Multi-index dynamic fusion traffic hidden danger identification and completion method and system

Abstract

The invention discloses a multi-index dynamic fusion traffic hidden danger identification and completion method and system, and relates to the technical field of intelligent traffic systems, and the method comprises the steps: obtaining a multi-dimensional traffic data sequence containing missing values in a target monitoring region, the multi-dimensional traffic data sequence comprises a traffic flow index sequence, a driving behavior characteristic sequence and a road environment parameter sequence; and interpolating missing values in the multi-dimensional traffic data sequence based on a spatio-temporal dynamic graph completion network to generate a completed spatio-temporal data tensor. According to the method, the missing data are accurately completed through a space-time dynamic graph completion network, a space graph convolution layer of road network topology, a gating circulation unit of a modeling time sequence rule and a cross-index attention module, multi-index dynamic association is fully fused, and propagation delay and a space-time evolution rule are considered; and deep fusion features are extracted by using a shared feature encoder, hidden danger identification branches are linked, traffic flow, driving behaviors and environmental parameters are dynamically fused, and hidden danger dynamic formation paths are captured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation systems, in particular to a multi-index dynamic fusion traffic hazard identification and completion method and system. BACKGROUND

[0002] With the rapid development of intelligent transportation systems, the traffic monitoring system of highways and urban roads is increasingly improved. Through multi-source sensing devices, multi-dimensional data including traffic flow indicators, driving behavior characteristics, and environmental parameters can be collected, providing a data basis for accurate identification of traffic hazards.

[0003] However, in practical applications, the effective identification of traffic hazards still faces many technical challenges, mainly in the following aspects:

[0004] 1. Limitations of multi-index data missing and completion: Due to factors such as uneven distribution of equipment, signal interference, and faults, traffic monitoring data often has missing multi-index problems. Existing data completion methods mostly focus on single indicators or static spatial correlation modeling, without fully integrating the dynamic correlation between multi-indexes, and insufficient consideration of traffic flow propagation delay and spatio-temporal evolution rules, resulting in deviations between completed data and real traffic conditions, making it difficult to support subsequent hazard identification.

[0005] 2. Insufficient multi-index fusion in hazard identification: Traditional traffic hazard identification models mostly rely on single type data, lacking dynamic fusion capability for multi-indexes. For example, some models only use static traffic flow parameters such as average speed and congestion index to judge risks, ignoring the synergistic effects of driving behavior sequences and environmental factors; other models attempt to integrate multi-source data, but do not consider the time evolution rules of different indicators, such as the gradual change process before the occurrence of hazards, making it difficult to capture the dynamic formation path of hazards.

[0006] 3. Disconnection between completion and identification: In existing technologies, data completion and hazard identification are often treated as independent processes. The completion process is not optimized for hazard identification needs, such as not preserving indicator features sensitive to hazards, while the identification model does not consider the uncertainty of data missing, leading to cumulative completion errors affecting identification accuracy, making it difficult to meet the needs of hazard early warning in complex traffic environments.

[0007] Therefore, how to accurately complete the missing multi-index data and based on the dynamic fusion information of the completed multi-indexes, to build a traffic hazard identification mechanism considering spatio-temporal correlation and dynamic evolution rules, has become a key technical problem to improve the risk control capability of intelligent transportation systems. SUMMARY

[0008] In view of the deficiencies of the prior art, the application provides a multi-index dynamic fusion traffic hazard identification and completion method and system, which solves the problems of multi-index data missing and completion limitations in the prior art, insufficient multi-index fusion in hazard identification, and disconnection between completion and identification processes.

[0009] To achieve the above object, the application is implemented by the following technical scheme: a multi-index dynamic fusion traffic hazard identification and completion method, comprising:

[0010] Obtaining a multi-dimensional traffic data sequence containing missing values in a target monitoring area, wherein the multi-dimensional traffic data sequence includes a traffic flow index sequence, a driving behavior feature sequence, and a road environment parameter sequence;

[0011] Based on a spatio-temporal dynamic graph completion network, the missing values in the multi-dimensional traffic data sequence are interpolated to generate a completed spatio-temporal data tensor; wherein the spatio-temporal dynamic graph completion network establishes a spatial graph convolution layer according to the road network topology structure to capture the spatial correlation of traffic indicators, and couples a gating recurrent unit suitable for processing time series data to model the time evolution law and propagation delay between multi-indices;

[0012] The completed spatio-temporal data tensor is input into a joint optimization model composed of a shared feature encoder, a hazard identification branch, and a data reconstruction branch;

[0013] The shared feature encoder extracts a deep fusion feature vector of the completed spatio-temporal data tensor;

[0014] The deep fusion feature vector is input into the hazard identification branch to generate a traffic hazard risk level of the target monitoring area;

[0015] The deep fusion feature vector is input into the data reconstruction branch to generate a reconstructed spatio-temporal data tensor;

[0016] The identification loss between the traffic hazard risk level output by the hazard identification branch and the preset hazard label, and the reconstruction loss between the reconstructed spatio-temporal data tensor and the completed spatio-temporal data tensor are calculated, a joint loss function is constructed based on the weighted sum of the identification loss and the reconstruction loss, and the joint loss function is back-propagated to synchronously update the model parameters of the spatio-temporal dynamic graph completion network, the shared feature encoder, the hazard identification branch, and the data reconstruction branch.

