Road transport vehicle driving risk uncertainty prediction method and system

Abstract

The invention provides a road transport vehicle driving risk uncertainty prediction method and system, and belongs to the technical field of vehicle driving risk prediction technology, and the method comprises the steps: obtaining an alarm number space-time sequence, a traffic flow space-time sequence and a multi-graph static adjacency matrix of a current time step; and processing the obtained alarm number time-space sequence, the traffic flow time-space sequence and the multi-graph static adjacency matrix by using a pre-trained road transport vehicle driving risk uncertainty prediction model to obtain the road transport vehicle alarm number in the future time step. According to the invention, accurate short-time prediction of the driving risk of the road transport vehicle is realized, and powerful technical support is provided for improving the reliability and real-time performance of vehicle driving risk early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle driving risk prediction, and in particular to a method and system for predicting uncertainty in driving risks of road transport vehicles based on deep learning. Background Art

[0002] In recent years, with the development of the internet and mobile devices, the vast amount of alarm and trajectory data recorded by onboard monitoring systems of road transport vehicles has provided a rich source of training data for prediction models, effectively overcoming the scarcity of accident data. This massive amount of data has enabled the widespread application of deep learning methods in the field of driving risk prediction. Regarding data sources, researchers are increasingly focusing on integrating multi-source data and multi-dimensional features. By incorporating various external factors such as the geographical environment, road conditions, and social environment, they are constructing more complex spatiotemporal prediction models.

[0003] In the road traffic system, road transport vehicles face a higher risk of accidents in adverse weather and high-risk road conditions due to the complex differences in vehicle types and cargo types. The consequences of accidents are also more serious than those of other types of vehicles. Therefore, the safety supervision of road transport vehicles has become a key and difficult issue in traffic safety management.

[0004] Existing risk prediction models are generally based on historical road traffic accident data. However, traffic accident data is rare and lagged, and is highly unevenly distributed across time and space. This severely impacts the model's training efficiency and generalization capabilities, leading to problems such as low prediction accuracy, poor real-time warning performance, and coarse spatiotemporal granularity. Furthermore, most existing models are deterministic point prediction models or single-distribution parameterized models. The former struggles to capture risk uncertainty, while the latter relies on specific statistical distributions and has weak generalization capabilities, making them difficult to meet practical application needs. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting the uncertainty of road transport vehicle driving risks based on deep learning, so as to solve at least one technical problem existing in the above-mentioned background technology.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting uncertainty in driving risk of road transport vehicles, comprising:

[0008] Obtain the spatiotemporal sequence of the number of alarms, the spatiotemporal sequence of traffic flow, and the multi-graph static adjacency matrix at the current time step;

[0009] The pre-trained road transport vehicle driving risk uncertainty prediction model is used for processing the obtained alarm number space-time sequence, vehicle flow space-time sequence and multi-graph static adjacency matrix to obtain the road transport vehicle alarm number at the future time step; wherein the road transport vehicle driving risk uncertainty prediction model comprises a dynamic graph construction layer, a multi-graph fusion module, a time and space feature extraction module and a multi-quantile prediction output module; the dynamic graph construction layer is used for extracting the dependency relationship between the alarm number and the vehicle flow and constructing a road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is used for calculating the feature information weight coefficient of each space graph structure and performing multi-graph fusion; the time feature extraction module is used for processing the alarm number space-time sequence and extracting the time dependency relationship thereof; the space feature extraction module is used for processing the fused graph and the alarm number space-time sequence and extracting the space dependency relationship thereof; and the multi-quantile prediction output module is used for fusing the feature information of all time steps and space graph structures, using a multilayer perceptron to estimate the median as the central reference, and obtaining the prediction output on multiple quantiles through the cumulative non-negative increment or decrement of the quantile.

[0010] As a further limitation of the first aspect of the application, the terrain and elevation data of the grid are obtained, the geographic environment vector is constructed, the cosine similarity method is used to measure the similarity of two vectors, and the geographic environment similarity adjacency matrix with the grid as the node is constructed.

[0011] As a further limitation of the first aspect of the application, the number of various POIs of the grid is obtained, and the POI vector is constructed, wherein the dimension of the POI vector is the number of POI categories, and the value is the POI density of each category in the grid; the cosine similarity method is used to measure the similarity of two POI vectors, and the social environment similarity adjacency matrix is constructed; the data of various grades of roads and bridges and tunnels of the grid are obtained, and a six-dimensional road network structure vector is constructed, including road density information, bridge density and tunnel density, and the cosine similarity is used to construct the road network structure similarity adjacency matrix; the central point longitude and latitude coordinates of each grid are obtained, the distance between each grid is calculated, and the grid distance adjacency matrix is constructed.

[0012] As a further limitation of the first aspect of the application, the dynamic graph construction layer comprises: trajectory space-time sequence data is firstly input in parallel to two feature extraction blocks block1 and block2, each block comprises a convolution layer, a batch normalization layer and a ReLU activation function, and the output is two feature vectors Then, a batch matrix multiplication, an activation function and a normalization processing are used to obtain an N*N matrix M flow , which realizes adaptive generation of the trajectory flow relationship dynamic graph between different grids according to the historical trajectory data.

