation ensures consistency between short-term predictions and real-time observations. Through these combined effects, the thermal state estimation of this invention remains stable even in scenarios involving simultaneous vehicle load changes, drastic weather changes, and traffic interventions, providing reliable input for transportation risk assessment and ensuring the accuracy and timeliness of risk optimization decisions.

[0092] Preferably, the causal relationship modeling algorithm determines the causal structure through a constrained acyclic graph optimization method and updates the edge weight parameters through backpropagation using a differentiable ordinary differential equation solver.

[0093] The core task of causal relationship modeling algorithms is to infer the directional dependence among three types of variables—traffic behavior, meteorological conditions, and road surface thermal state—in road environments lacking controllable experiments, thereby providing quantifiable intervention effects for the risk optimization stage. In this invention, the algorithm uses an acyclic graph structure as a priori and searches the adjacency matrix using a continuous optimization method, ensuring that the graph structure both fits the observed data and satisfies the acyclic constraint. Specifically, the total number of nodes is denoted as d, and the adjacency matrix is ​​denoted as... If we solve directly on the discrete graph space, we will face combinatorial explosion; therefore, we introduce a differentiable acyclic constraint function:

[0094] h(A) = tr(e) A⊙A )-d

[0095] Where tr(·) represents the trace operation, and ⊙ represents the Hadamard product. An acyclic graph requires h(A) = 0. To satisfy this condition, this invention constructs a continuous optimization objective:

[0096]

[0097] For the prediction error term, λ c λ is the acyclic constraint coefficient, and β is the sparse regularization coefficient. The prediction error term uses mean squared error. The network input includes traffic density, average vehicle speed, salt application dosage, relative humidity, wind speed, and the road surface thermal state tensor; the output is the thermal state increment for the next time step. During optimization, λ... c The adjacency matrix is ​​gradually increased using a logarithmic strategy, causing it to approximate an acyclic space at the end of convergence. After causal relationship inference is completed, it is necessary to evaluate the cumulative impact of the intervention variables over time. To this end, this invention employs a differentiable ordinary differential equation solver to convert the causal graph into a continuous dynamical system:

[0098]

[0099] x represents the node state vector, and θ represents the system parameters to be learned. Traffic intervention and meteorological drivers act on the system dynamics through the adjacency matrix A. The solver uses the polarized trapezoidal integral method to maintain numerical stability and calculates the gradient of the loss with respect to the terminal state using the adjoint variable method during the backpropagation phase. Then the accompanying variable a(t) satisfies:

[0100]

[0101] The gradient of the adjacency matrix is ​​given by:

[0102]

[0103] Given that the gradient is differentiable with respect to A, the causal graph can be updated via backpropagation in a single solution process, avoiding the non-smoothness problem caused by iterative switching of the discrete graph.

[0104] At the application level, traffic intervention nodes include salt application amount, lane speed limits, and traffic density; meteorological nodes include air temperature, relative humidity, and wind speed; and road surface thermal state nodes use surface temperature and latent heat flux at the corresponding coordinates in the thermal state tensor as observations. After causal inference is completed, counterfactual analysis can be performed. For example, in a road section with an 80% probability of icing, adjusting the salt application amount from 0 to 5 grams per square meter, and then re-integrating the dynamic system after the change in the edge weights of the corresponding causal graph, predicts that the road surface temperature will increase by 0.8 degrees Celsius after 20 minutes, thus quantifying the intervention effect.

[0105] Example: A 120-meter-long bridge deck was selected, and traffic detection radar and an automatic salting device were installed. During the testing phase, the traffic flow density was manually controlled to switch between peak and trough values, and salting was performed after 30 minutes. The adjacency matrix was updated online using a rolling window of 4 minutes and a step size of 30 seconds. The test lasted for 2 hours. The graph structure tended to stabilize in the first 20 minutes, and the error without acyclic constraints decreased to 1×10⁻⁶. -6 The comparative data show that the root mean square error of the difference between the model-predicted temperature rise after salting and the measured value is 0.12 degrees Celsius, which is 65% lower than the error of the traditional regression method without a differentiable solver. Moreover, it maintains the same accuracy during the peak traffic flow stage, indicating that the causal graph accurately captures the coupling effect of traffic density on sensible and latent heat fluxes.

[0106] This invention determines the causal structure through constrained acyclic graph optimization and updates edge weights via backpropagation using a differentiable ordinary differential equation solver, achieving interpretable path modeling from intervention variables to hot-state nodes. During network training, loops and redundant edges are explicitly penalized, resulting in a sparse and directionally unique causal graph, facilitating integration with human knowledge. The gradient propagation mechanism in continuous solution improves convergence speed and numerical stability. The counterfactual predictions output by the causal graph can be directly input into subsequent optimization modules, providing a quantified intervention benefit function for the optical coherence ising machine. This enables the system to integrate the causal effects of traffic and meteorology during the decision-making stage, thereby achieving higher predictability and resource utilization efficiency in transportation risk optimization.

[0107] Preferably, the four-dimensional variational assimilation minimizes both observation bias and background field bias within a preset time window, and employs a quasi-Newton iterative strategy to solve for the corrected digital twin state tensor.

