Road cascade congestion and secondary accident risk early warning method and related equipment

Abstract

This invention belongs to the field of intelligent transportation technology. It provides a method and related equipment for early warning of cascading congestion and secondary accidents on roads. The method includes: acquiring heterogeneous data and performing vectorized embedding, nonlinear mapping, and cross-modal comparative learning through a heterogeneous data processing pipeline to obtain a multimodal fusion input tensor; performing dual-stream lightweight feature extraction on the multimodal fusion input tensor to obtain spatial and temporal features; fusing the spatial and temporal features through an ST-Gating gated neural network and a cross-modal cross-attention mechanism based on cue learning to obtain the fused spatiotemporal features; training the model using a Hawkes process to obtain an early warning model for cascading congestion and secondary accidents on roads, and then performing subsequent prediction and early warning. Through the above scheme, this invention reduces the prediction overhead of cascading congestion and accident early warning and improves accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method and related equipment for early warning of cascading road congestion and secondary accident risks. Background Technology

[0002] With the widespread deployment of the Internet of Things (IoT), vehicle-to-everything (V2X), edge computing, and 5G in urban infrastructure, intelligent transportation systems have become a core infrastructure for modern urban governance, energy conservation, emission reduction, and sustainable development. Among these, congestion prediction and risk warning under non-stationary road network conditions are key prerequisite technologies for achieving safe management of the entire road network, improving overall road network capacity, and providing accurate dynamic travel navigation for the public. In real-world, open, and highly complex urban road networks, traffic data exhibits unprecedented massive volume, high dimensionality, and multi-source heterogeneity. Accurate risk contagion prediction not only relies heavily on historical traffic data (typically expressed as multi-dimensional time-series tensors) collected by various sensors deployed at the road's front end (such as geomagnetic sensors, loop coils, microwave radar, and AI surveillance cameras), but also requires deep integration with highly uncertain external non-traffic factors, such as weather conditions (heavy rain, fog, snowfall), emergencies (traffic accidents, vehicle breakdowns), road construction projects, and even holiday effects and large-scale urban gatherings. Early traffic flow prediction methods were mainly based on linear statistical models, such as the Autoregressive Integral Moving Average (ARIMA) model and Kalman filtering. These traditional methods strictly assume that the data is stationary and are completely unable to handle the nonlinear and highly volatile characteristics of real traffic flow.

[0003] Current mainstream traffic prediction models have revealed the following profound logical flaws and technical pain points in both their theoretical framework and engineering practice: (1) Extremely low modality utilization and misalignment of deep semantics: Most existing traffic prediction models use a single modality input, that is, only feeding the numerical traffic flow tensor collected by sensors into the network. This crude fusion method, which lacks a deep semantic alignment mechanism, causes the update directions of the model gradients to interfere with each other, resulting in extremely poor information fusion effect. The model is simply unable to explore the potential causal mapping relationship between different modalities in a unified feature manifold space. Business example: A sudden rainstorm (weather text classification data) will cause the traffic capacity of a certain low-lying road section (numerical tensor data) to drop sharply in the next few hours. However, due to the lack of semantic alignment, the model cannot establish an effective mathematical mapping between the "rainstorm vector" and the "risk contagion gradient" of the specific intersection.

[0004] (2) The computational disaster of serial architecture and attention mechanism: Currently, mainstream spatiotemporal prediction models generally adopt a "spatiotemporal serial" structure in network topology. In particular, the temporal memory features of long sequences will experience irreversible decay and over-smoothing after complex spatial graph Laplace smoothing operations. In addition, the widely introduced multi-head self-attention mechanism has... The computational complexity is excessive, making the model extremely bloated and completely unable to achieve lightweight deployment on edge nodes with limited computing power (such as intersection monitoring and RSU). Business example: During the five-day National Day Golden Week peak travel period, the key time period features of the earliest two days are easily covered or diluted by the short-term features of the following three days, resulting in a complete failure to capture long-term patterns.

[0005] (3) Limitations of Spatiotemporal Feature Interaction Failure and Static Network Topology in Extreme Disaster Scenarios: In real fluid dynamics traffic flow, spatial topology dependence and temporal evolution trends are often highly dynamically coupled. However, existing methods not only lack a dynamic adaptive gating mechanism to measure the relative importance of a node at a specific moment, but also heavily rely on a pre-set static adjacency matrix, completely failing to reflect the dynamic changes in road network connection topology caused by traffic accidents and temporary traffic control. Business Example: When a serious chain-reaction rear-end collision occurs at a major intersection in a city, the future traffic flow of that section will be in an extremely abnormal physical deadlock state. At this time, if the intersection is temporarily closed in both directions by the traffic police, not only will the local road network topology be completely severed physically and logically, but the traffic flow cycle pattern of the section over the past few days will also be basically invalidated. If the model still rigidly relies on the static pre-set matrix for spatial diffusion and uniformly assigns weights to spatiotemporal factors, it will be unable to actively block the invalidated historical time features, resulting in a blindly optimistic and erroneous prediction, and will be unable to effectively prevent the spread of cascading congestion. Summary of the Invention

[0006] Aimed at at least in addressing one of the technical problems existing in the prior art, this invention provides a method and related equipment for early warning of cascading road congestion and secondary accident risks, solving the problems of difficulty in fusion of multi-source heterogeneous data and feature loss under extreme disaster scenarios, improving the system resources required for early warning of cascading road congestion and secondary accident risks and increasing accuracy.