[0013] As a further limitation of the first aspect of the application, the temporal feature extraction module is composed of multiple temporal convolution network layers, each TCN layer followed by a batch normalization layer and a Dropout layer; the spatial feature extraction module is composed of multiple diffusion graph convolution network layers, each D-GCN layer followed by a Dropout layer.

[0014] As a further limitation of the first aspect of the application, the region to be predicted is rasterized, the number of alarms in each grid is calculated, and each node of the KDTree_weighted model corresponds to each grid, and the weight of each node is the number of alarms in the grid; after sorting the nodes according to the spatial position, a position that makes the weight sum of the left and right sub-trees equal or closest to equal is found as the division hyperplane using binary search, and the left and right sub-trees are divided, so that the logic recursion eventually realizes the relative balance of the node weight sum in each leaf node, and realizes the balanced, non-overlapping and efficient spatial grid division.

[0015] In a second aspect, the application provides a road transport vehicle driving risk uncertainty prediction system, comprising:

[0016] The acquisition module is configured to acquire the alarm number spatio-temporal sequence, the traffic flow spatio-temporal sequence and the multi-graph static adjacency matrix at the current time step.

[0017] The processing module is configured to process the acquired alarm number spatio-temporal sequence, traffic flow spatio-temporal sequence and multi-graph static adjacency matrix using a pre-trained road transport vehicle driving risk uncertainty prediction model to obtain the number of alarms for the road transport vehicle at the future time step; wherein the road transport vehicle driving risk uncertainty prediction model comprises a dynamic graph construction layer, a multi-graph fusion module, a time and space feature extraction module and a multi-quantile prediction output module; the dynamic graph construction layer is configured to extract the dependency relationship between the alarm number and the traffic flow, and construct the road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is configured to calculate the feature information weight coefficients of each spatial graph structure and perform multi-graph fusion; the time feature extraction module is configured to process the alarm number spatio-temporal sequence and extract its temporal dependency relationship; the spatial feature extraction module is configured to process the fused graph and the alarm number spatio-temporal sequence and extract their spatial dependency relationship; and the multi-quantile prediction output module is configured to fuse the feature information of all time steps and spatial graph structures, use a multi-layer perceptron to estimate the median as the center reference, and obtain the prediction output on multiple quantiles by accumulating non-negative increments or decrements.

[0018] In a third aspect, the present application provides a non-transitory computer readable storage medium for storing computer instructions, which, when executed by a processor, implement the deep learning-based road transport vehicle driving risk uncertainty prediction method according to the first aspect.

[0019] In a fourth aspect, the present application provides a computer device comprising a memory and a processor, the processor and the memory being in communication with each other, the memory storing program instructions executable by the processor, and the processor invoking the program instructions to execute the deep learning-based road transport vehicle driving risk uncertainty prediction method according to the first aspect.

[0020] In a fifth aspect, the present application provides an electronic device comprising a processor, a memory and a computer program, wherein the processor is connected to the memory, and the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the deep learning-based road transport vehicle driving risk uncertainty prediction method according to the first aspect.

[0021] The present application has the advantages of high reliability and the ability to quantify uncertainty, taking into account the influence of geographical environment, social environment and road network structure on driving risk, and achieving accurate short-term prediction of road transport vehicle driving risk, thereby providing strong technical support for improving the reliability and real-time performance of vehicle driving risk warning and providing a solid theoretical foundation for improving road traffic safety from the time and space dimensions.

[0022] The advantages of the additional aspects of the present application will be more apparent from the following description section or will be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0024] Figure 1 The deep learning-based road transport vehicle driving risk uncertainty prediction method flowchart according to the embodiments of the present application.

[0025] Figure 2 The STQuantileNet model structure diagram according to the embodiments of the present application. DETAILED DESCRIPTION

[0026] Embodiments of the present application will be described in detail below with reference to the attached drawings, which are given by way of illustration and thus do not limit the present application. In addition, the present application can be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided by way of example so that this disclosure will convey the principles and concepts of the present application to those skilled in the art.

[0027] Those skilled in the art will appreciate that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0028] It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0029] As will be understood by one skilled in the art, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. It should be further understood that the terms "comprise," "comprises," "comprising," "include," "includes," "including," "contain," "contains," "containing," and any variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises, includes, contains or contains one or more features, integers, steps, operations, elements, or components can also contain or comprise other features, integers, steps, operations, elements, components, or groups thereof not expressly listed or inherent to such process, method, article, or apparatus.

[0030] In the description of the present application, the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like, mean that the particular feature, structure, material, or characteristic being described is included in at least one embodiment or example of the present application. Such phrases in context can refer to something generally or loosely related to the particular embodiments or examples. In many instances, such phrases are followed by additional language reciting one or more specific examples of implementing the feature, structure, material, or characteristic. Such phrases can refer to one or more implementations of the feature, structure, material, or characteristic.

[0031] In order to facilitate the understanding of the present application, the present application will be further explained with reference to the attached drawings and specific embodiments, which do not limit the embodiments of the present application.

[0032] Those skilled in the art will appreciate that the components in the drawings are only schematic and are not necessarily drawn to scale.

