Traffic resilience assessment method based on diffusion probability model data reconstruction

Abstract

The application belongs to the field of intelligent transportation, and proposes a traffic resilience evaluation method based on diffusion probability model data reconstruction, including the following steps: step 1: reconstructing original resilience index data based on a diffusion probability model; step 2: cooperatively using original resilience index data and reconstructed data for resilience index prediction; and step 3: establishing a comprehensive resilience evaluation algorithm based on the shortest path concept, taking into account functional resilience and structural resilience, to realize real-time and certain predictive urban traffic network resilience evaluation. The application can effectively judge the performance of the traffic system under the influence of low-probability sudden events, realize multi-granularity, comprehensive and dynamic traffic resilience evaluation, and guide decision-making to enhance the resistance, recovery and adaptability of the urban road network traffic system to sudden events.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation, specifically a traffic resilience assessment method based on diffusion probability model data reconstruction. Background Technology

[0002] Urban transportation networks serve as the vital links and arteries of urban economic and social activities, forming the foundation for the normal operation of a city. Low-probability emergencies affecting urban transportation networks, such as extreme weather and traffic accidents, can lead to traffic congestion and loss of accessibility, significantly impacting network performance and potentially causing prolonged paralysis in some vulnerable areas. Therefore, effectively assessing the resilience and recovery capabilities of urban transportation networks in the face of sudden disruptions has become an urgent requirement for promoting socio-economic development and improving people's livelihoods. Previous research on urban transportation network performance primarily relied on indicators such as vulnerability, robustness, reliability, and survivability. However, these indicators differ in their focus and perspective, failing to provide a comprehensive assessment of urban transportation network performance.

[0003] The concept of traffic resilience has emerged to encompass the entire range of assessments of urban traffic network performance. It can be used to measure the resistance, recovery, and adaptability of urban traffic systems to external shocks, and to comprehensively evaluate the reliability, vulnerability, and robustness of traffic systems in the face of sudden events.

[0004] Resilience index prediction technology:

[0005] Continuously recorded resilience indicators, especially functional indicators such as speed and flow, can be considered a time-series signal. Considering the spatial correlation of data observation points within the road network of a traffic system, its data structure can be modeled as a graph structure. Research on the evolution of traffic resilience indicators can be categorized as a data mining task of spatiotemporal graph signals. Classical data-driven methods include Vector Autoregression (VAR) and Autoregressive Integral Moving Average (ARIMA), which are early works on traffic data sequences. In recent years, deep learning methods have been widely used in time-series modeling tasks, especially graph neural networks, which have been widely used in the transportation field and have produced many excellent results. Common graph neural network modules include Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT). They are mostly based on the topological adjacency matrix of the graph to learn the spatial correlation features between graph vertices. According to the way temporal features are extracted, spatiotemporal graph deep learning models can be further divided into convolutional neural networks (CNNs), recurrent neural networks (RNNs), and temporal attention mechanisms.

[0006] Existing methods for predicting traffic resilience indicators lack the ability to perceive the non-periodic anomalies unique to low-probability events, which leads to a significant decrease in the predictive ability of resilience indicators under low-probability events such as heavy rainfall.

[0007] Resilience assessment techniques:

[0008] Currently, most methods for assessing traffic resilience rely on simulation or topology model analysis. Simulation-based methods pre-define resilience quantification indicators and calculate resilience performance by analyzing changes in these indicators before and after simulated disturbance events. Topology model analysis methods quantify and compare the topological properties before and after resilience disturbance events. These methods have several limitations. First, they offer a relatively singular perspective on resilience, typically using only functional or structural resilience to assess the resilience of the traffic system. Second, simulation-based methods are often sensitive to parameters; potential flaws in the model parameter design process can lead to results that deviate from reality. Furthermore, simulation-based methods suffer from poor real-time performance, failing to assess resilience changes promptly under sudden events and thus unable to guide decision-making in such situations. Summary of the Invention

[0009] To address the problem that existing resilience assessment techniques cannot comprehensively assess and accurately predict the performance of urban traffic networks under the influence of low-probability emergencies, this invention proposes a traffic resilience assessment method based on diffusion probability model data reconstruction.

[0010] For low-probability emergencies in transportation networks, this invention takes a traffic resilience perspective, using network resilience indicators and network topology diagrams as data sources. It applies deep learning methods for spatiotemporal graph data mining and captures abnormal features based on diffusion probability model data reconstruction. Even under low-probability emergencies, it can accurately predict the network functional resilience indicators. Based on the obtained predicted functional resilience indicators, a set of resilience assessment algorithms combining function and structure is used to finally conduct a comprehensive resilience quantitative assessment of the urban road network's ability to cope with low-probability emergencies with a certain predictive ability.

[0011] The technical solution of this application is as follows:

[0012] The traffic resilience assessment method based on diffusion probability model data reconstruction includes the following steps:

[0013] Step 1: Reconstruct the original resilience index data based on the diffusion probability model;

[0014] Step 2: Collaboratively utilize the original resilience index data and the reconstructed data to predict resilience indicators;

[0015] Step 3: Establish a comprehensive resilience assessment algorithm based on the shortest path concept that takes into account both functional and structural resilience, so as to achieve real-time and predictive urban traffic network resilience assessment.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0017] This invention utilizes the characteristic that the learning objective of the diffusion probability model is the probability distribution of the original data, and innovatively designs a traffic data reconstruction model based on the diffusion probability model. The aim is to separate the non-periodic anomaly features unique to low-probability events by reconstructing the data by restoring the probability distribution of the original data.

[0018] This invention utilizes the combined and differential features in the original and reconstructed data to design a multi-head output fusion resilience index prediction model. By adjusting the fusion weights using the separated non-periodic anomaly features, the prediction accuracy of the resilience index can be improved under low-probability events such as heavy rainfall.

[0019] This invention establishes a comprehensive resilience assessment algorithm based on the shortest path, which takes into account both functional and structural resilience. It is based on historical big data benchmarks and also makes comprehensive use of real-time reconstruction and prediction data to achieve a certain degree of predictive comprehensive resilience assessment of the transportation system.

[0020] This invention can effectively assess the performance of a traffic system under resilient disturbance events, enabling multi-granular, comprehensive, and dynamic traffic resilience assessment, and guiding decision-making to enhance the resistance, recovery, and adaptability of urban road network traffic systems to emergencies. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0022] Figure 2 This is a schematic diagram of the data reconstruction model structure from step 1.

[0023] Figure 3 A schematic diagram of the overall structure of the data reconstruction model and the resilience index prediction model;

[0024] Figure 4 Examples include speed prediction fitting curves under heavy rain and normal weather conditions;

[0025] Figure 5 A heatmap for velocity prediction was fitted to the example.

[0026] Figure 6 This is a flowchart for step 3, the overall road network resilience assessment.

[0027] Figure 7 Visualization of the topology of the traffic network for example;

[0028] Figure 8 Heatmaps of edge resilience and node resilience indices for the PEMS-BAY road network are provided as an example.

[0029] Figure 9 This is an example of a comprehensive resilience benchmark for weekdays and weekends.