[0017] Further, the multi-dimensional traffic data sequence containing missing values in the target monitoring area comprises:

[0018] obtain a traffic flow indicator sequence associated with the target monitoring area from a traffic management data platform, the traffic flow indicator sequence including average vehicle speed, lane occupancy rate, and traffic flow;

[0019] obtain a driving behavior feature sequence from a vehicle terminal or a roadside unit, the driving behavior feature sequence including the number of sudden accelerations, the number of sudden decelerations, the frequency of abnormal lane changes, and the standard deviation of following distance;

[0020] obtain a road environment parameter sequence from a meteorological service system or an environmental sensor, the road environment parameter sequence including visibility, rainfall intensity, road wetness coefficient, and crosswind speed.

[0021] Further, the spatio-temporal dynamic graph completion network further comprises a cross-indicator attention module located between the spatial graph convolution layer and the gated recurrent unit, for calculating dynamic correlation scores between the traffic flow indicator sequence, the driving behavior feature sequence, and the road environment parameter sequence, and assigning fusion weights to features of different indicators according to the dynamic correlation scores, to enhance the representation of sensitive indicators of hidden dangers.

[0022] Further, the spatio-temporal dynamic graph completion network supplements missing values in the multi-dimensional traffic data sequence to generate a completed spatio-temporal data tensor, comprising:

[0023] align the timestamps of each indicator in the multi-dimensional traffic data sequence, and perform mask processing on the missing values to generate an input data matrix and a corresponding missing mask matrix;

[0024] input the input data matrix and an adjacency matrix defined by the road network topology structure into the spatial graph convolution layer to generate spatial fusion features;

[0025] input the spatial fusion features into the cross-indicator attention module to generate weighted multi-indicator features;

[0026] input the weighted multi-indicator features into the gated recurrent unit to capture temporal dependencies, and output an initial completed data tensor containing interpolated values;

[0027] perform element-wise multiplication of the initial completed data tensor and the missing mask matrix to retain only the interpolated values at the original missing positions, and generate a final completed spatio-temporal data tensor.

[0028] Further, the shared feature encoder is a stacked self-attention layer and a feedforward neural network layer, which are used for deep nonlinear transformation of the completed spatio-temporal data tensor to capture complex high-order feature patterns before hidden danger occurs.

[0029] Further, the hidden danger identification branch is a fully connected classification network, which receives the deep fusion feature vector and outputs a probability distribution corresponding to four preset levels of "no hidden danger", "low-risk hidden danger", "medium-risk hidden danger" and "high-risk hidden danger" through a Softmax activation function, and takes the level with the highest probability as the traffic hidden danger risk level.

[0030] Further, the data reconstruction branch is a decoder network symmetrical to the shared feature encoder structure, which receives the deep fusion feature vector and up-samples it to the same dimension as the completed spatio-temporal data tensor to generate the reconstructed spatio-temporal data tensor.

[0031] Further, the joint loss function is constructed based on the weighted sum of the discrimination loss and the reconstruction loss, including:

[0032] The discrimination loss is calculated using a cross-entropy loss function;

[0033] The reconstruction loss is calculated using a mean square error function;

[0034] The first weight coefficient and the second weight coefficient are set, the discrimination loss is multiplied by the first weight coefficient, the reconstruction loss is multiplied by the second weight coefficient, and then the sum is taken to obtain the joint loss function;

[0035] Wherein, at the initial stage of model training, the second weight coefficient is increased to prioritize the accuracy of data completion, and at the later stage of model training, the first weight coefficient is increased to focus on improving the accuracy of hidden danger identification.

[0036] Further, the method further includes:

[0037] When the output traffic hidden danger risk level is "medium-risk hidden danger" or "high-risk hidden danger", a pre-trained attribution analysis model is called;

[0038] The attribution analysis model analyzes the deep fusion feature vector and calculates the contribution score of the traffic flow indicator sequence, the driving behavior feature sequence and the road environment parameter sequence to the current hidden danger risk level;

[0039] According to the contribution score, the dominant factor causing the hidden danger is determined, and the traffic hidden danger risk level and the dominant factor are pushed to the traffic control center.

[0040] The application further provides a traffic hidden danger identification and completion system based on multi-index dynamic fusion, comprising a processor and a memory, the memory is connected with the processor, the memory is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the memory to realize the multi-index dynamic fusion traffic hidden danger identification and completion method in any of the above.