[0007] One aspect of the present invention provides a method for early warning of cascading road congestion and secondary accident risks, comprising: The historical traffic flow tensor, time metadata, and discrete external environmental factors of the target traffic network are obtained and vectorized and embedded through a heterogeneous data processing pipeline. Then, nonlinear mapping and cross-modal comparative learning are performed to obtain a multimodal fusion input tensor. We employ a dual-stream lightweight feature extraction method using structurally symmetric, functionally orthogonal, and parallel spatial and temporal streams on the multimodal fusion input tensor to obtain spatial and temporal features. Spatial and temporal features are used to calculate interaction gating coefficients through ST-Gating gated neural networks and a cross-modal cross-attention mechanism based on cue learning. During the calculation, hard fusion reconstruction is performed using a temperature scaling factor to obtain the fused spatiotemporal features, where the temperature scaling factor is used to characterize the severity level of external extreme events. The initial static distance adjacency matrix of the target traffic network is obtained, and the dynamic residual matrix is ​​constructed by Hawkes process decay. The fused spatiotemporal features are then subjected to weighted evolution of the final adjacency matrix using the initial static distance adjacency matrix and the dynamic residual matrix to obtain a road cascading congestion and secondary accident risk warning model. The target traffic network is predicted by using a road cascading congestion and secondary accident risk early warning model, and the early warning results of road cascading congestion and secondary accident risks are obtained.

[0008] According to the aforementioned method for early warning of cascading congestion and secondary accidents, the historical traffic flow tensor, temporal metadata, and discrete external environmental factors of the target traffic network are obtained. These are then vectorized and embedded through a heterogeneous data processing pipeline, followed by nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor, including: Obtain historical traffic flow data collected by multiple sensors in the target traffic network, and convert the historical traffic flow data into a tensor. A high-dimensional feature vector is obtained by non-linearly mapping a 1D convolutional layer that slides along the time dimension. ,in For the number of sensors, For historical time steps, The number of feature channels of the sensor. To pre-define a unified target hidden dimension; The road network time metadata is then encoded using the Time2Vec sinusoidal activation function mechanism for absolute periodicity, resulting in: ; in, Indicates the first Tensors of the dimension of embedded features For the input time scalar, For the dimensional boundaries of the embedded features, and These are the frequency and phase parameters that the model can learn during backpropagation training. Using discrete states among the discrete factors of the external environment as indices, the discrete factors of the external environment are mapped to dense real-number vectors of a specific length and queried to obtain embedding vectors. These embedding vectors are then subjected to dimensionality increase and nonlinear transformation using a fully connected feedforward network layer including the ReLU activation function to obtain the tensor of the discrete factors of the external environment. ; Self-supervised learning is used to analyze high-dimensional feature vectors that are in the same time window. and discrete factors of the external environment tensor As positive sample pairs, high-dimensional feature vectors from different time windows and discrete factors of the external environment tensor As negative sample pairs, cross-modal contrastive learning and semantic alignment are performed using the InfoNCE loss function to obtain the multimodal fusion input tensor. , where the InfoNCE loss function for: ; in, The cosine similarity between feature vectors For temperature hyperparameters, For time windows.

[0009] According to the aforementioned method for early warning of cascading congestion and secondary accidents, the multimodal fusion input tensor employs a dual-stream lightweight feature extraction method using structurally symmetric, functionally orthogonal, and parallel spatial and temporal flows to obtain spatial and temporal features, including: The spatial flow employs a risk topology network based on Chebyshev polynomial approximation to partition the multimodal fusion input tensor. Each independent graph signal segment is processed, where the risky topology network is approximated by Chebyshev polynomial truncation unrolling to obtain the spatial feature tensor. for: ; in, This is the scaled, normalized Laplace matrix. for Chebyshev polynomials For the frequency domain convolution kernel parameter tensor that the network can learn, For multimodal fusion input tensors, Chebyshev polynomial truncation order, For activation functions; The time flow uses a time-selective state space to partition the multimodal fusion input tensor into... Given a set of independent risk sequences, a dynamic transition matrix is ​​generated from the risk sequences using discretized parameters dependent on the input. Based on the dynamic transition matrix, a linear recursive operator that evolves dynamically over time is constructed as follows: ; ; in, As a time feature, It is a dynamic transition matrix. It is obtained by linear mapping of input tensors through multimodal fusion.

[0010] According to the aforementioned method for early warning of cascading road congestion and secondary accidents, spatial and temporal features are used to calculate interaction gating coefficients through an ST-Gating gated neural network and a cross-modal cross-attention mechanism based on cue learning. During calculation, a temperature scaling factor is used for hard fusion reconstruction to obtain the fused spatiotemporal features, including: Tensor of discrete factors of external environment As contextual soft hint vectors, construct the query matrix. , where the query matrix From spatial feature tensor Temporal feature tensor Node embedding and time embedding The result is obtained by combining and linearly mapping the matrix; using the query matrix. The query matrix is ​​used for cross-attention operations, and is expressed as a discrete factor tensor of the external environment. for Perform cross-attention calculation to obtain the gating coefficient tensor. for: ; in, For cross-attention operations, For learnable weight matrix, This is the temperature scaling factor. For bias terms, For activation functions; Based on the gating coefficient tensor, point-pair weighted fusion is performed on the parallel two-stream features to obtain the fused spatiotemporal features. for: ; in, The Hadamard product is a two-stream feature package containing spatial features. and time characteristics .