[0030] Figure 10 This example demonstrates the calculation of changes in the global comprehensive resilience index and the global resilience loss within one day.

[0031] Figure 11 This example compares the trends of the comprehensive toughness index with those of the single speed index. Detailed Implementation

[0032] Related technical terms:

[0033] Resilient disturbance events: Low-probability sudden events in the transportation system, such as extreme weather and traffic accidents.

[0034] Resilience indicators are data used to quantitatively assess the resilience of a transportation system, reflecting the strength of a certain attribute or function. Resilience indicators are mainly divided into two types: structural resilience indicators and functional resilience indicators. Structural resilience indicators extract one or more quantitative indicators from the topology of the road network based on complex network theory. Functional resilience indicators are the characteristics exhibited by users of the transportation system, such as traffic speed, traffic flow, road occupancy, and travel time.

[0035] Spatiotemporal graph data: This type of data possesses both time and spatial domain characteristics, with the spatial domain characteristics represented using graph structures in computer data structures. The time domain representation is the time-varying curves of attributes within the data, while the spatial domain representation is a graph adjacency matrix. Data collected by sensor networks in transportation networks is a typical example of spatiotemporal graph data. In the time domain, each network sensor is a graph node, and the one or more traffic indicator data collected by each sensor are one or more time-varying curves. In the spatial domain, edges describing the correlations between sensors can be established based on parameters such as sensor distance; the set of these correlated edges can be described as a graph adjacency matrix.

[0036] Resilience index prediction technology: It applies statistical methods, or machine learning and deep learning methods, to build a prediction model based on the patterns of historical data. Then, it uses the prediction model to predict the time-varying curve of the resilience index for a period of time after a certain point in time, based on the time-varying curve of the resilience index for a period of time before that point in time.

[0037] Resilience assessment: For the entire process of a resilience disturbance event, based on the changes of one or more selected resilience indicators during the process, a reasonable statistical analysis algorithm is used to quantitatively assess the changes in the resilience of the transportation system.

[0038] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0039] like Figure 1As shown, the traffic resilience assessment method based on diffusion probability model data reconstruction includes three steps:

[0040] Step 1: Reconstruct the original resilience index data based on the diffusion probability model;

[0041] Step 2: Collaboratively utilize the original resilience index data and the reconstructed data to predict resilience indicators;

[0042] Step 3: Establish a comprehensive resilience assessment algorithm based on the shortest path concept that takes into account both functional and structural resilience, so as to achieve real-time and predictive urban traffic network resilience assessment.

[0043] The following examples, using the publicly available traffic dataset PEMS-BAY, specifically illustrate the steps of a traffic resilience assessment method based on diffusion probability model data reconstruction.

[0044] Step 1: Construct a data reconstruction model based on the principle of diffusion probability model to reconstruct the input original traffic data.

[0045] The data reconstruction model takes as input the original resilience index data and the prior traffic map adjacency matrix A, and outputs the reconstructed resilience index data. The purpose of this data reconstruction model is to utilize the principles and characteristics of the diffusion probability model to generate reconstructed data that conforms to the probability distribution of the original data. This allows for the extraction of non-periodic anomaly features from the differences between the original and reconstructed spatiotemporal map data, providing support for subsequent accurate predictions. The overall principle diagram of this process is as follows: Figure 2 .

[0046] In this embodiment, the resilience index is mainly derived from traffic speed, and the spatiotemporal map data refers to the collection of speed data collected by all speed sensors in the entire road network over a period of time and the adjacency matrix between the sensors.

[0047] Step 1.1: Input the raw resilience index data X 1:2L Temporal and spatial domain features are extracted to construct conditional information for auxiliary reconstruction.

[0048] The original toughness index data X 1:2L This includes resilience index information (traffic speed in this example) from N nodes across 2L consecutive time slices, taking the first L time slices X... 1:L Used for condition information The calculation takes the last L time slices X. L+1:2L As the fragment to be reconstructed, that is, the target of data reconstruction.

[0049] The condition information This includes information relating to the time and space domains.

[0050] The first L time slices X of the original toughness index data 1:L First, it is converted to the implicit representation H using a linear mapping Linear(*). X =Linear(X) 1:L ).

[0051] For the first L time slices X of the original toughness index data 1:L The spatial domain features are extracted using graph convolution and spatial attention mechanisms, and the resulting features are superimposed to obtain spatial correlation features.

[0052] For the first L time slices X of the original toughness index data 1:L The temporal features are extracted using a temporal attention mechanism.

[0053] The extracted temporal and spatial features are superimposed, and after normalization and linear mapping, they are adjusted to match the implicit representation H of the original resilience index data. X Same size, and with H X Overlay, conditional information The calculation formula can be described as formula (1).

[0054]

[0055] X in the formula 1:L Refers to the first L time slices of the input, Linear(*) refers to the linear mapping, H X For X 1:L The implicit representation obtained after linear mapping, Norm refers to normalization, Attn Spa (*) and Attn Tem (*) refers to spatial attention and temporal attention mechanisms, respectively. GCN(*,A) refers to graph convolution based on the traffic graph adjacency matrix A. This is conditional information. Both spatial and temporal attention employ the scaled dot product self-attention method for calculating X. 1:L The implicit representation H obtained through linear mapping X The calculation formulas for spatial attention and temporal attention mechanisms are (2) and (3), respectively.

[0056]

[0057]

[0058] In the formula H X It is X 1:L The implicit representation of linear mappings, It is a learnable spatial self-attention transformation matrix. It is a learnable temporal self-attention transformation matrix, d is the number of channels after mapping, and Softmax(*) is the exponential normalization method.

[0059] Step 1.2: For the fragment X to be reconstructed in the original resilience index data L:2L Positive addition of Gaussian random noise ∈ t Obtain t-step noisy data Compared with the condition information obtained in step 1.1 A common input to a noise prediction neural network With conditional information As an aid, to accurately predict the Gaussian random noise ∈ ∈ a specified number of steps t added during the forward process. t The goal is to train a noise prediction neural network. Able to add noise to data in t steps Embedding representation of the number of noise-adding steps t embedding and condition information For positively added Gaussian random noise ∈ t Make accurate predictions.

[0060] The purpose of this step is to learn the latent characteristics of noise, thereby enabling an understanding of the overall probability distribution of the data. The underlying principle of this step is that the diffusion probability model models the forward noise addition and reverse denoising processes as two T-step Markov processes, thus establishing a close relationship between the probability distribution characteristics of the noise and the probability distribution characteristics of the original data.

[0061] The forward process of a diffusion probability model is called the diffusion process. A random sample conforming to the original data distribution is denoted as . After a T-step diffusion process (where T is sufficiently large), a sample approximating a Gaussian distribution will be obtained. Due to the properties of Markov processes, the prior probability distribution q during forward diffusion can be described by formulas (4) and (5).

[0062]

[0063]

[0064] In the formula β t It is a hyperparameter that controls the noise addition ratio. Here, I represents the Gaussian noise term, and I refers to the identity matrix. The data after t steps of noise addition is denoted as... Where α t =1-β t , And ∈ t It is random standard Gaussian noise.