[0033] The core objective of the present application is to provide more accurate and reasonable driving risk prediction for road transport vehicles. The present application selects real-time alarm data and trajectory data of the vehicle-mounted monitoring system, collects geographic, social environment and road network data, establishes a multi-source database and dynamic and static graphs. In view of the uneven spatial distribution characteristics of the alarm data, a tree model with increased weight is proposed, and an adaptive grid division method for balancing the distribution of alarm numbers among grids is realized. In view of the risk prediction and uncertainty capture problem, a multi-quantile regression model integrating multi-spatiotemporal graph neural networks is innovatively proposed, realizing short-time accurate prediction of the driving risk of road transport vehicles, and effectively capturing the uncertainty of the driving risk, providing a theoretical basis and technical support for improving the road traffic risk identification efficiency and perfecting the accident prevention mechanism.

[0034] In order to realize the prediction of the uncertainty of the driving risk of road transport vehicles, the risk prediction method based on deep learning and multi-quantile regression prediction provided by the present application comprises: acquiring trajectory data and alarm data of road transport vehicles, and analyzing the spatiotemporal distribution characteristics thereof. In space, the alarm events present a significantly uneven, highly clustered distribution characteristic in some areas. In time, the trajectory and alarm data both present a periodic variation rule in units of "day", and there is a close positive correlation between the two. Considering the spatial distribution characteristics, a self-adaptive grid division model KDTree_weighted is proposed to realize the balanced distribution of alarm numbers among grids, realizing balanced, non-overlapping and efficient spatial grid division. Considering the time distribution characteristics, an input time slice structure composed of three parts is constructed, which are the adjacent time slice, the previous day time slice and the previous week time slice. The prediction model is effectively assisted to fully learn the short-term dynamic change and long-term periodicity of risk alarms. Geographic environment data, POI data and road network structure data are acquired, and multi-source data fusion and processing are performed. In order to capture the spatial dependence relationship, the grid is taken as a node, and a geographic environment similarity graph, a social environment similarity graph, a road network structure similarity graph and a grid distance graph are constructed. A multi-quantile regression model STQuantileNet based on spatiotemporal graph convolution is proposed to realize the prediction of the alarm number of road transport vehicles. The input data of the model is composed of the trajectory number of road transport vehicles, the alarm number and the three static graphs, and the output is the multiple quantile values of each grid at T_out time steps in the future, realizing the prediction of the risk uncertainty. Through the analysis and judgment of the prediction results, comparative experiments and ablation experiments are carried out to verify the effectiveness of the model and the application effect of each part.

[0035] The road transport vehicle driving risk uncertainty prediction method provided by the application has good prediction effect, high reliability, and the ability of uncertainty quantification, considers the influence of geographical environment, social environment and road network structure on driving risk, realizes accurate short-time prediction of road transport vehicle driving risk, provides strong technical support for improving the reliability and real-time performance of vehicle driving risk warning, and provides a solid theoretical basis for improving road traffic safety from the time and space dimensions.

[0036] Embodiment 1

[0037] In this embodiment 1, first, a road transport vehicle driving risk uncertainty prediction system is provided, comprising: an acquisition module for acquiring an alarm quantity spatio-temporal sequence, a traffic flow spatio-temporal sequence and a multi-graph static adjacency matrix of a current time step; a processing module for processing the acquired alarm quantity spatio-temporal sequence, traffic flow spatio-temporal sequence and multi-graph static adjacency matrix by using a pre-trained road transport vehicle driving risk uncertainty prediction model to obtain the road transport vehicle alarm quantity at a future time step; wherein the road transport vehicle driving risk uncertainty prediction model comprises a dynamic graph construction layer, a multi-graph fusion module, a time and space feature extraction module and a multi-quantile prediction output module; the dynamic graph construction layer is used to extract the dependency relationship between the alarm quantity and the traffic flow, and construct the road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is used to calculate the feature information weight coefficients of each spatial graph structure and perform multi-graph fusion; the time feature extraction module is used to process the alarm quantity spatio-temporal sequence and extract its time dependency relationship; the space feature extraction module is used to process the fused graph and the alarm quantity spatio-temporal sequence and extract its spatial dependency relationship; the multi-quantile prediction output module is used to fuse the feature information of all time steps and spatial graph structures, use the multi-layer perceptron to estimate the median as the central reference, and obtain the prediction output on multiple quantiles by accumulating non-negative increments or decrements.

[0038] The training method of the road transport vehicle driving risk uncertainty prediction model described in this embodiment comprises: acquiring training data; the training data comprises road transport vehicle alarm quantity spatio-temporal sequences, road transport vehicle flow spatio-temporal sequences and multi-graph static adjacency matrices in a plurality of historical time intervals. Based on the alarm quantity spatio-temporal sequence, traffic flow spatio-temporal sequence and multi-graph static adjacency matrix as input, and based on the road transport vehicle alarm quantity at a future time step as output, the driving risk uncertainty prediction model is trained.

[0039] In this embodiment, based on the above system, a road transport vehicle driving risk uncertainty prediction method is realized, comprising: the road transport vehicle driving risk uncertainty prediction model comprises an input layer, a road transport vehicle flow dynamic graph construction layer (hereinafter referred to as a "dynamic graph construction layer"), a multi-graph fusion module, a time and space feature extraction module, and a multi-quantile prediction output module.