[0041] Advantages

[0042] The application is based on a spatiotemporal dynamic graph completion network, combines a spatial graph convolution layer of road network topology, a gating recurrent unit for modeling time sequence rules and a cross-index attention module, accurately completes missing data, fully fuses multi-index dynamic correlations and considers propagation delay and spatiotemporal evolution rules, extracts deep fusion features by using a shared feature encoder, links a hidden danger identification branch, dynamically fuses traffic flow, driving behavior and environmental parameters, captures a hidden danger dynamic formation path, synchronously updates the completion network and identification model parameters by using a joint loss function, forms a closed loop optimization, avoids completion error accumulation, and significantly improves the accuracy and reliability of hidden danger identification in a complex traffic environment. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The application provides a method flowchart. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0045] Please refer to Figure 1 The application provides a multi-index dynamic fusion traffic hidden danger identification and completion method, comprising:

[0046] Obtaining a multi-dimensional traffic data sequence containing missing values in a target monitoring area, the multi-dimensional traffic data sequence comprising a traffic flow index sequence, a driving behavior feature sequence and a road environment parameter sequence;

[0047] Based on a spatiotemporal dynamic graph completion network, the missing values in the multi-dimensional traffic data sequence are interpolated to generate a completed spatiotemporal data tensor; wherein the spatiotemporal dynamic graph completion network establishes a spatial graph convolution layer according to a road network topology structure, to capture the correlation of traffic indexes in space, and couples a gating recurrent unit suitable for processing time sequence data, to model the time evolution rules and propagation delay among multi-indexes;

[0048] The completed spatiotemporal data tensor is input into a joint optimization model composed of a shared feature encoder, a hazard identification branch and a data reconstruction branch;

[0049] Deep fusion feature vectors of the completed spatiotemporal data tensor are extracted by the shared feature encoder;

[0050] The deep fusion feature vectors are input into the hazard identification branch to generate a traffic hazard risk level of the target monitoring area;

[0051] The deep fusion feature vectors are input into the data reconstruction branch to generate a reconstructed spatiotemporal data tensor;

[0052] An identification loss between the traffic hazard risk level output by the hazard identification branch and a preset hazard label, and a reconstruction loss between the reconstructed spatiotemporal data tensor and the completed spatiotemporal data tensor are calculated, a joint loss function is constructed based on a weighted sum of the identification loss and the reconstruction loss, and the joint loss function is back propagated to synchronously update model parameters of the spatiotemporal dynamic graph completion network, the shared feature encoder, the hazard identification branch and the data reconstruction branch.

[0053] The multi-dimensional traffic data sequence refers to a time series data set containing traffic flow indicators, driving behavior characteristics and road environment parameters, which can be collected by a sensor network, a vehicle terminal and a meteorological system to comprehensively cover multi-source heterogeneous information of traffic states.

[0054] The spatiotemporal dynamic graph completion network refers to a hybrid model combining graph convolution and recurrent neural network, which specifically uses a spatial graph convolution layer to capture road network topological correlation and a gated recurrent unit to model time sequence evolution rules to solve the problem of spatiotemporal correlation modeling under data missing.

[0055] The shared feature encoder refers to a feature extraction module in a multi-task learning framework, which is implemented by using a self-attention mechanism and a feedforward neural network to extract deep fusion features related to hazards from the completed data.

[0056] The hazard identification branch refers to a prediction module based on a classification network, which is implemented by using a fully connected layer and a Softmax function to map the fusion features to a risk level probability distribution.

[0057] The data reconstruction branch refers to a decoder structure symmetrical to the encoder, which specifically uses an up-sampling layer to restore data dimensions to verify the effectiveness of feature extraction and constrain the optimization direction of the model.

[0058] The joint loss function refers to a weighted objective function of identification and reconstruction tasks, which is implemented by combining cross-entropy and mean square error to balance the collaborative optimization of completion accuracy and identification performance.

[0059] The core innovation of the application is to construct a multi-index dynamic fusion end-to-end joint optimization framework, to realize high-precision interpolation of missing data through a space-time dynamic graph completion network, and to utilize a shared feature encoder to synchronously drive hidden danger identification and data reconstruction tasks, forming a closed-loop optimization mechanism for completion and identification, effectively solving the error accumulation problem caused by the separation of data missing processing and hidden danger identification in traditional methods, and significantly improving the risk warning reliability in complex traffic scenarios.

[0060] The working process and principle of the application are as follows: first, a multi-dimensional traffic data sequence with missing values in a target monitoring area is obtained, including a traffic flow index sequence, a driving behavior feature sequence, and a road environment parameter sequence. Then, the missing values in the multi-dimensional traffic data sequence are interpolated based on a space-time dynamic graph completion network to generate a completed space-time data tensor. The space-time dynamic graph completion network establishes a spatial graph convolution layer according to the road network topology structure, captures the spatial correlation of traffic indicators, and couples a gated recurrent unit to model the time evolution law and propagation delay among multiple indicators.

[0061] Next, the completed space-time data tensor is input into a joint optimization model composed of a shared feature encoder, a hidden danger identification branch, and a data reconstruction branch. The shared feature encoder extracts a deep fusion feature vector of the completed space-time data tensor. The deep fusion feature vector is input into the hidden danger identification branch and the data reconstruction branch respectively to generate a traffic hidden danger risk level of the target monitoring area and a reconstructed space-time data tensor.