[0011] According to the aforementioned method for early warning of cascading road congestion and secondary accidents, the initial static distance adjacency matrix of the target traffic network is obtained. A dynamic residual matrix is ​​constructed using Hawkes process decay. The fused spatiotemporal features are then weighted and evolved into a final adjacency matrix using the initial static distance adjacency matrix and the dynamic residual matrix. Finally, the Huber Loss objective loss function is used for prediction optimization, resulting in a model for early warning of cascading road congestion and secondary accidents, including: Obtain the initial static distance adjacency matrix of the topological physical structure of the target transportation network. Construct two learnable node embedding representation matrices. ; During the modulation of the Hadamard product, a dynamic residual matrix is ​​constructed using the exponential time decay factor of the Hawkes process, where the exponential time decay factor... for: ; ;

[0012] in, For the risk dissipation rate parameter, To use the discrete factors tensor of the external environment Dynamic variables obtained through real-time mapping For the moment when an external emergency occurs, This is the current simulation moment; For dynamic residual matrix, This is the pooling average vector of the external features. This is the scaling factor; Represents the hyperbolic tangent function. Represents the linear rectification activation function; For the dynamic residual matrix and the initial static distance adjacency matrix Perform learnable weighted summation to generate a true dynamic adjacency matrix that participates in the convolution calculation of the current frame. for: ; in, These are weight parameters; The fused spatiotemporal features After performing residual connection and layer normalization processing, the predicted values ​​are obtained by predicting multiple consecutive time steps in the future through a two-layer nonlinear connection network. By repeatedly performing predictive learning, a road-cascade congestion and secondary accident risk warning model is obtained.

[0013] According to the aforementioned method for early warning of cascading road congestion and secondary accident risks, the method further includes: The Huber Loss objective loss function is used for prediction optimization, where the Huber Loss objective loss function is: ; in, For predicted values, The actual tags returned by the sensors. The threshold value is used.

[0014] Another aspect of the present invention provides a road cascading congestion and secondary accident risk warning device, comprising: The first module is used to obtain the historical traffic flow numerical tensor, time metadata, and discrete external environmental factors of the target traffic network, and to perform vectorized embedding through a heterogeneous data processing pipeline, and then perform nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor. The second module is used to perform dual-stream lightweight feature extraction on the multimodal fusion input tensor using structurally symmetric, functionally orthogonal, and parallel spatial and temporal streams to obtain spatial and temporal features. The third module is used to calculate the interaction gating coefficients of spatial and temporal features through the ST-Gating gated neural network and the cross-modal cross-attention mechanism based on cue learning, and to perform hard fusion reconstruction through the temperature scaling factor during the calculation to obtain the fused spatiotemporal features, where the temperature scaling factor is used to characterize the severity level of external extreme events; The fourth module is used to obtain the initial static distance adjacency matrix of the target traffic network, construct the dynamic residual matrix using Hawkes process decay, and perform a weighted evolution of the final adjacency matrix on the fused spatiotemporal features using the initial static distance adjacency matrix and the dynamic residual matrix to obtain the road cascading congestion and secondary accident risk warning model. The fifth module uses a road cascading congestion and secondary accident risk early warning model to predict the target traffic network and obtain early warning results for road cascading congestion and secondary accident risks.

[0015] Another aspect of the present invention provides an electronic device, including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described above.

[0016] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the methods described above.

[0017] The beneficial effects of this invention are as follows: Through cross-modal semantic alignment technology, it transforms unstructured information such as severe weather and special traffic control instructions into high-dimensional semantic latent vectors capable of direct matrix operations with the underlying physical tensor of traffic, greatly enriching the information sources for decision-making and fundamentally eliminating the early warning ceiling caused by the lack of information dimensions; the dual-stream parallel network structure completely isolates the extraction processes of temporal and spatial feature streams in the physical computation graph. Furthermore, it revolutionarily introduces Selective SSM to replace the traditional attention mechanism, thereby... The linear complexity breaks the quadratic computation barrier, reducing memory footprint and computing power requirements, thus clearing engineering obstacles for lightweight, low-power deployment of edge computing nodes (such as RSUs); by introducing a dynamic adjustment of the dissipation rate based on multimodal disaster semantics... Utilizing Hawkes dynamic graph technology and combined with an ST-Gating hard gating mechanism with temperature scaling, this model can instantly switch attention mechanisms like a human traffic police officer when facing non-stationary extreme scenarios such as sudden severe typhoon disasters and major traffic accidents that close roads. It can modify the mathematical connection topology of the underlying road network in real time, achieving millisecond-level blocking and exponential self-healing of risk shock waves. In complex real-world industrial scenarios, it has good industrial application prospects and practical deployment value. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the road cascading congestion and secondary accident risk warning process according to an embodiment of the present invention.

[0019] Figure 2 This is a diagram illustrating the overall system architecture and the logical framework for secondary accident risk warning data flow in an embodiment of the present invention.

[0020] Figure 3 This is a detailed analytical diagram of the multimodal data semantic alignment and latent space mapping pipeline module structure according to an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the fusion computing structure of the parallel dual-stream feature extraction network and the spatiotemporal interactive gating (ST-Gating) unit in an embodiment of the present invention.

[0022] Figure 5 This is a diagram of the adaptive dynamic adjacency matrix generation and risk contagion graph structure evolution and update mechanism based on Hawkes process driven by an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of a road cascading congestion and secondary accident risk warning device according to an embodiment of the present invention. Detailed Implementation

[0024] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] Please combine Figure 1 and Figure 2 ,in Figure 1 This is a schematic diagram of the road cascading congestion and secondary accident risk warning process according to an embodiment of the present invention. Figure 2 This is a diagram illustrating the overall system architecture and the logical framework for secondary accident risk early warning data flow in this invention. Figure 1 The road cascading congestion and secondary accident risk warning process includes, but is not limited to, steps S100~S500: S100 acquires the historical traffic flow numerical tensor, time metadata, and discrete external environmental factors of the target traffic network, and performs vectorized embedding through a heterogeneous data processing pipeline, followed by nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor.

[0026] In some embodiments, reference Figure 3 The diagram shown is a detailed analysis of the multimodal data semantic alignment and latent space mapping pipeline module structure. This embodiment of the invention acquires historical traffic flow values ​​collected by multiple sensors in the target traffic network. To capture extremely short-term local dynamic features and increase the dimensionality of the space, the tensor of the historical traffic flow values ​​is... A high-dimensional feature vector is obtained by non-linearly mapping a 1D convolutional layer that slides along the time dimension. ,in For the number of sensors, For historical time steps, The number of feature channels of the sensor. To pre-define a unified target hidden dimension, this method can effectively smooth out hardware measurement noise in the initial data acquisition process.