[0065] In a Markov process of reverse process, the parameterized posterior probability distribution p θ It can be described as formulas (6) and (7).

[0066]

[0067]

[0068] In the formula It is the expectation of the parameterized Gaussian noise term. This is the variance coefficient of the Gaussian noise term. The reverse process utilizes conditional information. The graph adjacency matrix A is used to assist in reverse denoising. In both the forward and reverse processes, the transition probabilities of the Markov processes described by equations (5) and (7) are modeled as Gaussian noise terms. Therefore, learning the probability distribution pattern of added noise is equivalent to mastering the distribution characteristics of the original data.

[0069] To implement the reverse denoising process using deep learning, the reverse process formula needs to be parameterized. This allows the backpropagation and gradient descent methods of deep learning to automatically update the neural network weight parameters and learn the probability distribution of noise. The Gaussian noise expectation in the reverse process formula (7) can be parameterized as formula (8), and the variance coefficient can be parameterized as formula (9).

[0070]

[0071]

[0072] In formula (8) Let θ represent a noise prediction neural network, where the learnable neural network parameters are denoted as θ. (Noise prediction neural network) All inputs include t-step noisy data Condition Information Given the prior traffic map adjacency matrix A and the number of noise addition steps t, the output is the predicted noise ∈ θ .

[0073] The noise prediction neural network The specific architecture is a multi-layered stacked residual network, with each layer including a spatiotemporal attention mechanism and a graph convolution module. First, for the number of noise-adding steps *t*, it is mapped to a 128-dimensional embedding encoding *t*. embedding This converts the step count information into a feature vector that is easy for neural networks to process, and the calculation formula is (10).

[0074]

[0075] Noisy data Condition Information The graph adjacency matrix A, along with t obtained from the noise addition steps t, embedding The inputs are then fed into a multi-layered residual network. The network consists of l layers, each with the same temporal attention, spatial attention, and graph convolution modules as in step 1.1. In each layer, the inputs are first... t embedding , After mapping to a uniform size and stacking them, the overall input is obtained. The temporal attention latent representation H obtained after passing through the temporal attention mechanism Tem Then, the obtained hidden representation is simultaneously subjected to graph convolution and spatial domain attention computation to obtain the spatial attention hidden representation H. Spa After being superimposed, the layers are passed through two different gated activation functions, sigmoid and tanh, and the results are multiplied at corresponding positions to obtain the final hidden representation H of this layer. fin The data flow in each layer of a multilayer network can be expressed by formulas (11) to (14):

[0076]

[0077]

[0078] H Spa =GCN(H Tem ,A)+Attn Spa (H Tem (13)

[0079] H fin =sigmoid(H Tem +H Spa )⊙tanh(H Tem +H Spa (14)

[0080] The implicit representation H obtained at each layer fin The feature channel dimension will be divided into two parts, one for residual connection, i.e., with... Overlay, replacing the input that should have been entered in the next layer. Another copy is directly used as the output of the li-th layer skip connection. The average of the outputs from all skip connections of the layers is directly used as the final feature, which is then passed through a multilayer perceptron to obtain the predicted noise ∈ θ The process can be described by formula (15).

[0081]

[0082] In the formula, li refers to the layer number of the residual connection, and l is the total number of layers. is the skip connection output of the li-th layer, and MLP(*) is a multilayer perceptron composed of multiple fully connected layers.

[0083] Noise prediction neural network The training objective is to make the posterior probability distribution of the parameterized inverse denoising process learned by the neural network approximate the true distribution. Since the similarity between the two distributions is usually measured by KL divergence, the KL divergence term is calculated by combining the true posterior probability distribution term with the parameterized probability distribution term learned by the neural network, ignoring the weight term, resulting in a simplified expression, namely formula (16).

[0084]

[0085] in Let be the expectation of the probability distribution, ∈ t Add noise to the known t steps, while ∈ θ To predict noise.

[0086] In formula (16) Represents neural networks The loss function (Loss) is the learning objective and optimization objective of the neural network. Based on this formula, the training method of the neural network can be clearly defined, that is, in the neural network... During the training process, first record the Gaussian random noise ∈ t steps added during the forward steps. t The specific data is then used to obtain the output prediction noise ∈ through a neural network. θ As can be seen from formula (16), the final optimization objective is actually to add noise ∈ t With prediction noise ∈ θ The similarity is determined by continuously minimizing the difference between the added noise and the predicted noise to ensure that the weight parameters in the neural network are updated correctly.

[0087] Step 1.3: Utilize the predicted noise obtained in Step 1.2 ∈ θ Obtain reconstruction data

[0088] According to formulas (8) and (9) From the calculation formula, the recursive formula for reverse reconstruction data can be derived as follows:

[0089]

[0090] The term ∈ in the formula represents randomly sampled Gaussian noise to increase the generation degree of freedom. According to this recursive formula, for noisy data at step t... The reverse denoising process first utilizes a neural network. Predicting the inverse noise and substituting it into the calculation yields... By recursively applying this method, the noise-reduced reconstructed data can eventually be obtained.

[0091] For each input data segment X to be reconstructed L+1:2L Generally, it is necessary to repeat the complete data reconstruction process multiple times and then take the average result as the final result. This is because, as a generative model, the optimization objective of the diffusion probability model is essentially to minimize the KL divergence between the probability distribution learned by the neural network and the target probability distribution it fits, as shown in formula (16). Theoretically, it can only ensure that the generated samples are close to the true probability distribution, rather than an exact autoregression between the input training samples and the output generated samples. Therefore, its prediction results have a certain degree of randomness. In order to obtain more reliable reconstructed data, it is necessary to perform multiple data reconstructions and take the average of the results. This invention selects the average of three reconstructions as the final reconstructed data.

[0092] Step 2: Utilize the original resilience index data X and its reconstructed data Collaborative prediction resilience indicators.

[0093] Specifically, this step establishes a traffic resilience index prediction model, collaboratively utilizes the combined features of reconstructed data and original data to generate multiple outputs, and uses the difference features between reconstructed data and original data to control the fusion ratio of multiple outputs, thereby dynamically adjusting to achieve accurate prediction.

[0094] The resilience index prediction model, as described, predicts the resilience index for a future period based on the resilience index over an input period. The input to this model is the original resilience index data X and its reconstructed data. The output is a predicted resilience index.

[0095] In step 2, the data reconstruction model based on the diffusion probability model principle trained in step 1 will be used as a fixed computational program. Its internal neural network weight parameters will remain fixed and will no longer participate in the weight parameter update of the neural network in step 2. It will only be responsible for reconstructing the original resilience index data X to obtain the reconstructed data.