[0062] Finally, the discrimination loss between the traffic hidden danger risk level output by the hidden danger identification branch and the preset hidden danger label is calculated, as well as the reconstruction loss between the reconstructed space-time data tensor and the completed space-time data tensor. A joint loss function is constructed based on the weighted sum of the discrimination loss and the reconstruction loss, and the joint loss function is back-propagated to synchronously update the model parameters of the space-time dynamic graph completion network, the shared feature encoder, the hidden danger identification branch, and the data reconstruction branch.

[0063] Through the space-time dynamic graph completion network, the spatial correlation and time evolution law are considered simultaneously, which can more accurately complete the missing data. The joint optimization model combines data completion and hidden danger identification tasks, and realizes synchronous parameter update of the two tasks through the shared feature encoder and the joint loss function, reducing error accumulation. This method can effectively capture the dynamic interaction among multiple indicators and improve the accuracy of hidden danger identification.

[0064] As a preferred embodiment, the scheme of the application is implemented as follows:

[0065] Firstly, average speed, lane occupancy and traffic flow data of the target monitoring area are obtained from the traffic management data platform as traffic flow index sequences. The number of sudden accelerations, the number of sudden decelerations, the frequency of abnormal lane changes and the standard deviation of following time interval data are obtained from the vehicle terminal and the roadside unit as driving behavior feature sequences. Visibility, rainfall intensity, road wetness coefficient and crosswind speed data are obtained from the meteorological service system and environmental sensors as road environment parameter sequences.

[0066] Then, a spatio-temporal dynamic graph completion network is constructed. The network includes a spatial graph convolution layer, a cross-index attention module and a gated recurrent unit. The spatial graph convolution layer performs spatial feature fusion based on the adjacency matrix defined by the road network topology. The cross-index attention module calculates the dynamic correlation score between different index sequences and assigns a fusion weight. The gated recurrent unit captures the time dependence.

[0067] The multi-dimensional traffic data sequences are input into the spatio-temporal dynamic graph completion network to generate a completed spatio-temporal data tensor. Then a joint optimization model is constructed, including a shared feature encoder, a hidden danger identification branch and a data reconstruction branch. The shared feature encoder is composed of stacked self-attention layers and feedforward neural network layers, which performs nonlinear transformation on the completed spatio-temporal data tensor and extracts a deep fusion feature vector.

[0068] The hidden danger identification branch is a fully connected classification network, which outputs a traffic hidden danger risk level probability distribution through a Softmax function. The data reconstruction branch is a decoder network that up-samples the deep fusion feature vector to the original dimension.

[0069] Finally, the cross-entropy loss is calculated as the discrimination loss, and the mean square error is calculated as the reconstruction loss. A weight coefficient is set to construct a joint loss function. The reconstruction loss weight is increased in the early stage of training, and the discrimination loss weight is increased in the later stage. The parameters of each module are updated through back propagation to realize the joint optimization of data completion and hidden danger identification.

[0070] The present application further proposes a specific implementation method for obtaining multi-dimensional traffic data sequences with missing values in the target monitoring area, including obtaining traffic flow index sequences from the traffic management data platform, obtaining driving behavior feature sequences from vehicle terminals or roadside units, and obtaining road environment parameter sequences from meteorological service systems or environmental sensors.

[0071] Among them, the traffic flow index sequence includes average speed, lane occupancy and traffic flow, the driving behavior feature sequence includes the number of sudden accelerations, the number of sudden decelerations, the frequency of abnormal lane changes and the standard deviation of following time interval, and the road environment parameter sequence includes visibility, rainfall intensity, road wetness coefficient and crosswind speed. These indicators are obtained in real time by different data collection devices, for example, lane occupancy is measured by inductance coils buried in the road, the number of sudden accelerations is recorded by vehicle acceleration sensors, and the road wetness coefficient is monitored by infrared road sensors.

[0072] Specifically, the traffic management data platform periodically collects vehicle traffic data through flow detectors deployed along the road to generate traffic flow indicator sequences containing average vehicle speed. The vehicle terminal records the longitudinal acceleration changes of the vehicle at a millisecond-level sampling frequency, and counts the number of sudden accelerations and sudden decelerations within a unit time. The roadside unit monitors the lane-changing trajectory of the vehicle through the millimeter wave radar and calculates the abnormal lane-changing frequency. The meteorological service system collects real-time visibility and rainfall intensity data through the visibility meter and rain gauge, and the environmental sensor obtains the road wet slip coefficient through the friction coefficient detection device installed on the road. Each data source is aligned with a time stamp as an index to form a multi-dimensional traffic data sequence with a unified time reference, providing structured input data for subsequent network dynamic graph completion.

[0073] As a preferred embodiment, the scheme of the application is implemented as follows:

[0074] When obtaining the multi-dimensional traffic data sequence with missing values in the target monitoring area, the traffic flow indicator sequence associated with the target monitoring area is obtained from the traffic management data platform. The traffic flow indicator sequence includes average vehicle speed, lane occupancy rate, and traffic flow. The driving behavior feature sequence is obtained from the vehicle terminal or roadside unit. The driving behavior feature sequence includes the number of sudden accelerations, the number of sudden decelerations, the abnormal lane-changing frequency, and the standard deviation of the following distance. The road environment parameter sequence is obtained from the meteorological service system or the environmental sensor. The road environment parameter sequence includes visibility, rainfall intensity, road wet slip coefficient, and crosswind speed.