[0027] It is understandable that traffic flow exhibits strong regularity and periodicity (such as daily tidal cycles and weekly workday cycles). For timestamp metadata, this invention abandons traditional one-hot encoding and employs an innovative Time2Vec sinusoidal activation function mechanism for absolute periodic encoding, enabling the network to automatically capture and generalize periodic patterns. In this embodiment, the Time2Vec sinusoidal activation function mechanism is used to perform absolute periodic encoding on road network time metadata, resulting in: ; in, Indicates the first Tensors of the dimension of embedded features For the input time scalar, For the dimensional boundaries of the embedded features, and These are the frequency and phase parameters that the model can learn during backpropagation training.

[0028] The above method can perfectly capture the complex overlapping periodic patterns in traffic flow using a series of sine function terms with different frequencies and phases.

[0029] Using discrete states among the discrete factors of the external environment as indices, the discrete factors of the external environment are mapped to dense real-number vectors of a specific length and queried to obtain embedding vectors. These embedding vectors are then subjected to dimensionality increase and nonlinear transformation using a fully connected feedforward network layer including the ReLU activation function to obtain the tensor of the discrete factors of the external environment. This approach enables the creation of a unified and learnable embedding dictionary from highly discrete and unstructured external data, achieving adaptive embedding of external environmental features.

[0030] In some embodiments, self-supervised learning is used to combine high-dimensional feature vectors that are in the same time window. and discrete factors of the external environment tensor As positive sample pairs, high-dimensional feature vectors from different time windows and discrete factors of the external environment tensor As negative sample pairs, cross-modal contrastive learning and semantic alignment are performed using the InfoNCE loss function to obtain the multimodal fusion input tensor. , where the InfoNCE loss function for: ; in, The cosine similarity between feature vectors For temperature hyperparameters, For time windows.

[0031] After strong constraint projection through the contrastive learning mechanism in this embodiment of the invention, the features are depth-aligned in the latent space, thereby obtaining a unified multimodal input tensor. That is, when "extreme rainstorms" occur, the model not only obtains the text vector literally, but also generates a mathematical response from the underlying feature manifold that resonates with the "cascading congestion risk".

[0032] S200 employs a dual-stream lightweight feature extraction method, utilizing structurally symmetric, functionally orthogonal, and parallel spatial and temporal flows, to extract spatial and temporal features from the multimodal fusion input tensor.

[0033] In some embodiments, the spatial flow employs a risk topology network based on Chebyshev polynomial approximation to partition the multimodal fusion input tensor into... Each independent graph signal segment is processed, where the risky topology network is approximated by Chebyshev polynomial truncation unrolling to obtain the spatial feature tensor. for: ; in, This is the scaled, normalized Laplace matrix. for Chebyshev polynomials For the frequency domain convolution kernel parameter tensor that the network can learn, For multimodal fusion input tensors, Chebyshev polynomial truncation order, For activation functions; The time-flow method uses a time-selective state space to partition the multimodal fusion input tensor into... Given a set of independent risk sequences, a dynamic transition matrix is ​​generated from the risk sequences using discretized parameters dependent on the input. Based on the dynamic transition matrix, a linear recursive operator that evolves dynamically over time is constructed as follows: ; ; in, As a time feature, It is a dynamic transition matrix. It is obtained through a linear mapping of the multimodal fusion input tensor, i.e., a matrix. and It is the current input Dynamically generated through linear projection, this gives the model the ability to "selectively scan": when a traffic risk shockwave (such as sudden braking or abnormal traffic influx) is detected, The response is significantly amplified to write critical instantaneous information into the hidden state; while in the steady state, noise is selectively ignored. Most importantly, this process... The linear complexity enables high-fidelity memory extraction of ultra-long sequences, completely solving the computing power bottleneck at the edge in cascading congestion simulation.

[0034] S300 calculates the interaction gating coefficients of spatial and temporal features through the ST-Gating gated neural network and the cross-modal cross-attention mechanism based on cue learning. During the calculation, it performs hard fusion reconstruction through the temperature scaling factor to obtain the fused spatiotemporal features, where the temperature scaling factor is used to characterize the severity level of external extreme events.

[0035] In some embodiments, reference Figure 4 The diagram shown illustrates the fusion computational structure of the parallel dual-stream feature extraction network and the spatiotemporal interactive gating (ST-Gating) unit. This embodiment of the invention integrates the discrete factors of the external environment into the tensor... As contextual soft hint vectors, construct the query matrix. The query matrix From spatial feature tensor Temporal feature tensor Node embedding and time embedding The query matrix is ​​obtained by combining and linearly mapping the results. For cross-attention operations and in the form of the discrete factors of the external environment tensor for Perform cross-attention calculation to obtain the gating coefficient tensor. The calculation is as follows: ; in, Let be a spatial feature tensor; be a temporal feature tensor. Embedded representation for road network nodes; For time-embedded vectors; Query is composed of The query matrix obtained by combining and linearly mapping; For discrete factors of the external environment tensor. For cross-attention operations, The weight matrix is ​​a learnable weight matrix; This is the temperature scaling factor. For bias terms; For activation function, For the gating coefficient tensor; Based on the gating coefficient tensor, point-pair weighted fusion is performed on the parallel two-stream features to obtain the fused spatiotemporal features. for: ; in, The Hadamard product is a two-stream feature package containing spatial features. and time characteristics .

[0036] Understandably, during periods of stable traffic, Maintaining default values, the model adopts spatiotemporal features in a balanced manner. Once the system detects a severe cascading accident at a major road intersection leading to a physical deadlock... A forced, sharp downward adjustment. This change causes the Sigmoid function to exhibit a polarized "binary hard selection" effect, making... The value spikes instantly, approaching 1. At this point, the model intelligently and proactively "forgets" and discards historical time-cycle patterns that are now invalid. It relies heavily on the propagation of information from the current dynamic spatial topology to provide rapid early warnings with extremely low latency for impending cascading congestion and secondary accidents in the surrounding area.