[0096] Step 2 makes full use of the reconstructed data obtained in Step 1 The difference between the original data X and the reconstructed signal may contain unconventional anomalies, enabling accurate prediction of resilience index changes over a future period under both normal and resilient disturbance events. Traffic events are categorized into periodic and sudden events. Periodic fluctuations can be captured through feature extraction, while aperiodic sudden fluctuations are difficult to capture. Therefore, the difference between the reconstructed signal, which conforms to the overall data distribution characteristics, and the actual original signal can serve as anomaly detection. This continuous difference signal is converted into a weight control signal through certain operations, thereby controlling the weight ratio of the final weighted fusion of different output heads. By automatically selecting different fusion ratios under normal and abnormal situations, the output is dynamically adjusted to adapt to different scenarios. The overall model structure diagram of Step 1 and Step 2 is shown below. Figure 3 As shown:

[0097] The basic architecture of the resilience index prediction model is a multi-channel multilayer perceptron, which uses a spatiotemporal feature embedding mechanism and adds a GCN layer to better capture the temporal and spatial correlation of graph signal nodes.

[0098] Resilience index raw data X or reconstructed data This can be considered as spatiotemporal map data within a certain observation time window. The establishment of its spatiotemporal feature embedding includes three aspects: periodic features, spatial features, and real-time extracted features from the input data. Among these, the periodic features map the position of the observed data's time within a day and its position within a week into 32-dimensional unique codes, forming a total of 64-dimensional periodic feature codes E. Tem Spatial features are 32-dimensional randomly initialized feature codes E. Spa Additionally, the original resilience index data X and the reconstructed data will be input. Real-time feature extraction E after dimensionality reduction to 32 dimensions via convolution. sum This can be expressed as formula (18). Therefore, the resilience index consists of the original data X and the reconstructed data. The model is a comprehensive feature tensor E with 128 feature channels for each node. X All of its characteristics are E Tem E Spa E sum The splicing can be expressed as formula (19).

[0099]

[0100] E X =[E Tem E Spa E sum (19)

[0101] In the formula, [*,*,…,*] represents the concatenation of feature channels, and Conv represents the convolution operation.

[0102] The output structure of the resilience index prediction model is designed as a multi-head output with dynamically adjusted weights, i.e., multiple output heads Y with identical internal structures. i The output results are also provided, but the weights of each output mapping module are calculated from the weighted control parameter W, which is derived from the state of the current spatiotemporal graph data. i The final output result is obtained by weighted summation. This achieves the effect of real-time adjustment of the output strategy based on the signal state. Each output head Y... i The system consists of a multilayer perceptron composed of fully connected layers, a graph convolutional layer, and a convolutional output layer. The data flow can be expressed as formula (20):

[0103] Y i =Conv(MLP(E X )+GCN(MLP(E X ))) (20)

[0104] The control weight W of the output head i From input data X and reconstructed data difference feature tensor The calculation yielded the result. For the difference characteristic tensor... It also uses the spatiotemporal feature embedding method, except that it uses a 64-dimensional periodic feature code E. Tem With 32-dimensional randomly initialized feature encoding E Spa In addition, the method of difference superposition is used to enhance the real-time feature extraction, which is to extract the difference feature E in real time. diff As shown in formula (21), The final result is a concatenation of the three types of features mentioned above, as shown in formula (22). The control weight W of the output head... i From the difference characteristic tensor After multiple fully connected multilayer perceptron and convolutional dimensionality reduction operations, a multi-head output Y is formed. i Control weights W of the same size i , as in formula (23).

[0105]

[0106]

[0107]

[0108] The final integrated output is the multi-head output Y. i In control parameter W i Weighted fusion under control, as shown in formula (24).

[0109]

[0110] To demonstrate the practical performance of the resilience index prediction model proposed in this invention under both normal and heavy rain conditions, experiments were conducted on the publicly available dataset PEMS-BAY. This dataset is specifically for traffic speed, containing 308 detector nodes over a continuous year, with speed data collected for each observation point every 5 minutes. Inputting the speed index for the past hour (12 discrete points), the model predicts the speed index for the next hour (12 discrete points). Existing traffic index prediction models typically learn the periodic characteristics of data. However, when urban road networks are affected by sudden events such as heavy rain, these periodic characteristics may be disrupted, leading to potential deviations in prediction accuracy across different spatiotemporal levels. Therefore, to demonstrate the robustness and generalization ability of this invention under various weather conditions, road network data under different weather conditions were selected as the test set for experiments. Two visualization examples reflecting the model's predictive capabilities are provided below.

[0111] The fitting curves of the velocity resilience index prediction model of this invention, comparing the prediction results with the actual values ​​under non-rainstorm and rainstorm weather conditions, are shown in the figure below. Figure 4 As shown. Considering the periodicity of traffic data, a day with heavy rain was selected ( Figure 4 (Left) and a day with normal weather 7 days later ( Figure 4 (Right) Showing the fitted curves. As you can see, regardless of whether it's heavy rain or normal weather, the model's predicted curve fits the actual curve well, proving that the model's predictions are quite accurate.

[0112] The heatmap comparing the predicted and actual values ​​of the velocity toughness index prediction model in this embodiment of the invention is shown below. Figure 5 As shown. In the heatmap, the vertical axis of each subplot represents the node number, and the horizontal axis represents the time of day. Each intersection of the vertical and horizontal axes represents the velocity value of a node at a specific time. The magnitude of the velocity value is indicated by the intensity of the color of the intersection; the greater the velocity measured by the monitor at a given time point, the darker the color of that point on the heatmap. The corresponding legend is shown in the figure. Figure 5 The far right. Therefore, each horizontal line on the heatmap corresponds to the numerical change curve of the corresponding node throughout the day. From Figure 5 Comparing the left and right subgraphs, their light and dark distributions are roughly the same, proving that the model has a good effect on predicting each node at all times of the day.

[0113] To demonstrate the superiority of this invention over other traffic indicator prediction methods under different weather conditions, a comparative experiment was conducted on the publicly available PEMS-BAYS dataset using the latest prediction model. The experimental results show that the speed prediction values ​​of this invention are superior to similar models in terms of mean error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) under both normal and rainy weather conditions.

[0114] Table 1 shows the comparison results with well-known models.

[0115]

[0116]

[0117] As shown in Table 1, compared to current mainstream and advanced methods, this invention outperforms all other methods in predictive performance under both normal and heavy rain conditions. Among these, Graph-WaveNet... [5] Temporal dilated convolution is introduced to obtain the receptive field while ensuring the necessary temporal causality for prediction tasks. ASTGCN [6] The model retains one-dimensional convolution and GCN as temporal and spatial features, respectively, and introduces an adaptive spatiotemporal attention mechanism, using multi-channel fusion to capture periodic features. ST-Norm [7] Frequency division processing expands the dimensions of feature extraction. STID [8] The indivisibility problem of spatiotemporal data prediction was discovered. By adding spatiotemporal feature embedding to the input, better prediction accuracy can be achieved with lower computational cost.

[0118] Step 3: Design a comprehensive resilience assessment algorithm based on the shortest path that takes into account both functional and structural resilience. Based on historical big data benchmarks, it also makes comprehensive use of real-time reconstruction and prediction data to achieve a certain degree of predictive comprehensive resilience assessment of the transportation system.

[0119] The evaluation criteria of the resilience assessment algorithm include not only the predicted resilience index for a certain period of time obtained through steps 1 and 2, but also the traffic map adjacency matrix A, which reflects the topology of the traffic system, and historical resilience index data over a long period of time (i.e., all the deep learning datasets used in steps 1 and 2).