[0040] The input layer is used to select a sliding window of appropriate size to process the alarm number spatio-temporal sequence and the vehicle flow spatio-temporal sequence respectively, and receive four static graph adjacency matrices of geographical environment similarity graph, social environment similarity graph, road network structure similarity graph and grid distance graph; the dynamic graph construction layer is used to extract the dependency relationship between the alarm number and the vehicle flow, and construct the road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is used to calculate the feature information weight coefficients of each spatial graph structure, and perform multi-graph fusion; the time feature extraction module is used to process the alarm number spatio-temporal sequence and extract its time dependency relationship; the space feature extraction module is used to process the fused graph and the alarm number spatio-temporal sequence and extract its spatial dependency relationship; the multi-quantile prediction output module is used to fuse the feature information of all time steps and spatial graph structures, use a special multi-layer perceptron (MLP) to estimate the median as the central reference, and then obtain the subsequent quantiles by accumulating non-negative increments or decrements, and these increments or decrements are generated by parallel MLPs with ReLU activation functions, to obtain the prediction output on multiple quantiles.

[0041] The construction of the geographical environment similarity graph comprises: obtaining the terrain and elevation data of the grid, constructing a geographical environment vector, and using the cosine similarity method to measure the similarity of two vectors, thereby constructing the geographical environment similarity adjacency matrix M geo .

[0042] The construction of the social environment similarity graph comprises: obtaining the number of various types of POIs of the grid, constructing a POI vector, wherein the dimension of the POI vector is the number of POI categories, and the value is the density of each type of POI in the grid. The cosine similarity method is used to measure the similarity of two POI vectors, thereby constructing the social environment similarity adjacency matrix M poi .

[0043] The construction of the road network structure similarity graph comprises: obtaining the data of various levels of roads and bridges and tunnels of the grid, constructing a six-dimensional road network structure vector, including the density information of four levels of roads (expressway and trunk road, urban main road, urban secondary road, urban branch road), bridge density, and tunnel density, and using the cosine similarity to construct the road network structure similarity adjacency matrix M road .

[0044] The construction of the grid distance map comprises: obtaining the central point longitude and latitude coordinates of each grid, calculating the distance between each grid, and constructing a grid distance adjacency matrix M dist .

[0045] The composition of the dynamic graph construction layer comprises: the trajectory space-time sequence data is first input in parallel to two feature extraction blocks block1 and block2, each block comprises a convolution layer, a batch normalization layer and a ReLU activation function, and the output is two feature vectors Then, a batch matrix multiplication, an activation function and a normalization processing are used to obtain an N*N matrix M flow , so as to adaptively generate a trajectory flow relationship dynamic graph G between different grids according to historical trajectory data flow (V, E, M flow ).

[0046]

[0047] The composition of the time feature extraction module comprises: the time feature extraction module is composed of multiple time convolution network (TCN) layers, and each TCN layer is followed by a batch normalization layer and a Dropout layer.

[0048] The composition of the space feature extraction module comprises: the space feature extraction module is composed of multiple diffusion graph convolution network (D-GCN) layers, and each D-GCN layer is followed by a Dropout layer.

[0049] The embodiment also provides a KDTree_weighted adaptive grid division model for realizing balanced distribution of alarm numbers among grids, and the model comprises the following steps: performing grid processing on a research range, calculating alarm numbers in each grid, and taking each node of the KDTree_weighted model as corresponding to each grid, and taking the weight of each node as the alarm number in the grid; after sorting the nodes according to spatial positions, finding a position that makes the total weight of left and right child trees equal or closest to equal as a division hyperplane by using binary search, and dividing the left and right child trees, so that the total weight of nodes in each leaf node is relatively balanced, and balanced, non-overlapping and efficient spatial grid division is realized.

[0050] Embodiment 2

[0051] The embodiment provides a road transport vehicle driving risk uncertainty prediction method based on deep learning, as shown in Figure 1 , the method comprises the following steps:

[0052] Step 1: Obtain trajectory data and alarm data of road transport vehicles, as well as geographic environment data, social environment data, and road network structure data, preprocess the data, and establish a multi-source data set.

[0053] Step 2: Analyze the spatio-temporal distribution characteristics of the trajectory data and alarm data of road transport vehicles. In terms of spatial distribution characteristics, analyze the clustering and dispersion regions of alarm data in space. Alarm events exhibit significant unevenness and high clustering in some regions. In terms of temporal distribution characteristics, analyze the numerical changes over time at a 2-hour time granularity, as well as the correlation between the trajectory and alarm data trends. Both exhibit significant periodic changes in "days" and a close positive correlation between the trends.

[0054] Step 3: Considering the spatial distribution characteristics, propose a self-adaptive grid division model KDTree_weighted that achieves balanced distribution of alarm numbers among grids. Based on KDTree, add alarm number weights. Process the research area into 1km*1km grids, calculate the alarm numbers in each grid, and set each node of the KDTree_weighted model to correspond to each grid. The weight of each node is the alarm number in the grid. After sorting the nodes by spatial position, use binary search to find a position that makes the weight sums of the left and right subtrees equal or closest to equal as the division hyperplane. Recursively implement this logic to achieve balanced, non-overlapping, and efficient spatial grid division.