[0108] Four-dimensional variational assimilation, based on control theory, maximizes the consistency between the model evolution trajectory and the observation sequence by synchronously optimizing the digital twin state vector within a finite time window. This method is particularly important for highway ice disaster scenarios because the road surface thermal state exhibits highly nonlinear and rapidly changing characteristics under the combined effects of forced driving and traffic disturbances, making short-term prediction reliability unreliable relying solely on feedforward neural networks. This invention embeds four-dimensional variational assimilation into the inference link, utilizing real-time observations provided by event density, infrared temperature, and fiber optic temperature strain to dynamically correct the physical information neural network output, forming a time-rolling digital twin state tensor that provides a high-confidence prior for subsequent diffusion-generative models.

[0109] Let the assimilation window length be τ. w The window starts at time t0 and ends at time t0+τ. w Within this interval, the state vector evolves with time, according to the discrete model s. k+1 =M k (s k Generate the prediction field, where M k This represents the update operator after inference via a physical information neural network. Observations within the window are retrieved from a unified index data package, denoted as y. k Observation operator H k The state vector is mapped to the observation space, and this mapping includes the spatial interpolation kernel and the sensor response function. The cost function takes the classical form:

[0110]

[0111] Where s b Let B be the background field at the start of the window, and R be the background error covariance matrix. k Let M be the covariance matrix of the observation error at the k-th step. 0→k This indicates the evolution from t0 to t k The multi-step model combination operator. The first term measures the deviation of the corrected state from the background field, and the second term measures the deviation of the model trajectory from the observation sequence; the two together determine the optimal solution. To improve the convergence speed, this invention employs the bounded quasi-second-order quasi-Newton method from the quasi-Newton algorithm to perform low-rank updates on the Hessian. Let the gradient be... When the number of iterations is m, the search direction is p. m Depend on Confirmed, B m This is the Hessian approximation matrix. The update process utilizes difference pairings. With step size s m =s m -s m-1A correction term is constructed to achieve a successive approximation of second-order information without explicitly calculating the Hessian. Since the dimension of the state vector is usually above 100,000, using a low-rank approximation can significantly reduce memory usage.

[0112] The operator design considers the resolution differences of the three source data. The event density is first integrated into a uniform grid of 0.01 meters at the pixel scale using a convolution kernel, and then interpolated to the fiber optic measurement point locations. Infrared temperature has built-in camera extrinsic parameters and can be directly mapped. Fiber optic temperature and strain are observed from discrete measurement points and expanded on a tensor grid using a weight matrix. The above processing ensures H... k Continuously differentiable, satisfying the gradient requirement of the quasi-Newton method. Observation error covariance R k The diagonal matrix is ​​used, with its diagonal elements derived from the sum of the sensor calibration variance and the sliding window statistical variance. The background error covariance B adopts a broadband separable structure, and the main diagonal elements are estimated from the residual variance of the past 24 hours. The diagonal bandwidth corresponds to a spatial correlation scale of 5 meters and a temporal correlation scale of 30 seconds.

[0113] To maintain energy conservation under significant changes in vehicle load, this invention preserves the energy residual constraints of the physical information neural network during the assimilation process. Specifically, the road surface energy conservation residual is recalculated after each iteration. If the residual exceeds the threshold η, then add a regularization term. The cost function is derived, where γ adaptively scales with the residual. This strategy ensures that the assimilation results satisfy both observational and thermodynamic equilibrium.

[0114] Eighteen hours of data were selected from highway sections experiencing continuous low temperatures and snowfall, with an assimilation window length of 6 minutes and a step size of 30 seconds. Compared to a baseline without assimilation, the mean square error of surface temperature prediction decreased from 0.8 degrees Celsius to 0.2 degrees Celsius; the icing detection lead time increased from 90 seconds to 350 seconds. Accident retrospective data showed that the sliding window alignment plus four-dimensional variational assimilation scheme could provide a risk warning of ice surface 100 meters in front of vehicles 5 minutes in advance, providing 2 minutes more buffer time than traditional computational fluid dynamics rapid prediction schemes.

[0115] Example: The aforementioned multimodal acquisition system was deployed on a two-way four-lane bridge. The event camera sampling frequency was 25000 Hz, the infrared frame rate was 120 Hz, and the fiber optic sampling frequency was 2000 Hz. Four-dimensional variational assimilation was performed on the graphics processing unit, with a single-window iteration of 10 steps taking 800 milliseconds. The quasi-Newton iteration convergence threshold was set to 1×10⁻⁶. -4 The threshold was reached in an average of 7 iterations. The road management system pushes the assimilated digital twin state tensor to the diffusion model, refreshing the risk assessment every minute. After 7 days of actual operation, the number of ice surface slippage accidents decreased by 60%, and the salt dosage decreased by 25%, proving that the assimilated state tensor provides sufficient and accurate prior information for the optimization module.