[0075] Specifically, the traffic management data platform collects traffic flow indicators through devices such as loop detectors and video detectors distributed in the road network. The vehicle terminal can be an OBD device or a smartphone APP installed on the vehicle, which collects driving behavior data through acceleration sensors, gyroscopes, etc. The roadside unit is a smart sensing device installed on both sides of the road, which can collect vehicle trajectories through radar, cameras, etc. The meteorological service system provides regional weather forecast data, and the environmental sensor includes visibility meters, rain gauges, road condition detectors, etc., for collecting local fine-grained environmental parameters.

[0076] The application further proposes a cross-indicator attention module located between the spatial graph convolution layer and the gated recurrent unit, which is used to calculate the dynamic correlation scores between the traffic flow indicator sequence, the driving behavior feature sequence, and the road environment parameter sequence, and to assign fusion weights to different indicators according to the dynamic correlation scores to enhance the representation of sensitive indicators of hidden hazards.

[0077] The cross-index attention module adopts a multi-head attention mechanism, and the input of each attention head is the index feature matrix processed by the spatial graph convolution. The dynamic correlation score is calculated by the dot product of the query vector and the key vector. The query vector is generated from the index features at the current time step, and the key vector is generated from the other index features in the historical time window. The fusion weight is normalized by the Softmax function to make the sum of the weights of each index feature equal to 1. In specific implementation, the number of attention heads is set to 4, and the hidden layer dimension is set to 64 to balance the calculation efficiency and feature expression ability.

[0078] Specifically, the working process of the cross-index attention module includes the following steps: first, the traffic flow index feature matrix, the driving behavior feature matrix and the road environment parameter matrix output by the spatial graph convolution layer are input respectively. Second, matrix multiplication is performed on each two index matrices to generate a dynamic correlation score matrix between each two, and each element of the matrix represents the correlation strength between the two indexes at a specific time step. Third, the attention weight of each index is calculated according to the dynamic correlation score matrix, and a fused multi-index feature vector is generated by weighted summation. Finally, the weighted feature vector is input into the gated recurrent unit for time series modeling. For example, in the rain weather scenario, the module may assign higher fusion weights to the road surface wetness coefficient and the number of sudden decelerations, so as to strengthen the feature expression of these two indexes when completing the data. By dynamically adjusting the fusion weights of different indexes, the collaborative change pattern between multiple indexes before the occurrence of hidden dangers can be effectively captured, and the reliability of the completed data and the sensitivity of hidden danger identification can be improved.

[0079] As a preferred embodiment, the scheme of the application is implemented as follows:

[0080] The spatio-temporal dynamic graph completion network further includes a cross-index attention module located between the spatial graph convolution layer and the gated recurrent unit. Specifically, the cross-index attention module first receives the spatial fusion features output by the spatial graph convolution layer. Then, for each pair of indexes in the traffic flow index sequence, the driving behavior feature sequence and the road environment parameter sequence, the dynamic correlation score between them is calculated. This dynamic correlation score is obtained by calculating the dot product of the two index sequences and performing softmax normalization. Next, according to the calculated dynamic correlation score, the fusion weight of the feature of different indexes is assigned. Specifically, the dynamic correlation score is used as a weight coefficient to perform weighted summation on the spatial fusion features to generate weighted multi-index features. This weighted operation enhances the representation of sensitive indexes to hidden dangers, enabling the model to better capture the mutual influence between different indexes. Finally, the weighted multi-index features are input into the gated recurrent unit for further modeling of time series dependency.

[0081] The application further proposes to fill in missing values in multi-dimensional traffic data sequences based on a space-time dynamic graph completion network, generate a completed space-time data tensor, comprising:

[0082] Align the timestamps of each index in the multi-dimensional traffic data sequence, and perform mask processing on the missing values to generate an input data matrix and a corresponding missing mask matrix;

[0083] Input the input data matrix and the adjacency matrix defined by the road network topology structure into the spatial graph convolution layer to generate spatial fusion features;

[0084] Input the spatial fusion features into the cross-index attention module to generate weighted multi-index features; input the weighted multi-index features into the gated recurrent unit to capture the time sequence dependency, and output an initial completion data tensor containing the interpolated values;

[0085] Element-wise multiplication of the initial completion data tensor and the missing mask matrix, only retaining the interpolated values at the original missing positions, to generate the final completed space-time data tensor.

[0086] Wherein, the time stamp alignment unifies the time resolution of each index through interpolation or resampling method, to ensure the time consistency of subsequent processing; the mask processing uses a binary matrix to mark the missing value position, and the missing value in the input data matrix is filled with zeros. The spatial graph convolution layer uses the node connection relationship defined by the adjacency matrix to aggregate the traffic index information of adjacent nodes through graph convolution operation, to capture the spatial correlation. The cross-index attention module calculates the dynamic correlation score between the traffic flow index, driving behavior feature and road environment parameter, to generate weighted multi-index features, to enhance the representation of sensitive indicators of hidden dangers. The gated recurrent unit models the time evolution law and propagation delay of multi-index through update gate and reset gate mechanism, to output an initial completion data tensor containing the interpolated values. The element-wise multiplication operation restores the data at the original non-missing positions in the initial completion data tensor, only retains the interpolated values at the missing positions, to avoid covering the existing data.