[0120] Step 3 proposes functional resilience indices for graph nodes and edges based on the resilience indices from Steps 1 and 2 (specifically traffic speed in this invention). It also proposes structural resilience indices for graph nodes and edges based on the topological information contained in the graph adjacency matrix A. Furthermore, it aggregates the structural and functional resilience inherent in graph nodes and edges using the topological concept of shortest paths between node pairs, obtaining a comprehensive resilience index for each node. Finally, it statistically analyzes the loss of all nodes' comprehensive resilience indices during the entire resilience disturbance event compared to a historical baseline, thereby assessing the overall resilience level of the traffic system.

[0121] The flowchart for step 3 is as follows: Figure 6 As shown:

[0122] Step 3.1: Construct the spatial topology of the transportation network.

[0123] A topological model of the traffic network of the selected dataset is performed, and its visualization is as follows: Figure 7 As shown in the figure, each marked point represents a traffic observation point, and the line segment connecting a pair of observation points represents a logical edge between the two observation points.

[0124] Step 3.2: Establish a comprehensive resilience index algorithm, and use the resilience index reconstruction data and resilience index prediction data obtained in Step 1 and Step 2 to obtain the comprehensive resilience index of each node of the transportation network, providing a basis for the final resilience assessment.

[0125] The overall process of step 3.2 is as follows:

[0126] For each node in the road network topology diagram, calculate its node structure resilience index based on the topology diagram, and calculate its node function resilience index based on the resilience index reconstruction data obtained in step 1 and the resilience index prediction data obtained in step 2.

[0127] Then, for each edge in the road network topology graph, the edge structure resilience index is calculated based on the topology graph, and the edge function resilience index is calculated based on the function resilience index of the two nodes connected to the edge.

[0128] Subsequently, based on the concepts and algorithms related to shortest paths, the shortest path index between node pairs is calculated using the edge resilience index involved in the shortest path of node pairs.

[0129] Finally, nearest neighbor aggregation is performed on each node to obtain the overall resilience index of the node.

[0130] The node structural resilience index considers node centrality, and degree centrality and proximity centrality are selected. The calculation of the node structural resilience index is as shown in formulas (25)-(27). Among them, the degree centrality of formula (25) reflects the number of connections between a node and its surrounding nodes; while the proximity centrality of formula (26) reflects the step distance from a node to all other nodes; the node structural resilience index is determined by both degree centrality and proximity centrality, as shown in formula (27).

[0131] The node's functional resilience index is a stability-weighted node resilience index (referring to speed in this embodiment), as shown in formula (28). It takes into account both the node's real-time original speed and the stability of the node speed. The higher the real-time speed of the node, the stronger the node's resilience is considered. The stability of the node speed is determined by the difference between the reconstructed signal and the original signal. The larger the absolute difference, the more unstable the node speed is at this moment, and the worse the node's resilience is considered.

[0132]

[0133]

[0134] R structure (N i ) = C degree (N i )·C closeness (N i (27)

[0135] R function (N i ) = V node (N i )·(1-|δV reconstruct (i)|) (28)

[0136] In the formula, N is the total number of nodes. i Refers to the i-th node, i∈N, C degree (N i ) refers to node N i Normalized degree centrality, deg(Ni) represents node N i The degree, d ij R refers to the shortest path length between node i and node j. structure (N i ) and R function (N i ) respectively refer to node N i Structural and functional toughness indicators, of which V node (N i () refers to the raw velocity data, |δV reconstruct(i) | represents the normalized absolute difference between the reconstructed data obtained in step 1 and the original data, (1-|δV reconstruct (i) The factor reflects the stability of the velocity index during this period.

[0137] The edge structure resilience index, also known as the betweenness centrality of an edge, is the frequency of occurrence of a particular edge among all shortest paths. It reflects the probability of people traveling through this path in all trips in the road network and is calculated using formula (29). The edge function index is determined by the weighted sum of the velocity attributes of the two endpoints of the edge. The weight is the reciprocal of the degree of the endpoint. The design intention here is that the higher the degree of an endpoint, the less the edge is affected by that point. The calculation formula is (30). The edge comprehensive resilience index is the product of the edge structure index and the edge function index. The two are mutually restrictive, as shown in formula (31).

[0138]

[0139]

[0140]

[0141] In the formula, the edge formed by adjacent node pairs (i,j) is denoted as E. ij =e(i,j), C betweenness (E ij ) is edge E ij betweenness centrality, It is side E ij The structural resilience index is given by σ(s,t), which represents the number of shortest paths between any pair of nodes (s,t), and σ(s,t|e(i,j)) which represents the number of shortest paths between pair of nodes (s,t) that pass through edge e(i,j). It is side E ij Functional resilience index, It is side E ij The comprehensive resilience index.

[0142] The shortest path resilience index is calculated from the comprehensive resilience index of the edges. For any pair of nodes (i,j), for all shortest paths, the edge resilience indices of all edges in that path are calculated and summed. The average of these sums is then used as the shortest path index for that pair of nodes (i,j). The calculation formula is (32). After calculating the shortest path index for all node pairs, the shortest path resilience index of each node is calculated again, taking each node as the center. The process is as follows: the shortest path attribute of each node to all other k-order nearest neighbor nodes is first divided by the square of the corresponding shortest path to reduce the influence of distant nodes. The average is then calculated for each order in k, and the results are summed. The process can be described as formula (33). The shortest path resilience index of a node can be used as another attribute factor of the node. Multiplying it by the original node function-structure comprehensive attribute, the node's function-structure-shortest path comprehensive attribute is obtained. This process can be described as formula (34).

[0143] Specifically, the shortest path index between any pair of nodes (i,j) is calculated as shown in formula (32), and finally an N×N shortest path index matrix of node pairs is obtained, where each element of the matrix represents the shortest path index of the pair of nodes (i,j).

[0144]

[0145] In the formula It is the shortest path index for node pair (i,j). SPs refers to the set of all shortest paths, and sp refers to a specific shortest path. A shortest path sp may contain multiple edges E along the path. st .

[0146] Specifically, the formula for calculating the shortest path resilience index of a node is formula (33), which involves node pairs consisting of the node and its k-th nearest neighbors. Finally, considering the three elements of "structure, function, and shortest path", the formula for calculating the comprehensive resilience index of a node is formula (34).

[0147]

[0148] R comprehensive (N i ) = R function (N i )·R structure (N i )·R SPD (N i (34)

[0149] In the formula R SPD (N i ) is node N i Based on the shortest path resilience index of related node pairs, kNeighbors(N i ) refers to node N i The set of k-th order nearest neighbors, count(kNeighbors(N) i )) refers to node N iThe number of elements in the k-th order nearest neighbor set, d ij This refers to the shortest path length between node pairs (i,j).

[0150] The effectiveness of the resilience index established in this step can be qualitatively observed using heatmaps of the node comprehensive resilience index and the edge comprehensive resilience index. For example... Figure 8 In the heat map shown, Figure 8 (a) is a heatmap of the edge resilience index of the PEMS-BAY road network. Figure 8 (b) is a heatmap of node resilience indices; the darker the color of a node or edge, the greater its resilience index. It can be observed that nodes and edges near overpasses in the central urban area with higher node degrees generally have higher overall resilience indices, while those in the suburbs with lower node degrees and lower edge betweenness centrality generally have lower resilience values. Furthermore, the resilience indices of nodes on the same road exhibit a geographically gradual trend. These characteristics align with common sense and prior knowledge in the transportation field, demonstrating the rationality of the resilience index design.