[0116] This invention enables a digital twin to absorb incremental information from multimodal observations in real time while maintaining physical consistency by simultaneously minimizing observation bias and background field bias in four-dimensional variational assimilation. A quasi-Newton iterative strategy significantly reduces computational load while ensuring convergence accuracy, achieving second-level updates. Dynamic threshold regularization further prevents energy closure violations. Overall, this assimilation scheme provides high-precision, timely, and physically reliable state inputs for optimizing transportation risks under extreme weather conditions.

[0117] S3. Using the digital twin state tensor as a conditional input to the diffusion generation model, risk assessment results are generated through forward diffusion and reverse denoising, and high-frequency attention spectrum is extracted.

[0118] The diffusion generation model undertakes two tasks in this invention: first, it performs probabilistic reasoning on the digital twin state tensor to generate risk assessment results with spatial continuity and uncertainty characterization; second, it extracts high-frequency attention spectra by analyzing the frequency domain characteristics of the attention mechanism within the model, providing structural factor weights for subsequent optimization stages.

[0119] The digital twin state tensor describes continuous variables such as surface temperature, latent heat flux, sensible heat flux, and potential intervention effects inferred from the causal graph on the road grid, providing physically consistent and high-resolution priors for the diffusion model. The diffusion generation model first performs forward diffusion, i.e., progressively injecting zero-mean Gaussian noise into the k-axis tensor at discrete time steps. Let the original tensor be x0, and the noise scheduling sequence be {β...} k}, then the closed-form sampling expression for the forward process is:

[0120]

[0121] α k =1-β k

[0122] Where ∈ k For independent and identically distributed Gaussian noise, α k A cosine function is used to ensure a smooth change in noise intensity. When the forward diffusion continues for K steps, the tensor approximates an isotropic Gaussian distribution, and spatial correlation is completely destroyed.

[0123] The inverse denoising process introduces the concept of conditional diffusion: the encoding of the digital twin's state tensor is used as a conditional vector, which is injected into the graph neural network through cross-attention, guiding the model to follow prior physical constraints during reconstruction. The inverse process is approximated by a piecewise Euler-Maruyama formula:

[0124]

[0125] For noise estimation of the conditional scoring network output, σ k denoted by , z represents standard Gaussian noise, and c is the conditional vector. The scoring network employs a four-layer graph convolutional residual structure, with node features encoded from digital twin tensors and edge features encoded from causal graph adjacency matrices. Experiments show that this structure can learn complex icing generation patterns while preserving physical consistency.

[0126] The risk assessment results consist of two parts: an icing probability tensor and a total uncertainty tensor. The icing probability is obtained by calculating the activation function on the final reconstructed samples; the uncertainty tensor is derived from the weighted fusion of sample variance and attention entropy, which can quantify the predictive reliability. To further explore the model's interpretability, this invention performs a Fast Fourier Transform on the attention weight matrix of each inverse layer, extracting the squared amplitudes higher than 0.6 times the Nyquist frequency as the high-frequency attention spectrum. The high-frequency attention spectrum reveals the model's sensitivity to small-scale road structure and microclimate differences, and serves as a direct input for the subsequent optical coherence ising mechanism to construct road structure factors.

[0127] In terms of implementation, the forward diffusion step count is set to 500, noise scheduling uses a cosine function, and backsampling uses an improved linear multi-step method to reduce computational load. The scoring network has 64 channels per layer, the residual blocks contain gated linear units, and the number of parameters is kept below 1.2 million, enabling 100-millisecond inference on a single consumer-grade graphics processing unit. High-frequency attention spectrum extraction uses batch fast Fourier transform, achieving a time of less than 10 milliseconds.

[0128] Example: This invention was tested on a highway bridge deck in a mountainous area at an altitude of 1400 meters. The resolution of the input digital twin state tensor was set to 0.5 meters, and the update frequency was 1 minute. In the actual test results, 45 rapid icing events were captured, and the model issued risk warnings on average 6 minutes in advance, which is 2 minutes faster than the baseline using only convolutional neural networks. The correlation coefficient between the icing probability curve and the actual observed ice thickness change reached 0.87, and the total uncertainty curve was positively correlated with the prediction error, indicating that the uncertainty estimation was reasonable. After performing road segment weighted averaging on the high-frequency attention spectrum, the correlation coefficient with the distribution of historical accident hotspots was 0.79, indicating that the high spectrum can characterize the location of potentially dangerous structures.

[0129] The diffusion-generative model, through forward randomization and inverse conditional reconstruction, obtains a probabilistic description of complex icing processes; combined with digital twin priors, it effectively avoids overconfidence in unobserved areas. High-frequency attention spectra provide data-driven weights for structural factors, enabling subsequent optimization to focus more on high-risk road sections. Overall, the diffusion-fractal coupled inference module of this invention significantly improves the time lead and spatial resolution of risk assessment, providing reliable probability fields and sensitivity information for transportation risk optimization decisions.

[0130] Preferably, the diffusion generation model uses cosine noise scheduling in the forward diffusion stage, uses a graph convolutional network with residual connections in the reverse denoising stage, and extracts high-frequency components from the network attention matrix through fast Fourier transform to form a side-level high-frequency attention spectrum.