[0055] Step 4: Considering the temporal distribution characteristics, construct an input time slice structure composed of three parts: the adjacent time slice, the previous day time slice, and the previous week time slice. The adjacent time slice records n time slices from X t-n to X t (current time), capturing short-term traffic dynamics. The previous day time slice records n time slices from X t-d to X t-d+n , capturing regularity features in daily cycles. The previous week time slice records n time slices from X t-7d to X t-7d+n , further enhancing the model's ability to perceive periodic trends. This effectively assists the model in fully integrating short-term dynamics and long-term periodicity, improving prediction accuracy and spatio-temporal generalization ability for future driving risks.

[0056] Step 5: The impact factor variable system was constructed based on the multi-source data set. By analyzing the correlation between variables and using random forest to construct a simple regression model, the global importance of variables in the entire model was obtained. Finally, the social environment variables (the density of various POIs), the road network structure variables (the density of various grades of roads, bridges and tunnels), and the geographic environment variables (terrain, elevation) were selected as research variables.

[0057] Step 6: Based on the selected research variables, a grid was used as a node to construct three static graphs of social environment similarity, road network structure similarity, and geographic environment similarity using cosine similarity. Based on the latitude and longitude coordinates of the grid center, a grid distance graph was constructed.

[0058] Step 7: A multi-quantile regression model STQuantileNet based on spatio-temporal graph convolution was constructed, which input trajectory and alarm data, as well as four static graphs. The model can realize the prediction of the number of future road transport vehicle alarms and the quantification of uncertainty.

[0059] Construction of the social environment similarity graph adjacency matrix in Step 6. The correlation of the POI vectors of two grids can represent the social environment similarity of the two grids, where the dimension of the POI vector is the number of POI categories, and the value is the number of each category of POI in the grid. The cosine similarity method was used to measure the similarity of two POI vectors. The smaller the angle between the two vectors, the larger the cosine value, and the greater the similarity.

[0060] Construction of the geographic environment similarity graph adjacency matrix in Step 6. Geographic environment is closely related to vehicle driving risk. By constructing a two-dimensional geographic environment vector (terrain, elevation) for each grid, the geographic environment similarity between two grids was calculated using the cosine similarity method, thereby constructing the geographic environment similarity graph.

[0061] Construction of the road network structure similarity graph adjacency matrix in Step 6. Grids with similar road network structures often have similar traffic flow patterns and face similar driving risk probabilities. Based on the density information of four grades of roads (including expressways and trunk roads, urban main roads, urban secondary roads, and urban branch roads), bridge density, and tunnel density within each grid, a six-dimensional road network structure vector was constructed, and the road network similarity graph was constructed using the cosine similarity.

[0062] Construction of the grid distance graph adjacency matrix in Step 6. The grid distance graph depicts the spatial distance relationship between grids and is an important way for GCN graph convolutional neural networks to capture spatial information features. By calculating the distance between the center coordinates of the grids, the grid distance graph was constructed.

[0063] In Step 7, the multi-quantile regression model STQuantileNet based on spatio-temporal graph convolution was constructed. The framework of the model is as follows: Figure 2As shown, it mainly includes four key components: dynamic graph construction layer (AdyGraphBlock), multi-graph fusion module (Graph Fusion), time and space feature extraction module (STBlock), and multi-quantile prediction module (Boosting Quantile). The input data of the model consists of two parts: the number of trajectories and the number of alerts of road transport vehicles, both of which are two-dimensional spatio-temporal sequences. The trajectory data is first input into the dynamic graph construction layer to obtain a dynamic graph and three static graphs: grid distance graph, social environment similarity graph, and road network structure similarity graph Figure 1 The dynamic graph and the three static graphs are input into the multi-graph fusion module and the DGCN spatial feature extraction module, while the alert data sequence is input into the BTCN time feature extraction module. These two modules extract the spatio-temporal features of the data and fuse them, and finally input them into the multi-quantile regression module to realize regression prediction.

[0064] The road transport vehicle driving risk uncertainty prediction model in this embodiment is a deep neural network with various hyperparameters, including learning rate, optimizer, etc. By comparing the model effect evaluation indicators, the optimal hyperparameter value is selected to make the model effect optimal. The learning rate is set to 0.001 by default, and the Adam optimizer and StepLR learning rate scheduler are used. Every 25 epochs, the current learning rate is multiplied by the decay factor 0.85 to control the learning pace in the later training and improve the stability of the model. The training set, validation set, and test set are divided in the ratio of 8:1:1. The input sample size is set to 32 for each training, the training rounds are set to 500 rounds, and the output quantiles are set to [0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95].

[0065] To describe the prediction effect of the proposed road transport vehicle driving risk uncertainty prediction model (STQuantileNet model), four certainty indicators: mean absolute error (MAE), root mean square error (RMSE), goodness of fit (R 2 ), F1 score, and four indicators for evaluating the performance of uncertainty quantification: mean prediction interval width (MPIW), interval coverage probability (PICP), mean quantile error (MQE), and comprehensive evaluation indicator (CWC) are selected to evaluate the model performance. The calculation formula is:

[0066]

[0067] CWC = MPIW × [1 + λ · H (γ - PICP) · (γ - PICP) 2 ]

[0068]

[0069] In the formula, yi represents the true value, represents the predicted value, N is the number of samples, L i , i are the lower bound and upper bound of the prediction interval, respectively, if the true value y i of the i-th sample falls in the prediction interval [L i , i ], then c i = 1, otherwise c i = 0. γ is the confidence level, λ is the penalty coefficient to control the penalty degree when the coverage rate is insufficient, H(·) is the Heaviside function, H = 1 if γ-PICP>0, otherwise H = 0. Q is a set of quantiles (such as {0.1, 0.5, 0.9}), the predicted value of the model at quantile q, ρ q (·) is the quantile loss function.