[0037] S400: Obtain the initial static distance adjacency matrix of the target traffic network, construct the dynamic residual matrix using Hawkes process decay, and perform weighted evolution of the final adjacency matrix on the fused spatiotemporal features using the initial static distance adjacency matrix and the dynamic residual matrix to obtain the road cascading congestion and secondary accident risk warning model.

[0038] In some embodiments, reference Figure 5 The invention presents a Hawkes process-driven adaptive dynamic adjacency matrix generation and risk contagion graph structure evolution update mechanism. To accurately reflect the objective law of the decay of the influence of sudden events over time in the physical world, this embodiment obtains the initial static distance adjacency matrix of the target traffic network's topological physical structure. Construct two learnable node embedding representation matrices. ; During the modulation of the Hadamard product, a dynamic residual matrix is ​​constructed using the exponential time decay factor of the Hawkes process, where the exponential time decay factor... for: ; ; in, For the risk dissipation rate parameter, To use the discrete factors tensor of the external environment Dynamic variables obtained through real-time mapping The time of occurrence of external emergencies (such as heavy rain or accidents). This represents the current simulation moment; different catastrophic events correspond to different physical self-healing rates, thereby calculating a residual connected matrix possessing "temporal memory and self-healing capabilities". , For dynamic residual matrix, This is the pooling average vector of the external features. This is the scaling factor; Represents the hyperbolic tangent function. This represents the linear rectification activation function.

[0039] It is understood that, according to the above embodiments, the embodiments of the present invention have an environment semantic-driven and exponentially self-healing graph topology evolution mechanism, that is, when the system resolves the "major traffic accident" signal, At its absolute peak, the feature dimensions of the affected nodes are instantly and drastically scaled or even zeroed out. The graph topology completes "physical blocking" within milliseconds, effectively preventing the spread of erroneous spatial prediction information. However, as the incident is cleared and time passes, the decay factor... The weights of the blocked road network topology gradually "heal" and smoothly return to the normal static connectivity state as they decrease exponentially.

[0040] For the dynamic residual matrix and the initial static distance adjacency matrix Perform learnable weighted summation to generate a true dynamic adjacency matrix that participates in the convolution calculation of the current frame. for: ; in, These are weight parameters; The fused spatiotemporal features After performing residual connection and layer normalization processing, the predicted values ​​are obtained by predicting multiple consecutive time steps in the future through a two-layer nonlinear connection network. By repeatedly performing predictive learning, a road-cascade congestion and secondary accident risk warning model is obtained.

[0041] In some embodiments, the present invention addresses the industry pain point that real-world traffic flow sensing data inevitably contains a large number of measurement outliers, and specifically adopts Huber Loss as the core objective loss function: ; in The actual tags returned by the sensors. For predicted values, The threshold is used. This mechanism acts like a "pressure limiter" for model parameter updates, effectively isolating the malicious damage of extreme erroneous and dirty data to the overall network weights. Finally, the overall network parameters are iteratively updated end-to-end by an industrial-grade optimizer.

[0042] S500 uses a road congestion and secondary accident risk warning model to predict the target traffic network and obtain the road congestion and secondary accident risk warning results.

[0043] It is understood that the road-cascade congestion and secondary accident risk warning model obtained through the above training in the embodiments of the present invention can predict the target traffic network to obtain the road-cascade congestion and secondary accident risk warning results.

[0044] In some embodiments, in a cascading congestion risk warning scenario, the input first continuously collects 8-dimensional features such as vehicle speed, density, and occupancy at 50 key intersections of the target road network using radar equipment deployed on the roadside (e.g., with a sampling frequency of 10Hz), forming a historical traffic flow tensor with a time window of 2.4 seconds (24 frames × 0.1 seconds). Simultaneously, it connects to open APIs such as those from meteorological bureaus to access discrete labels for meteorological environments such as "extreme rainstorms," ​​achieving precise alignment in timestamps. During the processing phase, the system utilizes the InfoNCE cross-modal contrastive learning mechanism to map rainstorm labels into 128-dimensional dense vectors, achieving semantic space alignment with the traffic flow tensor. Simultaneously, it captures the traffic attenuation caused by water accumulation through a Hawkes process, dynamically adjusting the adjacency weights of affected intersections in real time. Finally, at the output, the system generates a heatmap of the entire road network risk evolution 45 minutes in advance, accurately predicting the trend of local congestion spreading "cascading" to upstream intersections. After the prediction results are generated, they guide traffic management departments to implement green wave traffic control in advance. Simultaneously, they are pushed to roadside units (RSUs) via a dedicated network, broadcast to onboard units (OBUs) via C-V2X protocol, and simultaneously displayed as a warning to third-party navigation apps, such as "High risk of congestion 3 kilometers ahead." Onboard terminals, three levels of alerts are triggered: visual (red indicator on dashboard), auditory (buzzer), and tactile (steering wheel vibration) to complete a closed-loop feedback for drivers. Experimental data based on the NGSIM I-80 dataset demonstrates that this solution has significant low-consumption advantages in terms of computing power. In a pure CPU environment, the single inference latency for a 50-node road network is only 128-143 milliseconds, and the mean absolute error (MAE) of location prediction is as low as 5.93 (risk value dimension 0-100). Furthermore, extreme weather conditions only increase computation time by 4.7% under different event severity levels (severity 0.0 to 1.0), effectively curbing congestion.