[0131] The diffusion generation model undertakes the dual tasks of probability-based fine-grained risk reasoning and sensitive area identification in the transportation risk optimization process of this invention. The model consists of three main lines: forward diffusion, inverse denoising, and frequency domain interpretation, which respectively implement randomization, conditional reconstruction, and attention spectrum extraction. The following explanation covers its principles, implementation details, and effects.

[0132] The forward diffusion phase first injects time-varying Gaussian noise into the digital twin state tensor, causing it to gradually degenerate into an isotropic distribution. In the k-th step, the noise intensity is controlled by a cosine function:

[0133]

[0134] Where β max β is the maximum injection coefficient, K is the total number of steps, and β is the maximum injection coefficient. k Let α be the noise figure at step k. The cosine scheduling curve maintains a low amplitude in the initial stage to preserve key information, and accelerates the injection of noise in the later stage to improve the randomization. k =1-β k The cumulative attenuation coefficient is The state tensor x0 is obtained at step k:

[0135]

[0136] Where ∈ k The noise is Gaussian noise with zero mean and one variance. By using cosine modulation, the noise injection is smoother, avoiding damage to the local coherence structure in the early stages, while improving the diversity of later sampling.

[0137] The inverse denoising stage uses a graph convolutional network with residual connections as the scoring function to progressively denoise degraded samples and reconstruct the risk probability field. The road surface thermal state tensor is meshed to form node features, with adjacency relationships defined by the road topology and causal graph edge weights. The graph convolutional kernels are activated using gated linear units, and an identity skip connection structure is added after each layer to ensure smooth gradient flow and preserve multi-scale features. The conditional vector is obtained from the digital twin state tensor through two fully connected layers and injected into each graph convolutional layer via cross-attention. The inverse denoising iteration is implemented using an improved linear multi-step method, sharing noise estimates every two time steps to improve computational efficiency.

[0138] During the denoising process, an attention matrix is ​​generated within the scoring network to balance cross-node information flow. To reveal the model's sensitivity to local road structure and microclimate differences, this invention performs a Fast Fourier Transform on the attention matrix along the edge index direction, retaining the squared amplitudes with normalized frequencies higher than 0.6 as high-frequency spectra. The resulting spectral density is aggregated according to edge numbers to obtain the edge-level high-frequency attention spectrum. Since high-frequency components correspond to rapidly changing local features, such as abrupt slope changes, bridge joints, or windy areas, this spectrum can be regarded as structural factor weights, used to penalize road edges with high risk and tight logical connections in subsequent optical coherence ising machine optimization.

[0139] In terms of implementation details, the number of forward steps is set to 500, and β... max The value is set to 0.995 to quickly approach pure noise within the last 30 steps. The graph convolutional network has 4 layers, each with 64 channels, and the residual connections adhere to the "input plus output" strategy. The cosine scheduling sequence and network parameters are loaded offline; only matrix multiplication and element-wise operations are performed online. High-frequency attention spectrum extraction uses batch fast Fourier transform, with a runtime of less than 20 milliseconds.

[0140] To evaluate the effectiveness of the model in risk inference, an experiment was conducted on a 500-meter-long section of a high-altitude, cold-weather bridge. The digital twin state tensor spatial resolution was set to 0.5 meters, with an update cycle of 60 seconds. Data was collected for one week, recording 51 rapid icing events. The diffusion model of this invention identified the risk areas on average 5.5 minutes in advance, 2 minutes earlier than the baseline based on a conventional convolutional network. The high-risk edges marked by the high-frequency attention spectrum overlapped with historical accident hotspots by 78%, with a 22% increase in newly identified risks at unmarked bridge joints, significantly improving the accuracy of structural factors.

[0141] In terms of resource consumption, the total inference time for both forward and reverse directions is 200 milliseconds, with high-spectrum extraction taking 16 milliseconds, meeting the time constraint of refreshing the dynamic optimization link every minute. Compared to a pure diffusion generation model without using conditional vectors, the cross-entropy loss in the icing region is reduced by 0.18, indicating that the prior of the digital twin's state tensor effectively guides the sampling distribution. Further analysis reveals a significant positive correlation between high-spectral energy and prediction error during the reverse process, suggesting that spectral features can reflect the uncertainty of the model's attention at rapidly changing points.

[0142] Example: Deploying this invention in an urban expressway tunnel area, the sensible heat flux of the road surface fluctuates significantly at high frequencies due to the influence of the ventilation system. The experiment was conducted over 24 hours during rainy and foggy weather, with data streams generated through sliding window integration. The diffusion model detected 33 high-frequency hotspots, 29 of which corresponded to water accumulation and localized icing areas recorded by monitoring. The sensitivity provided by the high-spectral density was 35% higher than that of the non-spectral optimization scheme. The system was further optimized by adding variable speed limits and intelligent salting devices to the hotspot edges based on the structure factor output by the optical coherence isolator. The accident rate decreased by 57% in the following week, demonstrating the value of high-spectral density indicators in the optimization phase.