[0070] Table 1 shows the performance of the STQuantileNet of the embodiment and five comparative models (STZINBNet, STZIGSNet, LSTM, RNN, TCN) in various indicators, wherein the STZINBNet, STZIGSNet and STQuantileNet have the same architecture in the spatio-temporal graph convolution part, and the difference lies in that the STZINBNet replaces the multi-quantile regression module with a zero-inflated negative binomial distribution, and the STZIGSNet replaces the multi-quantile regression module with a zero-inflated Gaussian distribution. The optimal model and the suboptimal model are marked in bold and underlined in the table, and in addition, since LSTM, RNN and TCN are deterministic prediction models, they are not applicable to the three uncertainty indicators of PICP, MPIW and CWC.

[0071] Table 1 Comparison of prediction performance of six models

[0072]

[0073] It can be seen that the STQuantileNet model performs excellently in deterministic indicators, the STZIGSNet and STZINBNet models perform worse than the STQuantileNet, and the remaining models (LSTM, RNN, TCN) perform relatively poorly. In terms of uncertainty indicators, the prediction interval MPIW of the STQuantileNet is the narrowest, although the coverage rate PICP is inferior to the other two models, but the CWC is the lowest, which shows that it can maintain a high coverage rate in a relatively narrow prediction interval, reflecting the superiority of the model in considering uncertainty and accuracy.

[0074] The STQuantileNet model, proposed for predicting road transport vehicle driving risk uncertainty, incorporates a variety of dynamic and static graphs. These include social environment similarity graphs, geographic environment similarity graphs, road network structure similarity graphs, and grid distance graphs; and dynamic graphs, including trajectory dynamic graphs. Through ablation experiments, we analyze the contribution of these five types of spatial relationship graphs to the model's predictive performance. Tables 2 and 3 show the meaning of the ablation model codes and their prediction results and evaluation metrics.

[0075] Table 2 Model codes and their meanings in ablation experiments

[0076]

[0077] Table 3 Ablation experiments

[0078]

[0079] As can be seen from Table 3, the STQuantileNet model proposed in this embodiment, which integrates multiple dynamic and static graphs, performs better than other variant models. This means that the five graphs, namely the social environment similarity graph, the geographical environment similarity graph, the road network structure similarity graph, the grid distance graph, and the trajectory dynamic graph structure, all contribute to the final prediction results of the model, indicating that the various grid space graphs modeled by the present invention are effective in extracting spatial dependencies.

[0080] The deep learning-based road transport vehicle driving risk uncertainty prediction method provided in this embodiment has high reliability. It takes into account the impact of factors such as urban spatial distribution, social environment, geographical environment, road environment, and the complex time-varying dependency relationship between the trajectory and alarms of road transport vehicles on vehicle driving risks, and has higher calculation accuracy. This embodiment provides scientific support and practical guidance for road transport vehicle driving safety management.

[0081] Example 3

[0082] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the above-mentioned method for predicting uncertainty in driving risk of road transport vehicles based on deep learning is implemented. The method includes:

[0083] Obtain the spatiotemporal sequence of the number of alarms, the spatiotemporal sequence of traffic flow, and the multi-graph static adjacency matrix at the current time step;

[0084] The pre-trained road transport vehicle driving risk uncertainty prediction model is used to process the obtained alarm quantity spatio-temporal sequence, traffic flow spatio-temporal sequence and multi-graph static adjacency matrix to obtain the road transport vehicle alarm quantity at the future time step; wherein the road transport vehicle driving risk uncertainty prediction model comprises a dynamic graph construction layer, a multi-graph fusion module, a time and space feature extraction module and a multi-quantile prediction output module; the dynamic graph construction layer is used to extract the dependency relationship between the alarm quantity and the traffic flow, and construct the road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is used to calculate the feature information weight coefficient of each spatial graph structure and perform multi-graph fusion; the time feature extraction module is used to process the alarm quantity spatio-temporal sequence and extract its time dependency relationship; the space feature extraction module is used to process the fused graph and the alarm quantity spatio-temporal sequence and extract its spatial dependency relationship; the multi-quantile prediction output module is used to fuse the feature information of all time steps and spatial graph structures, use the multi-layer perceptron to estimate the median as the central reference, and obtain the quantile by accumulating the non-negative increment or decrement, to obtain the prediction output on multiple quantiles.