[0087] Specifically, in the timestamp alignment step, if the time interval of the traffic flow indicator sequence is 5 minutes and the time interval of the driving behavior feature sequence is 10 seconds, the linear interpolation method is used to resample the driving behavior feature sequence to 5-minute intervals to ensure that the time dimensions of each indicator are consistent. In the mask matrix, the original data position is marked as 1, and the missing position is marked as 0. The spatial graph convolution layer uses the Chebyshev polynomial to approximate the graph convolution kernel, and the adjacency matrix is constructed according to the actual connection relationship of the nodes in the road network topology. The connection edge is established when the distance between nodes is less than 500 meters. In the cross-indicator attention module, the dynamic correlation score is calculated by the query-key value attention mechanism, and the fusion weight of the traffic flow indicator and the driving behavior feature is set to 0.6, and the fusion weight of the road environment parameter is set to 0.3. The hidden layer dimension of the gated recurrent unit is set to 64, and the time window length is set to 12 time steps. After the element-wise multiplication operation, the original non-missing position data in the completed spatio-temporal data tensor remains the original value, and the missing position is filled with the interpolated value, and finally a tensor with dimensions [N x T x C] is generated, where N is the number of road network nodes, T is the number of time steps, and C is the number of indicators.

[0088] As a preferred embodiment, the scheme of the application is implemented as follows:

[0089] The timestamps of each indicator in the multi-dimensional traffic data sequence are aligned, the missing values are masked, and an input data matrix and a corresponding missing mask matrix are generated. The input data matrix and the adjacency matrix defined by the road network topology structure are input into the spatial graph convolution layer to generate spatial fusion features. The spatial fusion features are input into the cross-indicator attention module to generate weighted multi-indicator features. The weighted multi-indicator features are input into the gated recurrent unit to capture the time sequence dependency, and an initial completed data tensor containing interpolated values is output. The initial completed data tensor and the missing mask matrix are multiplied element by element to retain only the interpolated values for the original missing positions, and a final completed spatio-temporal data tensor is generated.

[0090] Specifically, first, the traffic flow indicators, driving behavior features, and road environment parameters in the multi-dimensional traffic data sequence are aligned by timestamp to form a unified time series. For data points with missing values, 0 is filled and a corresponding missing mask matrix is generated, with 1 indicating valid data and 0 indicating missing. Then, the aligned data matrix and the adjacency matrix of the road network topology structure are input into the spatial graph convolution layer. The spatial graph convolution layer aggregates the features of adjacent road segments through the message passing mechanism to generate fusion features considering spatial correlation.

[0091] Further, the fused features are input into a cross-index attention module. This module calculates the dynamic correlation scores between different indices and assigns fusion weights to each index accordingly. For example, in the case of congestion, the correlation between vehicle speed and lane occupancy may be enhanced. The attention mechanism can adaptively adjust the importance of each index, highlighting the features that best explain the current traffic state.

[0092] Thus, the weighted multi-index features are sent to a gated recurrent unit. Through its internal gating mechanism, the gated recurrent unit can selectively retain or forget historical information, effectively modeling long-term temporal dependencies. This enables the model to capture the dynamic change patterns of traffic flow, such as the periodic patterns of morning and evening rush hours. The output of the gated recurrent unit is a data tensor that contains preliminary estimates of missing values.

[0093] Finally, this preliminary completed data tensor is element-wise multiplied with the original missing mask matrix. This step ensures that only the originally missing positions are filled with interpolated values, while the original valid data remains unchanged. The completed spatio-temporal data tensor obtained in this way not only preserves the authenticity of the original data, but also fills in the missing values, providing a complete data foundation for subsequent hazard identification.

[0094] The present application further proposes that the shared feature encoder is a stacked self-attention layer and a feedforward neural network layer, which are used to perform deep nonlinear transformation on the completed spatio-temporal data tensor to capture complex high-order feature patterns before the hazard occurs.

[0095] Among them, the self-attention layer dynamically aggregates global information by calculating the correlation weights between feature vectors at different spatio-temporal positions in the completed spatio-temporal data tensor; the feedforward neural network layer consists of two fully connected layers with a nonlinear activation function added in between, which is used to perform nonlinear mapping in the spatial dimension on the output of the self-attention layer. The stacking operation refers to the arrangement of the self-attention layer and the feedforward neural network layer in multiple layers, with the feature vectors output by each layer serving as the input to the next layer. The number of heads of the self-attention layer is set to 8, and the hidden layer dimension of the feedforward neural network layer is set to 4 times the input dimension, with residual connection and layer normalization operations to prevent gradient vanishing.