[0143] In summary, cosine noise scheduling provides progressive randomization to ensure sampling quality; graph convolutional networks with residual connections achieve stable denoising of high-dimensional states; and high-frequency attention spectrum mining of local sensitive edges provides a reliable quantitative basis for structural factor weighting. These three components, combined as the core module of this invention's risk reasoning, not only improve the accuracy of icing probability prediction but also inject interpretability into the decision optimization process, thereby achieving higher safety benefits and resource utilization efficiency in transportation risk optimization scenarios.

[0144] S4. Based on the risk assessment results and the high-frequency attention spectrum, calculate the road structure factor, construct a binary unconstrained optimization model, map the binary unconstrained optimization model to the optical coherent Ising machine for solving, obtain the spin state results, use the policy network to adjust and safely verify the spin state results to form an intervention scheduling scheme set, execute the intervention scheduling scheme set and write the execution effect and new observation data back to the unified index data packet, and cyclically update steps S2-S4.

[0145] The risk assessment results provide the icing probability tensor and total uncertainty tensor on the road grid, while the edge-level high-frequency attention spectrum reveals the model's sensitivity to local structural changes. This invention calculates the road structure factor by fusing these two factors, which is used to quantify the risk weight of a single road edge in global optimization. The specific calculation formula is as follows:

[0146] S(e)=ω1P ice (e)+ω2U tot (e)+ω3H ig (e),

[0147] Where P ice (e) represents the average probability of edge e freezing, U tot (e) represents the average uncertainty of this side, H high (e) represents the high-frequency attention spectrum energy, where ω1, ω2, and ω3 are normalized weights satisfying ω1 + ω2 + ω3 = 1. After calculation, the structure factors of all edges are linearly scaled to the interval [0, 1] for unified processing by the subsequent optimization solver.

[0148] The optimization problem consists of two types of binary variables: logistics vehicle route variable x e,k Indicates whether vehicle k passes through edge e; y is the intervention variable for de-icing vehicles. e This indicates whether to salt or scrape ice from edge e. The decision objective is to minimize the risk accumulation function and the operating cost function while satisfying traffic flow and operational resource constraints. This invention transforms the objective function into a quadratic expression solvable by an optical coherence Ising machine using a binary unconstrained optimization model:

[0149]

[0150] Where L(e) is the side length, C salt Cost per unit of salt, T k Let x be the travel time of vehicle k. e,* Let α, β, λ represent any vehicle passing through edge e, where α, β, λ are artificially adjustable weights. The formula does not contain constraint terms because constraints are implicitly satisfied through pre-defined mutual exclusion and capacity conditions between the binary variables, making it suitable for mapping to an optical coherent Ising machine. This machine uses a silicon photonic ring cavity as a classical spin node, achieves physical simulation of the coupling matrix through coherent optical coupling, and searches for low-energy spin configurations on a submicrosecond timescale. The mapping process transforms the binary variables into spins σ∈{-1,+1}. Let σ=2b-1, where b∈{0,1}, the binary unconstrained optimization expression can be directly converted into the Ising energy function. The optical coherent Ising machine uses a pump power ramp method to perform annealing on the coupling matrix J. ij With bias h i Simulated injection was performed. The number of nodes was set to 2048, which can accommodate variable embeddings for approximately 900 roadside elements and 50 vehicles. The annealing time was 2 microseconds, and the data was read 2000 times. The 20 lowest-energy spin states were selected as candidate schemes.

[0151] After the spin-state results are mapped back to binary variables, they are input into the policy network for secondary adjustments. The policy network is based on a transformer structure; its input includes spin variables, structural factors, and real-time vehicle positions. The output probability distribution is used to fine-tune variable values, making the scheme more closely reflect dynamic traffic conditions. Next, it enters the safety validator: first, a symbolic constraint solver verifies hard constraints such as traffic regulations, minimum vehicle distance, and construction safety distance; then, a deep confidence estimator calculates the confidence score of the policy network output. When both checks pass, the scheme is written into the intervention scheduling scheme set.

[0152] The intervention and dispatch plan is transmitted to logistics and de-icing vehicles via the vehicle-mounted 5G cellular network. After receiving the instructions, the vehicles perform route tracking and operation, and report progress and status in real time. The system calculates execution effectiveness indicators, including risk reduction rate, salt dosage, and arrival delay, and stores them in the execution log. Subsequently, the execution effectiveness and new observation data are written back to the unified index data packet every minute, triggering the physical information neural network to re-assimilate, the diffusion model to infer again, and the optical coherence ising machine to be optimized again, realizing the closed-loop rolling of steps S2 to S4.

[0153] Example: Two salt-spreading trucks and 20 cargo trucks were deployed on a 500-meter bridge section. The risk assessment refresh cycle was 60 seconds, optical coherent isolating machine annealing was 2 microseconds, policy network inference was 5 milliseconds, security verification was 1 millisecond, and end-to-end decision latency was 0.21 seconds. Compared to the heuristic shortest path plus fixed salt-spreading scheme, the icing accident rate decreased by 58%, and the salt dosage was reduced by 27%. The high-frequency attention spectrum weighted scheme prioritized salt spreading at bridgeheads, slopes, and windy areas, resulting in a calculated 31% improvement in resource utilization efficiency.