[0085] Embodiment 4

[0086] The embodiment 4 provides a computer device comprising a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the deep learning-based road transport vehicle driving risk uncertainty prediction method as described above, which comprises:

[0087] obtaining the alarm quantity spatio-temporal sequence, traffic flow spatio-temporal sequence and multi-graph static adjacency matrix at the current time step;

[0088] The pre-trained road transport vehicle driving risk uncertainty prediction model is used to process the obtained alarm quantity spatio-temporal sequence, traffic flow spatio-temporal sequence and multi-graph static adjacency matrix to obtain the road transport vehicle alarm quantity at the future time step; wherein the road transport vehicle driving risk uncertainty prediction model comprises a dynamic graph construction layer, a multi-graph fusion module, a time and space feature extraction module and a multi-quantile prediction output module; the dynamic graph construction layer is used to extract the dependency relationship between the alarm quantity and the traffic flow, and construct the road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is used to calculate the feature information weight coefficient of each spatial graph structure and perform multi-graph fusion; the time feature extraction module is used to process the alarm quantity spatio-temporal sequence and extract its time dependency relationship; the space feature extraction module is used to process the fused graph and the alarm quantity spatio-temporal sequence and extract its spatial dependency relationship; the multi-quantile prediction output module is used to fuse the feature information of all time steps and spatial graph structures, use the multilayer perceptron to estimate the median as the central reference, and obtain the quantile by accumulating the non-negative increment or decrement to obtain the prediction output on multiple quantiles.

[0089] Embodiment 5

[0090] The embodiment 5 provides an electronic device, comprising: a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device runs, the processor executes the computer program stored in the memory, so that the electronic device executes the instructions for realizing the road transport vehicle driving risk uncertainty prediction method based on deep learning as described above, which comprises:

[0091] obtaining the alarm quantity spatio-temporal sequence, the traffic flow spatio-temporal sequence and the multi-graph static adjacency matrix at the current time step;

[0092] The pre-trained road transport vehicle driving risk uncertainty prediction model is used for processing the obtained alarm quantity space-time sequence, vehicle flow space-time sequence and multi-graph static adjacency matrix to obtain the road transport vehicle alarm quantity at a future time step; wherein, the road transport vehicle driving risk uncertainty prediction model comprises a dynamic graph construction layer, a multi-graph fusion module, a time and space feature extraction module and a multi-quantile prediction output module; the dynamic graph construction layer is used for extracting the dependency relationship between the alarm quantity and the vehicle flow and constructing a road transport vehicle flow dynamic graph adjacency matrix; the multi-graph fusion module is used for calculating the feature information weight coefficients of each space graph structure and performing multi-graph fusion; the time feature extraction module is used for processing the alarm quantity space-time sequence and extracting the time dependency relationship thereof; the space feature extraction module is used for processing the fused graph and the alarm quantity space-time sequence and extracting the space dependency relationship thereof; and the multi-quantile prediction output module is used for fusing the feature information of all time steps and space graph structures, using a multilayer perceptron to estimate the median as a central reference, and obtaining the prediction output on multiple quantiles through cumulative non-negative increments or decrements.

[0093] In summary, the road transport vehicle driving risk uncertainty prediction method based on deep learning provided by the embodiment of the application comprises: based on the real-time vehicle trajectory and alarm data derived by the road transport vehicle on-board monitoring system, using the KDTree_weighted algorithm to perform adaptive grid division of the alarm quantity balance. The space-time distribution characteristics and numerical distribution characteristics are analyzed, four static graphs of grid distance graph, social environment similarity graph, road network structure similarity graph and geographical environment similarity graph are constructed, a dynamic graph construction layer is introduced to construct a trajectory dynamic graph, and spatial correlation characteristics are captured. A prediction model STQuantileNet fusing a space-time graph convolutional network and a multi-quantile regression technology is proposed, the dynamic graph convolutional network DGCN and the time convolutional network BTCN are introduced to realize effective extraction of the multi-space-time characteristics of the traffic flow and the alarm data, the multi-quantile regression model is used to effectively capture the uncertainty of the driving risk, the accuracy of the prediction result and the rationality of the confidence interval are taken into account, and scientific support and practical guidance are provided for the road transport vehicle driving safety management.

[0094] Those skilled in the art will understand that the embodiments of the application can be provided as a method, a system or a computer program product. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flows and / or blocks specified in the block diagrams. Figure 1 one or more flows and / or blocks. Figure 1 one or more flows and / or blocks.

[0096] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart or flows and / or blocks specified in the block diagrams. Figure 1 one or more flows and / or blocks. Figure 1 one or more flows and / or blocks.

[0097] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flows and / or blocks specified in the block diagrams. Figure 1 one or more flows and / or blocks. Figure 1 one or more flows and / or blocks.

[0098] The above description is only a specific implementation of the present application and is not intended to limit the protection scope of the present application. It should be understood by those skilled in the art that various modifications or changes can be made on the basis of the disclosed technical solutions without any inventive effort, and these modifications or changes should be covered within the protection scope of the present application.