[0096] Specifically, the completed spatio-temporal data tensor is first input into a self-attention layer to calculate the attention weights between different spatial nodes at different time steps and generate a spatio-temporal feature matrix with global dependency; the matrix is subjected to a position-by-position nonlinear transformation by a feedforward neural network layer to enhance the feature expression capability. The stacked structure gradually extracts the cross-spatio-temporal nonlinear correlations between the traffic flow indicators, driving behavior features and road environment parameters through the combination of multiple layers of self-attention and feedforward neural networks, such as the collaborative change pattern of the number of sudden decelerations before the occurrence of hidden dangers and the road surface wetness coefficient. The final output deep fusion feature vector contains high-order dynamic features in the hidden danger formation process, so that the hidden danger identification branch can make classification decisions based on richer feature information, thereby improving the accuracy of risk level determination.

[0097] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0098] The shared feature encoder is composed of multiple layers of self-attention layers and feedforward neural network layers. The self-attention layer adopts a multi-head attention mechanism, with 8 heads and a hidden layer dimension of 512. The feedforward neural network layer is composed of two fully connected layers, with 2048 neurons in the first layer and 512 neurons in the second layer, and the activation function is ReLU. The number of stacked layers is set to 6 layers.

[0099] The input completed spatio-temporal data tensor is first subjected to position encoding to add spatio-temporal position information. Then it is sequentially subjected to the alternating stacking of 6 layers of self-attention layers and feedforward neural network layers. The attention weight matrix calculated by each layer of self-attention reflects the correlation strength between different spatio-temporal position features. The feedforward neural network layer performs nonlinear transformation to enhance the expression capability of the model.

[0100] The features output by the last layer are subjected to global average pooling to obtain a deep fusion feature vector with fixed dimensions. The feature vector integrates the spatio-temporal correlations of multi-dimensional traffic data and captures complex high-order feature patterns before the occurrence of hidden dangers.

[0101] The present application further proposes that the hidden danger identification branch is a fully connected classification network, which receives the deep fusion feature vector and outputs a probability distribution corresponding to the four preset levels of "no hidden danger", "low-risk hidden danger", "medium-risk hidden danger" and "high-risk hidden danger" through a Softmax activation function, and takes the level with the highest probability as the traffic hidden danger risk level.

[0102] The full connection classification network comprises at least two hidden layers, and the number of neurons of each layer is sequentially reduced to gradually compress the feature dimension; the Softmax activation function normalizes the four-dimensional output of the last full connection layer to generate probability values of each hidden danger level; and a temperature coefficient is introduced in the probability distribution calculation process to adjust the sharpness of the probability distribution, so that the classification result jump caused by slight fluctuations in the features is avoided.

[0103] Specifically, the deep fusion feature vector is mapped to a low-dimensional hidden danger semantic space through multi-layer nonlinear transformation of the full connection classification network, the hidden layer uses the ReLU activation function to enhance the nonlinear expression ability, and the last full connection layer outputs a four-dimensional vector corresponding to the logistic regression values of the four hidden danger levels. The Softmax function converts the logistic values into probability distribution through exponential operation and normalization operation, wherein the temperature coefficient is set as a trainable parameter, the initial value is 1.0 and is dynamically adjusted during the training process to balance the classification sensitivity and stability. After the probability distribution is generated, the hidden danger level corresponding to the maximum probability value is selected as the final output, and the complete probability distribution is reserved for subsequent attribution analysis model. The design ensures the explicit association between the hidden danger level division result and the multi-index dynamic features through end-to-end probability modeling, and improves the classification confidence and interpretability.

[0104] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0105] The hidden danger identification branch is a full connection classification network. The full connection classification network receives the deep fusion feature vector and outputs a probability distribution corresponding to the four preset levels of "no hidden danger", "low-risk hidden danger", "medium-risk hidden danger" and "high-risk hidden danger" through a Softmax activation function, and the level with the highest probability is taken as the traffic hidden danger risk level.

[0106] Specifically, the full connection classification network comprises three full connection layers. The first layer receives a 256-dimensional deep fusion feature vector as input, performs nonlinear transformation through 1024 neurons, and outputs a 1024-dimensional intermediate feature. The second layer maps the 1024-dimensional intermediate feature to 512 dimensions. The third layer finally maps the 512-dimensional feature to a 4-dimensional output corresponding to the four preset hidden danger levels. The ReLU activation function is used between the first layer and the second layer to introduce nonlinearity, and the Softmax function is connected after the third layer to normalize the output into a probability distribution.

[0107] Thus, by the multi-layer nonlinear transformation, the full connection classification network can fully utilize the multi-dimensional traffic data information contained in the deep fusion feature, and learn the complex feature mode of different hidden danger grades. The application of the Softmax function ensures that the output satisfies the probability distribution property, facilitating intuitive understanding of the relative risk degree of each grade. Finally, the grade with the highest probability is selected as the traffic hidden danger risk grade, providing a quantitative risk assessment result for traffic management decision-making.

[0108] The data reconstruction branch is a decoder network symmetrical to the shared feature encoder structure. The decoder network receives the deep fusion feature vector and up-samples it to the same dimension as the completed spatio-temporal data tensor, generating a reconstructed spatio-temporal data tensor.