[0151] Step 3.3: Establish a global comprehensive resilience assessment method, thereby using the comprehensive resilience index obtained in Step 3.2 as a benchmark, and based on historical traffic big data, to give the final global comprehensive resilience assessment result for the comprehensive resilience capability of the traffic system throughout the entire process of resilience disturbance events.

[0152] The specific process of this step can be summarized as follows: Based on the node comprehensive resilience index in step 3.2, the comprehensive resilience index benchmark is calculated and saved using historical traffic big data; based on the comprehensive resilience index benchmark, through a global comprehensive resilience index statistical analysis, the comprehensive resilience assessment of the predicted resilience index obtained in step 2 can be realized throughout the entire process of resilience disturbance events.

[0153] Step 3.3.1: Establish a comprehensive resilience index benchmark using historical traffic big data.

[0154] Considering the distinct periodicity of the transportation system on a weekly basis, each day of the week is treated as a periodic research object, establishing seven independent comprehensive resilience index benchmarks for each of the seven days of the week. Each comprehensive resilience index benchmark can be further subdivided into comprehensive resilience index benchmarks for any given time of day. The comprehensive resilience index benchmarks are calculated using data from non-resilient disturbance event scenarios. In the experiments of this invention, the disturbance event scenario specifically refers to a "non-rainstorm scenario".

[0155] The calculation process for each comprehensive resilience index benchmark, taking the calculation of the "Tuesday comprehensive resilience index benchmark" based on the PEMS-BAY traffic speed resilience index dataset in the experiment of this invention as an example, can be described as follows: Extract all full-day data from the resilience index dataset for actual dates of Tuesday and non-rainstorm days. Since each full-day dataset includes 288 time slices, for all time slices at the same time in all full-day datasets, substitute them into the comprehensive resilience index algorithm in step 3.2, select a certain number of the most representative nodes, and take the average of all obtained comprehensive resilience indices. This average value is used as the comprehensive resilience index benchmark for the corresponding time slice on Tuesday. After performing this 288 times, the comprehensive resilience index benchmarks for all time slices on Tuesday are obtained. This calculation process can be described as the pseudocode shown in Table 2.

[0156] Table 2 shows the benchmark comprehensive resilience index calculated under non-rainstorm conditions.

[0157]

[0158]

[0159] Regarding the comprehensive resilience index benchmark calculation method described in step 3.3.1 and Table 2, when the selected number of nodes participating in the calculation is n... top When all key nodes are covered, the calculated benchmark value of the comprehensive resilience index can be considered a quantitative description of the global resilience capability benchmark of the road network; therefore, this value can be defined as the global comprehensive resilience index benchmark. When the selected nodes for calculation include only a single node (i.e., n = n...), the calculation is different. top (In the case where n = 1), the calculated benchmark value of the comprehensive resilience index can be defined as the single-node comprehensive resilience index benchmark. Correspondingly, in each single time slice used for comparison with the benchmark, n... top The average of the comprehensive resilience index of each node can also be defined as the global comprehensive resilience index of that time slice.

[0160] Some examples of the comprehensive resilience benchmark established in this step are as follows: Figure 9 As shown in the figure, the horizontal axis represents time of day, and the vertical axis represents the global comprehensive resilience index value. The example object selected is a weekday and a non-working day. The "obvious periodicity of the transportation system on a weekly basis" proposed in step 3.3.1 can also be seen from... Figure 9 This is fully demonstrated in the data. Observing the resilience benchmark derived from historical averages, it can be seen that the patterns of resilience changes between weekdays and weekends show significant differences, such as... Figure 9 As shown, weekday commuting peaks exhibit distinct "troughs," while weekend commutes show a more gradual pattern throughout the day. This demonstrates the rationale and necessity of establishing different resilience indicators over a 7-day cycle.

[0161] Step 3.3.2: Based on the comprehensive resilience index benchmark, conduct a resilience assessment of the entire process of the rainstorm resilience disturbance event.

[0162] For each time slice of the entire process of the studied rainstorm resilience disturbance event, the corresponding comprehensive resilience index benchmark can be found from step 3.3.1. The comprehensive resilience index throughout the entire process of the resilience disturbance event is subtracted from the comprehensive resilience index benchmark, and the differences are accumulated to obtain the quantitative evaluation value of the resilience disturbance event process. The pseudocode description of this process is shown in Table 3.

[0163] Table 3 shows the calculated results of the overall resilience assessment.

[0164]

[0165] A computational example of the resilience assessment algorithm established in this step is shown below. Figure 10 As shown. Figure 10 (a) On a specific day of heavy rain, based on the resilience index established in step 3.2, the global comprehensive resilience assessment method proposed in step 3.3 is used to obtain a comparison between the global resilience benchmark and the global comprehensive resilience index after global statistical analysis. Figure 10 (b) is actually a visualization of the resilience assessment algorithm operation process proposed in step 3.3.2, reflecting the difference between the global resilience benchmark and the global comprehensive resilience index at each moment, i.e., the global resilience loss. By accumulating and integrating the global resilience loss throughout the entire resilience disturbance event, a quantitative assessment result of the entire resilience disturbance event can be obtained.

[0166] To demonstrate the superiority of the comprehensive resilience index proposed in step 3.2 over a single index in step 3.3, a time period including rainfall is selected. Using both the comprehensive resilience index and the single traffic speed index, and employing the resilience benchmark algorithm proposed in step 3.3.1, the corresponding comprehensive resilience index benchmark and speed index benchmark are calculated for a given observation node. The resilience index variation curve for a specific day including the rainfall period is then plotted, as shown below. Figure 11 The gray area in the graph represents the period of rain, and each curve reflects the ratio of the real-time indicator at the corresponding moment to the historical baseline. Figure 11In this diagram, because the velocity index and the comprehensive index have different value ranges, the vertical axis uses a ratio of a certain index to its benchmark for a unified representation. When the ratio is less than 1, it indicates a state of resilience loss, and the lower the ratio, the worse the resilience. It can be observed that during the rainfall period from 9:00 to 11:00 in the diagram, the calculation of the nodal comprehensive resilience index is essentially a central aggregation based on topological structure, with a wider sensing range, thus more accurately capturing the resilience loss during this period. In contrast, the single velocity index does not have a sensing range outside the node itself, and therefore cannot accurately characterize the resilience loss during this period.

[0167] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

[0168] References

[0169] [1]Afrin T,Yodo NA survey of road traffic congestion measurestowards a sustainable and resilient transportation system[J].Sustainability,2020,12(11):4660.

[0170] [2]Zhou Y, Wang J, Yang H. Resilience of transportation systems: concepts and comprehensive review[J]. IEEE Transactions on Intelligent Transportation Systems, 2019, 20(12): 4262-4276.