[0154] This optimization process maps the probabilistic risk field and structurally sensitive information into a physically solvable binary model, using optical hardware acceleration to obtain an approximate optimal solution, and ensuring feasibility and real-time performance through a policy network and security verification. A closed-loop write-back mechanism enables the system to continuously learn the intervention effects, progressively improving the structural factor weights and cost coefficients, achieving self-driven optimization of transportation risks.

[0155] Preferably, the road structure factor is calculated by weighting the icing probability tensor, the total uncertainty tensor, and the edge-level high-frequency attention spectrum, and is used as a coupling matrix element to construct a binary unconstrained optimization model.

[0156] The road structure factor integrates three pieces of information—the icing probability tensor, the total uncertainty tensor, and the edge-level high-frequency attention spectrum—into a single weight to guide the optical coherence ising machine in prioritizing high-risk edge segments during optimization. Its calculation employs a weighted linear superposition:

[0157] S(e)=ω1P ice (e)+ω2U tot (e)+ω3H high (e)

[0158] Where P ice (e) is the average probability of icing, U tot (e) is the average uncertainty, H high (e) is the high-frequency attention spectrum energy; the weights satisfy ω1+ω2+ω3=1. The three features are first normalized to zero mean and unit variance, and then the weights are assigned using the inverse information entropy method. The lower the entropy value and the higher the discrimination, the greater the weight.

[0159] When projecting the risk tensor onto the graph edges, polygon area weighting is used. The attention spectrum is mapped to 0 and 1 after preserving the squared amplitudes with normalized frequencies greater than 0.6 through a Fast Fourier Transform. The structure factor is recalculated every minute and... Exponential smoothing reduces mutations.

[0160] Injecting structural factors into the binary unconstrained optimization model: Energy function E = -∑ i＜j J ij b i b j -∑ i h i b i In the middle, b i For binary decision variables, the coupling coefficient is set. bias h i =S(i), and then linearly scaled to [-1, 1] to meet the hardware amplitude limit. In this way, the high-risk edge contributes more negative energy, guiding the Ising machine to spin-flip the corresponding variable.

[0161] The test section was 1000 meters long with 180 edges. The highest-risk edge had an icing probability of 0.92, an uncertainty of 0.18, and a high-spectral density of 0.64. Entropy equilibrium yielded ω1 = 0.46, ω2 = 0.24, ω3 = 0.30, a structure factor of 0.70, and an average of 0.32. After annealing for 2 microseconds using an optical coherent isolator, the edge's spin was reduced to 1, and the policy network only needed to fine-tune the salting order. Actual operation showed a 23% improvement in accident warning accuracy, a 19% reduction in salting amount, and that salting resources were concentrated in the high-risk section covering 15% of the area, further reducing the accident rate by 12%. The entropy equilibrium and sliding update mechanism of the structure factor ensured robust weight allocation, achieving a balance between safety, resources, and timeliness in the optimization results.

[0162] Preferably, the policy network adjusts the network weights using a soft update method after receiving the spin state results. The security verification is completed by jointly determining the symbolic constraint solution and the confidence of the deep model, and an intervention scheduling scheme set is formed after the verification is passed.

[0163] The policy network handles the final-hop mapping from the spin-state output of the optical coherent ising machine to the executable scheduling scheme. Its core function is to fine-tune the spin-state results while ensuring road safety constraints, making the scheme more closely match real-time traffic conditions and highly interpretable. The network structure employs a four-layer attention transformer. The input sequence is composed of three parts: first, a spin variable vector arranged in road edge order; second, road structure factors with the same index; and third, real-time location codes for vehicles and de-icing equipment. Through a multi-head attention mechanism, the network can capture high-order dependencies between variables across long topological distances, while explicitly focusing on high-risk edges using structure factors.

[0164] The output variables of an optical coherent ising machine belong to a discrete binary set. If hard-labeled inputs are directly applied to a neural network, gradient propagation will terminate. Therefore, this invention adds Gaussian noise perturbation to the network front-end and uses gated bilinear interpolation to map the discrete variables to the continuous domain, balancing differentiability and interpretability. Let the spin variable σ∈{-1,+1}, and the perturbed expression is:

[0165] v=σ+∈

[0166]

[0167] Where τ is the noise standard deviation, which can be adaptively scaled according to the spin state energy: the higher the energy, the larger τ, which promotes more exploration by the network; the lower the energy, the smaller τ, which keeps the solution close to the global optimum.

[0168] The network output is the probability distribution of the path variables and intervention variables for each edge, which is thresholded to obtain a binary decision. To avoid drastic parameter fluctuations in dynamic traffic, this invention employs a soft update strategy. Let the old weights be θ. old The new gradient updates the weights to θ. new Soft updates use exponential sliding:

[0169] θ←(1-η)θ old +ηθ new

[0170] Where η∈(0,1) is the update coefficient, which is dynamically adjusted during the online phase based on the decision returns of the previous window. The higher the returns, the smaller η becomes to stabilize high-quality strategies; the lower the returns, the larger η becomes to accelerate strategy migration. The return function is defined as the difference between the risk reduction rate and the operating cost, summarized after time discounting.