Claims

1. A method for predicting uncertainty in driving risk of road transport vehicles, characterized in that: include: Obtain the spatiotemporal sequence of the number of alarms, the spatiotemporal sequence of traffic flow, and the multi-graph static adjacency matrix at the current time step; The pre-trained road transport vehicle travel risk uncertainty prediction model is used to process the acquired spatiotemporal sequence of the number of alarms, the spatiotemporal sequence of vehicle flow, and the multi-graph static adjacency matrix to obtain the number of road transport vehicle alarms at future time steps; wherein the road transport vehicle travel risk uncertainty prediction model includes a dynamic graph construction layer, a multi-graph fusion module, a temporal and spatial feature extraction module, and a multi-quantile prediction output module; the dynamic graph construction layer is used to extract the dependency between the number of alarms and vehicle flow, and construct the road transport vehicle flow dynamic graph adjacency matrix; The multi-graph fusion module is used to calculate the feature information weight coefficients of each spatial graph structure and perform multi-graph fusion; the time feature extraction module is used to process the spatiotemporal sequence of the alarm quantity and extract its temporal dependency; the spatial feature extraction module is used to process the fusion graph and the spatiotemporal sequence of the alarm quantity and extract their spatial dependency; the multi-quantile prediction output module is used to fuse the feature information of all time steps and spatial graph structures, use a multi-layer perceptron to estimate the median as a central reference, and obtain the quantiles by accumulating non-negative increments or decrements to obtain prediction outputs at multiple quantiles.

2. The method for predicting uncertainty of road transport vehicle driving risk according to claim 1, characterized in that: Obtain the terrain and altitude data of the grid, construct the geographical environment vector, use the cosine similarity method to measure the similarity between two vectors, and construct the geographical environment similarity adjacency matrix with the grid as the node.

3. The road transport vehicle driving risk uncertainty prediction model training method according to claim 1 is characterized in that: Obtain the number of various POIs in the grid and construct a POI vector, where the dimension of the POI vector is the number of POI categories and the value is the density of POIs of each category in the grid; use the cosine similarity method to measure the similarity of two POI vectors and construct a social environment similarity adjacency matrix; obtain the data of roads, bridges and tunnels of various levels in the grid, construct a six-dimensional road network structure vector, including road density information, bridge density, and tunnel density, and use cosine similarity to construct a road network structure similarity adjacency matrix; obtain the longitude and latitude coordinates of the center point of each grid, calculate the distance between each grid, and construct a grid distance adjacency matrix.

4. The road transport vehicle driving risk uncertainty prediction model training method according to claim 1 is characterized in that: The dynamic graph construction layer includes: the trajectory spatiotemporal sequence data is first input in parallel to two feature extraction blocks block1 and block2. Each block contains a convolution layer, a batch normalization layer and a ReLU activation function, and the output is two feature vectors Then use batch matrix multiplication, activation function and normalization to get an N×N matrix M flow , to achieve adaptive generation of dynamic graphs of trajectory flow relationships between different grids based on historical trajectory data.

5. The road transport vehicle driving risk uncertainty prediction model training method according to claim 1 is characterized in that: The temporal feature extraction module consists of multiple temporal convolutional network layers, each TCN layer is followed by a batch normalization layer and a dropout layer; The spatial feature extraction module consists of multiple Diffusion Graph Convolutional Network layers, and each D-GCN layer is followed by a Dropout layer.

6. The road transport vehicle driving risk uncertainty prediction model training method according to claim 1 is characterized in that: The area to be predicted is rasterized, and the number of alarms in each grid is calculated. Each node of the adaptive grid partitioning model corresponds to each grid, and its weight is the number of alarms in the grid. After sorting the nodes according to their spatial positions, a binary search is used to find a position that makes the sum of the weights of the left and right subtrees equal or closest to equal as the dividing hyperplane, and the left and right subtrees are divided. This logical recursion eventually achieves a relative balance in the sum of the node weights in each leaf node, realizing balanced, non-overlapping, and efficient spatial grid partitioning.

7. A road transport vehicle driving risk uncertainty prediction system, characterized in that: include: The acquisition module is used to obtain the spatiotemporal sequence of the number of alarms, the spatiotemporal sequence of traffic flow, and the multi-graph static adjacency matrix at the current time step; A processing module is used to process the acquired spatiotemporal sequence of the number of alarms, the spatiotemporal sequence of vehicle flow, and the multi-graph static adjacency matrix using a pre-trained road transport vehicle travel risk uncertainty prediction model to obtain the number of road transport vehicle alarms at future time steps; wherein the road transport vehicle travel risk uncertainty prediction model includes a dynamic graph construction layer, a multi-graph fusion module, a temporal and spatial feature extraction module, and a multi-quantile prediction output module; the dynamic graph construction layer is used to extract the dependency between the number of alarms and vehicle flow and construct the road transport vehicle flow dynamic graph adjacency matrix; The multi-graph fusion module is used to calculate the feature information weight coefficients of each spatial graph structure and perform multi-graph fusion; the time feature extraction module is used to process the spatiotemporal sequence of the alarm quantity and extract its temporal dependency; the spatial feature extraction module is used to process the fusion graph and the spatiotemporal sequence of the alarm quantity and extract their spatial dependency; the multi-quantile prediction output module is used to fuse the feature information of all time steps and spatial graph structures, use a multi-layer perceptron to estimate the median as a central reference, and obtain the quantiles by accumulating non-negative increments or decrements to obtain prediction outputs at multiple quantiles.

8. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the road transport vehicle driving risk uncertainty prediction method as described in any one of claims 1-6 is implemented.

9. A computer device, characterized in that: It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the road transport vehicle driving risk uncertainty prediction method as described in any one of claims 1-6.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the method for predicting the uncertainty of road transport vehicle driving risks as described in any one of claims 1 to 6.

Images