[0045] In high-risk secondary accident early warning scenarios, this invention relies on multi-source fusion to acquire data at the input end: lane-level traffic flow mutation signals are collected at 0.1-second intervals using geomagnetic coils (e.g., a speed drop >30% triggers an anomaly marker), combined with video surveillance analysis to identify discrete labels for sudden accidents such as "multi-vehicle rear-end collisions," and integrated with real-time accident locations and timestamps reported by the traffic police alarm system to achieve spatiotemporal alignment of the three heterogeneous data sources. During the processing phase, a severe accident label instantly triggers a polarization adjustment of the external temperature scaling factor, activating a spatiotemporally interactive gated neural network (ST-Gating). This hard gating mechanism actively blocks invalid historical time-cycle features, causing the model to fully shift to accident-controlled spatial topological features, and completing an emergency physical blockage of the underlying road network mathematical connections within milliseconds. Finally, at the output end, the system comprehensively assesses the cascading collision risk of vehicles approaching from behind and implements tiered feedback prevention and control measures: it issues an "emergency braking suggestion" command to roadside units (RSUs) and vehicle terminals within 500 meters behind the accident site and triggers a high-risk alarm combining sound, light, and vibration; it pushes a "suggested lane change and avoidance" message to the navigation terminals of drivers within 500 meters to 2 kilometers; and it sends detour planning routes to terminals within 2 to 5 kilometers. Cross-time period tests based on the NGSIM real trajectory dataset verified the accuracy of its cascading collision risk estimation. The model's location prediction accuracy covers 50 highway segment nodes, and the MAE is stable at 5.7-6.1 (with cross-time period fluctuations <0.5). In terms of computing power, the scheme maintains computational resource consumption at a level that can be run on a single-core CPU, and the time efficiency achieves a single inference time of less than 0.15 seconds (frame rate stable at 41-42 FPS). The end-to-end latency from accident detection to the delivery of warnings at all levels to the vehicle owner's client is controlled within 200 milliseconds (including 128-143 milliseconds of inference plus less than 60 milliseconds of network transmission), achieving extremely low latency and high precision prevention and control of secondary accidents without GPU acceleration.

[0046] Figure 6 This is a schematic diagram of a road cascaded congestion and secondary accident risk early warning device according to an embodiment of the present invention. The device includes a first module 610, a second module 620, a third module 630, a fourth module 640, and a fifth module 650.

[0047] The system comprises three modules: The first module acquires historical traffic flow tensors, temporal metadata, and discrete external environmental factors of the target traffic network. It then performs vectorized embedding through a heterogeneous data processing pipeline, followed by nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor. The second module performs lightweight dual-stream feature extraction on the multimodal fusion input tensor using structurally symmetric, functionally orthogonal, and parallel spatial and temporal flows to obtain spatial and temporal features. The third module calculates interaction gating coefficients for the spatial and temporal features using an ST-Gating gated neural network and a cue-based cross-modal attention mechanism. The system performs hard fusion reconstruction using a temperature scaling factor during computation to obtain the fused spatiotemporal features, where the temperature scaling factor characterizes the severity level of external extreme events. The fourth module obtains the initial static distance adjacency matrix of the target traffic network, constructs a dynamic residual matrix using Hawkes process decay, and performs a weighted evolution of the final adjacency matrix on the fused spatiotemporal features using the initial static distance adjacency matrix and the dynamic residual matrix to obtain a road cascading congestion and secondary accident risk warning model. The fifth module predicts the target traffic network using the road cascading congestion and secondary accident risk warning model to obtain the road cascading congestion and secondary accident risk warning results.

[0048] For example, with the cooperation of the first, second, third, fourth, fifth, and sixth modules in the device, the embodiment device can implement any of the aforementioned methods for early warning of cascaded road congestion and secondary accident risks. This involves acquiring the historical traffic flow tensor, temporal metadata, and discrete external environmental factors of the target traffic network, and embedding them into vectors through a heterogeneous data processing pipeline. This is followed by nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor. The multimodal fusion input tensor is then subjected to dual-stream lightweight feature extraction using structurally symmetric, functionally orthogonal, and parallel spatial and temporal flows to obtain spatial and temporal features. These spatial and temporal features are then processed through ST-Gating. A gated neural network and a cross-modal cross-attention mechanism based on cue learning are used to calculate the interaction gating coefficients. During computation, a temperature scaling factor is used for hard fusion reconstruction to obtain the fused spatiotemporal features, where the temperature scaling factor characterizes the severity level of external extreme events. An initial static distance adjacency matrix of the target traffic network is obtained, and a dynamic residual matrix is ​​constructed using Hawkes process decay. The fused spatiotemporal features are then weighted and evolved using the initial static distance adjacency matrix and the dynamic residual matrix to obtain a road cascading congestion and secondary accident risk warning model. This model is used to predict the target traffic network, yielding warning results for road cascading congestion and secondary accident risks. The beneficial effects of this invention are: through cross-modal semantic alignment technology, it transforms unstructured information such as severe weather and special traffic control instructions into high-dimensional semantic latent vectors that can directly perform matrix operations with the underlying physical tensor of traffic, greatly enriching the information sources for decision-making and fundamentally eliminating the warning ceiling caused by the lack of information dimensions; the dual-stream parallel network structure completely isolates the extraction processes of temporal and spatial feature streams on the physical computation graph. And it revolutionarily introduces Selective SSM to replace the traditional attention mechanism, in order to The linear complexity breaks the quadratic computation barrier, reducing memory footprint and computing power requirements, thus clearing engineering obstacles for lightweight, low-power deployment of edge computing nodes (such as RSUs); by introducing a dynamic adjustment of the dissipation rate based on multimodal disaster semantics... Utilizing Hawkes dynamic graph technology and combined with an ST-Gating hard gating mechanism with temperature scaling, this model can instantly switch attention mechanisms like a human traffic police officer when facing non-stationary extreme scenarios such as sudden severe typhoon disasters and major traffic accidents that close roads. It can modify the mathematical connection topology of the underlying road network in real time, achieving millisecond-level blocking and exponential self-healing of risk shock waves. In complex real-world industrial scenarios, it has good industrial application prospects and practical deployment value.