[0171] Safety verification comprises two parts: logical constraint checking and deep confidence assessment. Logical constraints, based on a symbolic constraint solver, transform traffic regulations, vehicle dynamics limitations, and equipment capacity into satisfiability propositions. Common constraint examples include: only one de-icing truck is allowed on the same road segment at the same time; the amount of salt spread must not exceed the loading limit; logistics vehicles and de-icing trucks must maintain a minimum distance of d. min The solver uses conflict-driven Boolean learning to achieve millisecond-level verification.

[0172] To avoid the network outputting logically feasible solutions with low confidence in weakly constrained scenarios, this invention introduces a deep confidence estimator. This estimator is a three-layer multilayer perceptron, taking the attention weights and output probabilities of the last layer of the policy network as input, and outputting a confidence value γ∈[0,1]. If γ is lower than a threshold γ... min The proposed solution will be discarded and reverted to the energy suboptimal spin solution for fine-tuning. This step will be repeated up to three times to ensure the real-time nature of the decision-making process.

[0173] After solving for symbolic constraints and jointly determining the deep confidence level, an intervention scheduling scheme set is formed. The schemes are issued according to a priority queue: when there are insufficient salt-spreading trucks and queuing is required, the product of the road structure factor and the icing probability is given priority. The logistics vehicle route adjustment adopts graph cost rewriting. If the original route has passed through high-risk edges, it is adjusted to the alternative route; otherwise, the original route is maintained to reduce delays.

[0174] Example: A 10km high-speed test section was equipped with 3 salt-spreading vehicles and 50 logistics vehicles, with a peak vehicle density of 500 vehicles per hour. The optical coherent ising machine output 1850 spin variables. After fine-tuning by the policy network, the number of path changes accounted for 18%, and the salt-spreading scheduling order was rearranged twice. The soft update coefficient was initially set to 0.05, automatically decaying to 0.02 after a significant increase in returns in the 4th round. The average time for logical constraint checks was 0.4 milliseconds, the confidence assessment time was 0.1 milliseconds, and the total policy generation time was controlled within 85 milliseconds. In 12 hours of actual operation, the system processed 720 rounds of decisions without issuing any instructions that violated traffic regulations.

[0175] Compared to the hard update strategy, the soft update strategy reduced the variation of network parameters by 63%, the fluctuation of scheduling schemes by 55%, and the average arrival delay of logistics vehicles by 8%. Incident retrospective analysis shows that during a sudden freezing rain event, the system generated three rounds of scheduling schemes within 60 seconds, ultimately reducing road closure time by 14 minutes compared to the static scheduling baseline.

[0176] In summary, the policy network achieves continuous fine-tuning of the spin-state results through noise injection and residual attention; soft updates ensure the stability of online learning; the symbolic constraint solver and deep confidence assessment jointly verify the legality and reliability of the decision; and the final output set of intervention scheduling schemes achieves a good balance between safety, efficiency and real-time performance, providing high-quality terminal decisions for transportation risk optimization.

[0177] like Figure 2 As shown, a transportation risk optimization system is used to implement the transportation risk optimization method described above. The system includes:

[0178] The sensing module is used to collect multimodal observation data, perform pulse coding, noise reduction, demodulation, and temporal-spatial synchronization on the collected data, and generate a unified index data packet. The sensing module consists of an event camera, an infrared thermal imager, and a distributed fiber Bragg demodulator. The front-end embedded FPGA completes pulse coding, digital filtering, and demodulation in real time. The on-chip synchronous clock and gigabit Ethernet interface encapsulate the processed data into a unified index data packet and send it to the edge computing node.

[0179] The assimilation module is used to input the unified index data package into the physical information neural network, adjust the parameters of the physical information neural network according to the road surface energy conservation equation, obtain the road surface thermal state tensor, and construct a traffic meteorological intervention causal graph using a causal relationship modeling algorithm. A digital twin state tensor is obtained through four-dimensional variational assimilation. The edge computing node of the assimilation module uses a 32GB GPU and a 16-core CPU to run the physical information neural network in collaboration. The energy conservation calculation is completed by the vectorized instruction set on the CPU, and the GPU accelerates training and inference. High-bandwidth DDR5 memory ensures that the road surface thermal state tensor is loaded into the four-dimensional variational assimilation solver in milliseconds.

[0180] The inference module is used to generate a diffusion model with the state tensor of the digital twin as a conditional input, generate risk assessment results through forward diffusion and reverse denoising, and extract high-frequency attention spectrum. The conditional diffusion model of the inference module and the graph convolutional denoising network are deployed on the same GPU and perform 16-bit mixed precision operations through the tensor core. The high-frequency attention spectrum extraction calls the onboard FFTIP core, with a typical single-batch inference latency of about 0.2s.

[0181] The optimization module is used to calculate the road structure factor based on the risk assessment results and the high-frequency attention spectrum, construct a binary unconstrained optimization model, map the binary unconstrained optimization model to the optical coherent ising machine to solve the spin state results, and form an intervention scheduling scheme set through policy network adjustment and security verification. The binary unconstrained optimization matrix of the optimization module is constructed on the CPU and then passed through the optical coherent ising machine via PCIe 4.0. The optical chip contains 2048 tunable ring cavities. The spin state output is completed within 2μs during the annealing process. The policy network runs on a second GPU, and the soft-updated weights are stored in NVMe SSD for the next round of calls. The security verification logic is executed in parallel by the SMT thread on the CPU.