[0171] [3]Ho J,Jain A,Abbeel P.Denoising diffusion probabilistic models[J].Advances in neural information processing systems,2020,33:6840-6851.

[0172] [4]Kong Z,Ping W,Huang J,et al.Diffwave:A versatilediffusion modelfor audio synthesis[J].arXiv preprint arXiv:2009.09761,2020.

[0173] [5]Wu Z,Pan S,Long G,et al.Graph wavenet for deep spatial-temporalgraph modeling[J].arXiv preprint arXiv:1906.00121,2019.

[0174] [6]Guo S,Lin Y,Feng N,et al.Attention based spatial-temporal graphconvolutional networks for traffic flow forecasting[C] / / Proceedings of theAAAI conference on artificial intelligence.2019,33(01):922-929.

[0175] [7]Deng J,Chen X,Jiang R,et al.St-norm:Spatial and temporalnormalization for multi-variate time series forecasting[C] / / Proceedings ofthe 27th ACM SIGKDD conference on knowledge discovery&data mining.2021:269-278.

[0176] [8]Shao Z,Zhang Z,Wang F,et al.Spatial-temporal identity:A simple yeteffective baseline for multivariate time series forecasting[C] / / Proceedingsof the 31st ACM International Conference on Information&KnowledgeManagement.2022:4454-4458。

Claims

1. A traffic resilience assessment method based on diffusion probability model data reconstruction, characterized in that, It includes three steps: Step 1: Reconstruct the original resilience index data based on the diffusion probability model; Step 2: Collaboratively utilize the original resilience index data and the reconstructed data to predict resilience indicators; Step 3: Establish a comprehensive resilience assessment algorithm based on the shortest path concept that takes into account both functional and structural resilience, so as to achieve real-time and predictive urban traffic network resilience assessment. Step 1 involves constructing a data reconstruction model based on the principle of diffusion probability modeling to reconstruct the input raw traffic data. The input data for the data reconstruction model are the original resilience index data and the prior traffic map adjacency matrix. The output data is the reconstructed resilience index data; the data reconstruction model utilizes the principles and characteristics of the diffusion probability model to generate reconstructed data that conforms to the probability distribution of the original data, thereby obtaining non-periodic anomaly features from the differences between the original spatiotemporal map data and the reconstructed spatiotemporal map data, providing support for subsequent accurate predictions; specifically including: Step 1.1: Input the raw resilience index data Temporal and spatial domain features are extracted to construct conditional information for auxiliary reconstruction. ; The original toughness index data include On a continuous time slice The resilience index information of each node, taking the previous value. A time slice Used for condition information The calculation, after taking A time slice As the segment to be reconstructed, i.e., the target of data reconstruction; The condition information This includes information relating to the time and space domains; Step 1.2: The segment to be reconstructed from the original resilience index data. Adding Gaussian random noise in the positive direction get Step-by-step noisy data The condition information obtained in step 1.1 A common input to a noise prediction neural network conditional information As an aid, to accurately predict the specified number of steps added during the forward process. Gaussian random noise As the target; the trained noise prediction neural network Able to base on Step-by-step noisy data Embedded representation of the number of noise-adding steps t and condition information Adding Gaussian random noise in the positive direction To make accurate predictions, output prediction noise. ; Step 1.3: Utilize the predicted noise obtained in Step 1.2 Obtain reconstruction data .

2. The method as described in claim 1, characterized in that, Step 1.1: Original toughness index data First, through linear mapping Convert to implicit representation ; For the original toughness index data The spatial domain features are extracted using graph convolution and spatial attention mechanisms, and the resulting features are superimposed to obtain spatial correlation features. For the original toughness index data The temporal features are extracted using a temporal attention mechanism; The extracted temporal and spatial features are superimposed, and after normalization and linear mapping, they are adjusted to the implicit representation of the original resilience index data. Same size, and with Overlay, conditional information The calculation formula is described as formula (1): In the formula The input before A time slice, Refers to linear mapping, for The implicit representation obtained through linear mapping. Normalization and These refer to spatial attention and temporal attention mechanisms, respectively. Refers to the adjacency matrix based on the traffic map Graph convolution, Conditional information; both spatial and temporal attention employ the scaled dot product self-attention calculation method. The implicit representation obtained through linear mapping The calculation formulas for spatial attention and temporal attention mechanisms are respectively formulas (2) and (3): In the formula yes The implicit representation of linear mappings, , , It is a learnable spatial self-attention transformation matrix. , , It is a learnable temporal self-attention transition matrix. The number of channels after mapping. This is an exponential normalization method.

3. The method as described in claim 1, characterized in that, Step 1.2: In a forward-diffusion Markov process, the prior probability distribution Described as formulas (4) and (5): In the formula It is a hyperparameter that controls the noise addition ratio. This is a Gaussian noise term. Refers to the identity matrix. The data after adding noise is denoted as ,in , ,and It is random standard Gaussian noise; In a Markov process of reverse process, the parameterized posterior probability distribution Described as formulas (6) and (7): In the formula It is the expectation of the parameterized Gaussian noise term. These are the variance coefficients of the Gaussian noise term; the reverse process utilizes conditional information. And graph adjacency matrix To assist in reverse denoising; in both the forward and reverse processes, formulas (5) and (7) describe the transition probabilities of the Markov process, which are both modeled as Gaussian noise terms. ; The Gaussian noise expectation parameter in the reverse process formula (7) is parameterized as formula (8), and the variance coefficient is parameterized as formula (9); In formula (8) Let represent a noise prediction neural network, where the learnable neural network parameters are denoted as . ; The noise prediction neural network The specific architecture is a multi-layered stacked residual network, with each layer including a spatiotemporal attention mechanism and a graph convolution module; firstly, for the number of noise addition steps... Map it to a 128-dimensional embedding code. This converts the step count information into a feature vector that is easy for neural networks to process, and the calculation formula is (10). Noisy data Conditional information Graph adjacency matrix Together with the number of noise-adding steps The converted The data flow in each layer of the multi-layered residual network is expressed as formulas (11) to (14): The implicit representation obtained at each layer The feature channel dimension will be divided into two parts, one for residual connection, i.e., with... Overlay, replacing the input that should have been entered in the next layer. Another one is directly used as the first Layer skip connection output The average value of the outputs from all skip connections of the layers is directly used as the final feature, which is then passed through a multilayer perceptron to obtain the prediction noise. The process is described by formula (15); In the formula The layer number of the residual connection. This represents the total number of floors. For the first The layer skip connection output, A multilayer perceptron consisting of multiple fully connected layers; Noise prediction neural network The expression for the loss function is given by formula (16). in The expected value of the probability distribution. For known Add noise step by step, and To predict noise.

4. The method as described in claim 1, characterized in that, Step 1.3: According to formulas (8) and (9) From the calculation formula, the recursive formula for reverse reconstruction data is derived as follows: In the formula The term is randomly sampled Gaussian noise to increase the generation degree of freedom; according to this recursive formula, for those in... Noisy data with added noise The reverse denoising process first utilizes a neural network. Predicting the inverse noise and substituting it into the calculation yields... By recursively applying this method, the noise-reduced reconstructed data can eventually be obtained. ; For each input data segment to be reconstructed The complete data reconstruction process is repeated multiple times, and the average result is taken as the final result.