[0049] This invention also provides an electronic device, which includes a processor and a memory; The memory stores the program; The processor executes a program to perform the aforementioned method for warning of cascading road congestion and secondary accidents; the electronic device has the function of carrying and running the software system for warning of cascading road congestion and secondary accidents provided in the embodiments of the present invention, such as a personal computer, minicomputer, mainframe, workstation, network or distributed computing environment, standalone or integrated computer platform, or communicating with charged particle tools or other imaging devices, etc.

[0050] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the road cascading congestion and secondary accident risk warning method described above.

[0051] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in the embodiments of this invention. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0052] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method for cascading road congestion and secondary accident risk warning.

[0053] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0054] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0055] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0056] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0057] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0058] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0059] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0060] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A road cascade congestion and secondary accident risk early warning method, characterized in that, include: The historical traffic flow tensor, time metadata, and discrete external environmental factors of the target traffic network are obtained and vectorized and embedded through a heterogeneous data processing pipeline. Then, nonlinear mapping and cross-modal comparative learning are performed to obtain a multimodal fusion input tensor. We employ a dual-stream lightweight feature extraction method using structurally symmetric, functionally orthogonal, and parallel spatial and temporal streams on the multimodal fusion input tensor to obtain spatial and temporal features. Spatial and temporal features are used to calculate interaction gating coefficients through ST-Gating gated neural networks and a cross-modal cross-attention mechanism based on cue learning. During the calculation, hard fusion reconstruction is performed using a temperature scaling factor to obtain the fused spatiotemporal features, where the temperature scaling factor is used to characterize the severity level of external extreme events. The initial static distance adjacency matrix of the target traffic network is obtained, and the dynamic residual matrix is ​​constructed by Hawkes process decay. The fused spatiotemporal features are then subjected to weighted evolution of the final adjacency matrix using the initial static distance adjacency matrix and the dynamic residual matrix to obtain a road cascading congestion and secondary accident risk warning model. The target traffic network is predicted by a road cascade congestion and secondary accident risk early warning model, and the results of road cascade congestion and secondary accident risk early warning are obtained. The multimodal fusion input tensor is subjected to dual-stream lightweight feature extraction using structurally symmetric, functionally orthogonal, and parallel spatial and temporal flows to obtain spatial and temporal features, including: The spatial flow employs a risk topology network based on Chebyshev polynomial approximation to partition the multimodal fusion input tensor. Each independent graph signal segment is processed, where the risky topology network is approximated by Chebyshev polynomial truncation unrolling to obtain the spatial feature tensor. for: in, This is the scaled, normalized Laplace matrix. for Chebyshev polynomials For the frequency domain convolution kernel parameter tensor that the network can learn, For multimodal fusion input tensors, Chebyshev polynomial truncation order, For activation functions; The time flow uses a time-selective state space to partition the multimodal fusion input tensor into... Given a set of independent risk sequences, a dynamic transition matrix is ​​generated from the risk sequences using discretized parameters dependent on the input. Based on the dynamic transition matrix, a linear recursive operator that evolves dynamically over time is constructed as follows: in, As a time feature, It is a dynamic transition matrix. It is obtained through a linear mapping of the input tensor via multimodal fusion; The process involves calculating interaction gating coefficients for spatial and temporal features using an ST-Gating gated neural network and a cross-modal cross-attention mechanism based on cue learning, and then performing hard fusion reconstruction using a temperature scaling factor during calculation to obtain the fused spatiotemporal features, including: Tensor of discrete factors of external environment As contextual soft hint vectors, construct the query matrix. , where the query matrix From spatial feature tensor Temporal feature tensor Node embedding and time embedding The result is obtained by combining and linearly mapping the matrix; using the query matrix. The query matrix is ​​used for cross-attention operations, and is expressed as a discrete factor tensor of the external environment. for Perform cross-attention calculation to obtain the gating coefficient tensor. for: in, For cross-attention operations, For learnable weight matrix, This is the temperature scaling factor. For bias terms, For activation functions; Based on the gating coefficient tensor, point-pair weighted fusion is performed on the parallel two-stream features to obtain the fused spatiotemporal features. for: in, The Hadamard product is a two-stream feature package containing spatial features. and time characteristics .

2. The method for early warning of cascading road congestion and secondary accident risks according to claim 1, characterized in that, The process involves acquiring historical traffic flow tensors, temporal metadata, and discrete external environmental factors of the target traffic network, embedding them into vectors through a heterogeneous data processing pipeline, and then performing nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor, including: Obtain historical traffic flow data collected by multiple sensors in the target traffic network, and convert the historical traffic flow data into a tensor. A high-dimensional feature vector is obtained by non-linearly mapping a 1D convolutional layer that slides along the time dimension. ,in For the number of sensors, For historical time steps, The number of feature channels of the sensor. To pre-define a unified target hidden dimension; The road network time metadata is then encoded using the Time2Vec sinusoidal activation function mechanism for absolute periodicity, resulting in: in, Indicates the first Tensors of the dimension of embedded features For the input time scalar, For the dimensional boundaries of the embedded features, and These are the frequency and phase parameters that the model can learn during backpropagation training. Using discrete states among the discrete factors of the external environment as indices, the discrete factors of the external environment are mapped to dense real-number vectors of a specific length and queried to obtain embedding vectors. These embedding vectors are then subjected to dimensionality increase and nonlinear transformation using a fully connected feedforward network layer including the ReLU activation function to obtain the tensor of the discrete factors of the external environment. ; Self-supervised learning is used to analyze high-dimensional feature vectors that are in the same time window. and discrete factors of the external environment tensor As positive sample pairs, high-dimensional feature vectors from different time windows and discrete factors of the external environment tensor As negative sample pairs, cross-modal contrastive learning and semantic alignment are performed using the InfoNCE loss function to obtain the multimodal fusion input tensor. , where the InfoNCE loss function for: in, The cosine similarity between feature vectors For temperature hyperparameters, For time windows.