[0182] The update module executes the intervention scheduling scheme set, writes the execution results and new observation data into the unified index data packet, and triggers the cyclical update of the assimilation module, the inference module, and the optimization module. The industrial-grade gateway of the update module aggregates vehicle CAN signals and 5G feedback, writes the execution results into the unified index data packet; the same gateway schedules gRPC calls to activate the next cycle of the assimilation, inference, and optimization modules, achieving a self-driven closed loop with a 60-second cycle.

[0183] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A transportation risk optimization method, characterized in that, Includes the following steps: S1. Collect multimodal observation data, and perform pulse coding, noise reduction, demodulation, and temporal-spatial synchronization on the data to generate a unified index data package; S2. Input the unified index data package into the physical information neural network, adjust the network parameters according to the road surface energy conservation equation, obtain the road surface thermal state tensor, and construct the traffic meteorological intervention causal graph using the causal relationship modeling algorithm. Obtain the digital twin state tensor through four-dimensional variational assimilation. S3. Using the digital twin state tensor as a conditional input to the diffusion generation model, risk assessment results are generated through forward diffusion and reverse denoising, and high-frequency attention spectrum is extracted. S4. Based on the risk assessment results and the high-frequency attention spectrum, calculate the road structure factor, construct a binary unconstrained optimization model, map the binary unconstrained optimization model to the optical coherent Ising machine for solving, obtain the spin state results, use the policy network to adjust and safely verify the spin state results to form an intervention scheduling scheme set, execute the intervention scheduling scheme set and write the execution effect and new observation data back to the unified index data packet, and cyclically update steps S2-S4.

2. The method according to claim 1, characterized in that, The multimodal observation data consists of an event stream generated by an event camera, an infrared radiation temperature sequence generated by an infrared thermal imaging module, and a road surface temperature and strain sequence generated by a distributed fiber Bragg sensor.

3. The method according to claim 1, characterized in that, When generating a unified index data package, the event stream, infrared radiation temperature sequence, and road surface temperature sequence are aligned with the strain sequence using a fixed time-length sliding window, and index fields are constructed using a unified time identifier and geographic coordinates.

4. The method according to claim 1, characterized in that, The physical information neural network includes at least one layer of leakage integration triggering neurons and at least one fully connected layer using a nonlinear activation function, and uses the residual of the road energy conservation equation as a loss function term during training.

5. The method according to claim 1, characterized in that, The causal relationship modeling algorithm determines the causal structure through a constrained acyclic graph optimization method and updates the edge weight parameters through backpropagation using a differentiable ordinary differential equation solver.

6. The method according to claim 1, characterized in that, Four-dimensional variational assimilation simultaneously minimizes observation bias and background field bias within a preset time window, and employs a quasi-Newton iterative strategy to solve for the corrected digital twin state tensor.

7. The method according to claim 1, characterized in that, The diffusion generation model uses cosine noise scheduling in the forward diffusion stage and a graph convolutional network with residual connections in the reverse denoising stage. It also extracts high-frequency components from the network attention matrix through fast Fourier transform to form a side-level high-frequency attention spectrum.

8. The method according to claim 1, characterized in that, The road structure factor is calculated by weighting the icing probability tensor, the total uncertainty tensor, and the edge-level high-frequency attention spectrum, and is used as a coupling matrix element to construct a binary unconstrained optimization model.

9. The method according to claim 1, characterized in that, After receiving the spin state results, the policy network adjusts the network weights using a soft update method. Security verification is completed by jointly determining the symbolic constraint solution and the confidence of the deep model, and an intervention scheduling scheme set is formed after the verification is passed.

10. A transportation risk optimization system, used to implement the transportation risk optimization method according to any one of claims 1-9, characterized in that, The system includes: The sensing module is used to collect multimodal observation data, perform pulse coding, noise reduction, demodulation and temporal-space synchronization on the collected data, and generate a unified index data package; The assimilation module is used to input the unified index data package into the physical information neural network, adjust the parameters of the physical information neural network according to the road surface energy conservation equation, obtain the road surface thermal state tensor, and construct a traffic meteorological intervention causal graph using a causal relationship modeling algorithm, and obtain the digital twin state tensor through four-dimensional variational assimilation. The inference module is used to take the digital twin state tensor as a conditional input to the diffusion generation model, generate risk assessment results through forward diffusion and reverse denoising, and extract high-frequency attention spectrum; The optimization module is used to calculate the road structure factor based on the risk assessment results and the high-frequency attention spectrum, construct a binary unconstrained optimization model, map the binary unconstrained optimization model to the optical coherent Ising machine to solve the spin state results, and form an intervention scheduling scheme set through policy network adjustment and safety verification. The update module is used to execute the intervention scheduling scheme set, write the execution effect and new observation data into the unified index data packet, and trigger the cyclic update of the assimilation module, the inference module and the optimization module.

Images