5. The method as described in claim 1, characterized in that, Step 2: A traffic resilience index prediction model is established, which uses the combined features of reconstructed data and original data to generate multiple outputs, and uses the difference features between reconstructed data and original data to control the fusion ratio of multiple outputs, thereby dynamically adjusting to achieve accurate prediction. The resilience index prediction model predicts resilience indices for a future period based on resilience indices received over a given period. The basic architecture of the model is a multi-layer perceptron, employing a spatiotemporal feature embedding mechanism and adding a GCN layer to capture the temporal and spatial correlations of graph signal nodes. The input to this model is the raw resilience index data. Rather than rebuilding data The output is a predicted resilience index; Raw data of resilience index Or rebuild data For spatiotemporal map data within a certain observation time window, the establishment of its spatiotemporal feature embedding includes three aspects: periodic features, spatial features, and real-time extracted features from the input data. Among them, the periodic features map the position of the observed data's time within a day and its position within a week into 32-dimensional unique codes, forming a total of 64-dimensional periodic feature codes. Spatial features are 32-dimensional randomly initialized feature codes. Additionally, the raw data of the resilience index will be input. With reconstruction data Real-time feature extraction after dimensionality reduction to 32 dimensions via convolution. , expressed as formula (18); therefore, the original data of the resilience index and reconstructing data Modeled as a comprehensive feature tensor with 128 feature channels for each node All its characteristics are , , The concatenation of these elements is represented by formula (19). In the formula This represents the splicing of characteristic channels. Represents the convolution operation; The output structure of the resilience index prediction model is a multi-head output with dynamically adjusted weights, that is, multiple output heads with the same internal structure. The output results are also provided, but the weights of each output mapping module are calculated from the weighted control parameters derived from the state of the current spatiotemporal graph data. The final output result is obtained by weighted summation, thereby achieving the effect of real-time adjustment of the output strategy based on the signal state of the graph; each output head The data flow consists of a multilayer perceptron composed of fully connected layers, a graph convolutional layer, and a convolutional output layer, and is represented by formula (20): Output head control weight From input data With reconstruction data difference feature tensor The calculation shows that for the difference characteristic tensor... It also uses the spatiotemporal feature embedding method, except that it uses 64-dimensional periodic feature encoding. With 32-dimensional random initialization feature encoding In addition, the method of difference superposition is used to enhance real-time feature extraction, specifically for real-time difference feature extraction. , as in formula (21). The final result is a concatenation of the three types of features mentioned above, as shown in formula (22); the control weights of the output head. From the difference characteristic tensor After multiple fully connected multilayer perceptron and convolutional dimensionality reduction operations, a multi-head output is formed. Control weights of the same size For example, in formula (23); The final overall output is multi-head output. In control parameters Weighted fusion under control, as shown in formula (24): 。 6. The method as described in claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Construct the spatial topology of the transportation network; The traffic network of the selected dataset is modeled in topology. Each marked point in the topology graph represents a node, i.e., a traffic observation point. The line segment connecting a pair of observation points represents a logical edge between the two observation points. Step 3.2: Establish a comprehensive resilience index algorithm, and use the resilience index reconstruction data and resilience index prediction data obtained in Step 1 and Step 2 to obtain the comprehensive resilience index of each node of the transportation network, so as to provide a basis for the final resilience assessment. Step 3.3: Using the comprehensive resilience index obtained in Step 3.2 and based on historical traffic big data, give the final global comprehensive resilience assessment result for the comprehensive resilience capability of the traffic system throughout the entire process of resilience disturbance events.

7. The method as described in claim 6, characterized in that, Step 3.2 specifically includes: For each node in the road network topology diagram, calculate its node structure resilience index based on the topology diagram, and calculate its node function resilience index based on the resilience index reconstruction data obtained in step 1 and the resilience index prediction data obtained in step 2. Then, for each edge in the road network topology graph, the edge structure resilience index is calculated based on the topology graph, and the edge function resilience index is calculated based on the function resilience index of the two nodes connected to the edge. Subsequently, based on the shortest path algorithm, the shortest path index between node pairs is calculated using the edge resilience index involved in the shortest path of the node pair. Finally, nearest neighbor aggregation is performed on each node to obtain the node's comprehensive resilience index. The node structural resilience index takes into account node centrality, and selects degree centrality and proximity centrality, calculated as shown in formula (25)-formula (27); where, the degree centrality of formula (25) reflects the number of connections between a node and its surrounding nodes; while the proximity centrality of formula (26) reflects the step distance from a node to all other nodes; the node structural resilience index is determined by both degree centrality and proximity centrality, as shown in formula (27). The functional resilience index of a node is a weighted node resilience index, as shown in formula (28). In the formula, The total number of nodes. Refers to the first 1 node , Pointer node Normalized degree centrality Represents a node The degree, Pointer node With nodes The shortest path length between them. and Each refers to a node The structural and functional resilience indicators, among which Refers to the raw speed data. | represents the normalized absolute difference between the reconstructed data obtained in step 1 above and the original data. The factor reflects the stability of the speed index during this period; The edge structure resilience index, also known as the betweenness centrality of an edge, is the frequency of occurrence of a certain edge among all shortest paths, reflecting the probability of all trips in the road network passing through this path. It is calculated using formula (29). The edge function index is determined by the weighted sum of the velocity attributes of the two endpoints of the edge, with the weight being the reciprocal of the degree of the endpoints. The calculation formula is (30). The edge comprehensive resilience index is the product of the edge structure and edge function indices, which are mutually constrained, as shown in formula (31). In the formula, adjacent node pairs The edge notation is , It is the edge betweenness centrality, It is the edge Structural toughness index, Represents any pair of nodes The number of shortest paths between them, and Represents node pairs The shortest path between them is through the edges The number of items, It is the edge Functional resilience index, It is the edge The comprehensive resilience index; The shortest path resilience index is calculated from the comprehensive resilience index of the edges; Any node pair The calculation method for the shortest path index between them is shown in formula (32), and finally a result is obtained. The shortest path index matrix for node pairs, where each element represents a node pair. The shortest path metric; In the formula It is a node pair The shortest path metric It refers to the set of all shortest paths. Refers to a shortest path. Edge containing multiple paths ; The formula for calculating the shortest path resilience index of a node is formula (33). The calculation process involves the node and its... The node pairs are composed of nearest neighbors; finally, the calculation formula for the node comprehensive resilience index, which considers both "structure-function-shortest path", is formula (34). In the formula It is a node Based on the shortest path resilience index of related node pairs Pointer node of nearest neighbor set, Pointer node of The number of elements in the nearest neighbor set of order 1. Node pairs The shortest path length between them.

8. The method as described in claim 6, characterized in that, Step 3.3 specifically includes: Step 3.3.1: Establish a comprehensive resilience index benchmark using historical traffic big data; Step 3.3.2: Based on the comprehensive resilience index benchmark, conduct a resilience assessment of the entire process of the resilience disturbance event.

Images