3. The method for early warning of cascading road congestion and secondary accident risks according to claim 1, characterized in that, The process involves obtaining the initial static distance adjacency matrix of the target traffic network, constructing a dynamic residual matrix using Hawkes process decay, and then performing a weighted evolution of the final adjacency matrix on the fused spatiotemporal features using the initial static distance adjacency matrix and the dynamic residual matrix. Finally, the Huber Loss objective loss function is used for prediction optimization to obtain a road cascading congestion and secondary accident risk warning model, including: Obtain the initial static distance adjacency matrix of the topological physical structure of the target transportation network. Construct two learnable node embedding representation matrices. ; During the modulation of the Hadamard product, a dynamic residual matrix is ​​constructed using the exponential time decay factor of the Hawkes process, where the exponential time decay factor... for: in, For the risk dissipation rate parameter, To use the discrete factors tensor of the external environment Dynamic variables obtained through real-time mapping For the moment when an external emergency occurs, This is the current simulation moment; For dynamic residual matrix, This is the pooling average vector of the external features. This is the scaling factor; Represents the hyperbolic tangent function. Represents the linear rectification activation function; For the dynamic residual matrix and the initial static distance adjacency matrix Perform learnable weighted summation to generate a true dynamic adjacency matrix that participates in the convolution calculation of the current frame. for: in, These are weight parameters; The fused spatiotemporal features After performing residual connection and layer normalization processing, the predicted values ​​are obtained by predicting multiple consecutive time steps in the future through a two-layer nonlinear connection network. By repeatedly performing predictive learning, a road-cascade congestion and secondary accident risk warning model is obtained.

4. The road cascading congestion and secondary accident risk early warning method according to claim 3, characterized in that, The method further includes: The Huber Loss objective loss function is used for prediction optimization, where the Huber Loss objective loss function is: in, For predicted values, The actual tags returned by the sensors. The threshold value is used.

5. A road cascade congestion and secondary accident risk early warning device, characterized in that, include: The first module is used to obtain the historical traffic flow numerical tensor, time metadata, and discrete external environmental factors of the target traffic network, and to perform vectorized embedding through a heterogeneous data processing pipeline, and then perform nonlinear mapping and cross-modal comparative learning to obtain a multimodal fusion input tensor. The second module is used to perform dual-stream lightweight feature extraction on the multimodal fusion input tensor using structurally symmetric, functionally orthogonal, and parallel spatial and temporal streams to obtain spatial and temporal features. The third module is used to calculate the interaction gating coefficients of spatial and temporal features through the ST-Gating gated neural network and the cross-modal cross-attention mechanism based on cue learning, and to perform hard fusion reconstruction through the temperature scaling factor during the calculation to obtain the fused spatiotemporal features, where the temperature scaling factor is used to characterize the severity level of external extreme events; The fourth module is used to obtain the initial static distance adjacency matrix of the target traffic network, construct the dynamic residual matrix using Hawkes process decay, and perform a weighted evolution of the final adjacency matrix on the fused spatiotemporal features using the initial static distance adjacency matrix and the dynamic residual matrix to obtain the road cascading congestion and secondary accident risk warning model. The fifth module uses a road cascading congestion and secondary accident risk warning model to predict the target traffic network and obtain the road cascading congestion and secondary accident risk warning results. The multimodal fusion input tensor is subjected to dual-stream lightweight feature extraction using structurally symmetric, functionally orthogonal, and parallel spatial and temporal flows to obtain spatial and temporal features, including: The spatial flow employs a risk topology network based on Chebyshev polynomial approximation to partition the multimodal fusion input tensor. Each independent graph signal segment is processed, where the risky topology network is approximated by Chebyshev polynomial truncation unrolling to obtain the spatial feature tensor. for: in, This is the scaled, normalized Laplace matrix. for Chebyshev polynomials For the frequency domain convolution kernel parameter tensor that the network can learn, For multimodal fusion input tensors, Chebyshev polynomial truncation order, For activation functions; The time flow uses a time-selective state space to partition the multimodal fusion input tensor into... Given a set of independent risk sequences, a dynamic transition matrix is ​​generated from the risk sequences using discretized parameters dependent on the input. Based on the dynamic transition matrix, a linear recursive operator that evolves dynamically over time is constructed as follows: in, As a time feature, It is a dynamic transition matrix. It is obtained through a linear mapping of the input tensor via multimodal fusion; The process involves calculating interaction gating coefficients for spatial and temporal features using an ST-Gating gated neural network and a cross-modal cross-attention mechanism based on cue learning, and then performing hard fusion reconstruction using a temperature scaling factor during calculation to obtain the fused spatiotemporal features, including: Tensor of discrete factors of external environment As contextual soft hint vectors, construct the query matrix. , where the query matrix From spatial feature tensor Temporal feature tensor Node embedding and time embedding The result is obtained by combining and linearly mapping the matrix; using the query matrix. The query matrix is ​​used for cross-attention operations, and is expressed as a discrete factor tensor of the external environment. for Perform cross-attention calculation to obtain the gating coefficient tensor. for: in, For cross-attention operations, For learnable weight matrix, This is the temperature scaling factor. For bias terms, For activation functions; Based on the gating coefficient tensor, point-pair weighted fusion is performed on the parallel two-stream features to obtain the fused spatiotemporal features. for: in, The Hadamard product is a two-stream feature package containing spatial features. and time characteristics .

6. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the road cascading congestion and secondary accident risk warning method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The storage medium stores a program, which is executed by a processor to implement the road cascading congestion and secondary accident risk warning method as described in any one of claims 1-4.

Images