[0109] The decoder network includes the same number of deconvolution layers and reverse self-attention layers as the encoder. The output channel number of each deconvolution layer is consistent with the input channel number of the corresponding convolution layer of the encoder. The weight matrix dimension of the reverse self-attention layer is symmetrical to that of the encoder self-attention layer, ensuring that the feature vector dimension is reversibly transformed in the encoding and decoding process. The up-sampling operation uses a bilinear interpolation algorithm, and the scaling factor is the inverse of the down-sampling multiple of the encoder, achieving accurate recovery of the spatial dimension.

[0110] Specifically, the shared feature encoder captures the dependency between different time steps and spatial nodes in the completed spatio-temporal data tensor through the self-attention layer, and the feedforward neural network layer further extracts nonlinear features. The reverse self-attention layer of the decoder network models the inverse dependency relationship of the deep fusion feature vector, and the deconvolution layer gradually expands the spatial dimension of the feature vector. For example, if the encoder down-samples the input data from 128x128 to 16x16 through three convolution layers, the decoder up-samples 16x16 back to 128x128 through three deconvolution layers. The bilinear interpolation algorithm is inserted between the deconvolution layers to recover the detail information in a smooth transition manner. Due to the symmetry of the encoding and decoding structure, the deep fusion feature vector is reversibly transformed into a tensor of the same dimension as the original data in the reconstruction process, maximally preserving the spatio-temporal feature mode sensitive to hidden dangers and ensuring the accuracy of the reconstruction loss calculation, and then optimizing the parameter consistency of the completion and discrimination models through the joint loss function backpropagation.

[0111] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0112] The data reconstruction branch is a decoder network symmetrical to the shared feature encoder structure. The decoder network receives the deep fusion feature vector and up-samples it to the same dimension as the completed spatio-temporal data tensor, generating a reconstructed spatio-temporal data tensor.

[0113] Specifically, the decoder network adopts a multi-layer deconvolutional structure. The first deconvolutional layer receives the deep fusion feature vector, and up-samples it from the dimension of 1x1xC to the dimension of H / 4xW / 4xC / 2. The second deconvolutional layer further up-samples the feature map to the dimension of H / 2xW / 2xC / 4. The last deconvolutional layer up-samples the feature map to the dimension of HxWxC of the original completed spatio-temporal data tensor. Wherein, H, W and C represent the height, width and channel number of the completed spatio-temporal data tensor, respectively.

[0114] Further, after each deconvolutional operation, a batch normalization layer and a ReLU activation function are added to enhance the non-linear expression ability of the model. Thus, the decoder network can gradually recover the spatio-temporal structure information contained in the deep fusion feature vector, and generate a reconstructed spatio-temporal data tensor with the same dimension as the original completed spatio-temporal data tensor.

[0115] The application further proposes to construct a joint loss function based on the weighted sum of the discriminative loss and the reconstruction loss, including calculating the discriminative loss by using a cross-entropy loss function; calculating the reconstruction loss by using a mean square error function; setting a first weight coefficient and a second weight coefficient, multiplying the discriminative loss by the first weight coefficient, multiplying the reconstruction loss by the second weight coefficient, and then summing to obtain the joint loss function; in the early stage of model training, increasing the second weight coefficient to prioritize the accuracy of data completion, and in the later stage of model training, increasing the first weight coefficient to focus on improving the accuracy of hazard discrimination.

[0116] Wherein, the cross-entropy loss function is used to measure the difference between the hazard risk grade prediction result and the true label, and the mean square error function is used to evaluate the fitting degree between the reconstructed data and the completed data. The sum of the first weight coefficient and the second weight coefficient is a fixed value, the second weight coefficient is set to 0.8 and the first weight coefficient is set to 0.2 in the early stage of training; the second weight coefficient is adjusted to 0.3 and the first weight coefficient is adjusted to 0.7 in the later stage of training. The weight adjustment strategy is realized by an exponential decay function, and the second weight coefficient decreases linearly with the increase of the training round.

[0117] Specifically, in the early stage of model training, the data completion module is not yet stable, at this time the reconstruction loss dominates, forcing the model to prioritize learning the interpolation ability of missing data, ensuring that the completed spatio-temporal data tensor has high fidelity. As the training round progresses, the completion module tends to converge, at this time the weight of the reconstruction loss is gradually reduced, and the weight of the discriminative loss is simultaneously increased, so that the model shifts the optimization focus to the discrimination accuracy of the hazard risk grade. By dynamically adjusting the contribution proportion of the two losses, the transmission of completion error to the discrimination module is avoided, and the extraction ability of hazard sensitive features is strengthened, ultimately realizing the synchronous improvement of completion accuracy and discrimination performance.

[0118] As a preferred embodiment, the scheme of the present application is implemented as follows: in the model training process, the cross-entropy loss function is used to calculate the difference between the risk level probability distribution output by the hidden danger identification branch and the true label, and the mean square error function is used to measure the deviation of the reconstructed spatio-temporal data tensor and the completed spatio-temporal data tensor in the numerical dimension. The first weight coefficient is initially set to 0.2, and the second weight coefficient is initially set to 0.8, and the sum of the two is kept at 1. During the first 50 epochs of training, the second weight coefficient decreases linearly to 0.3, and the first weight coefficient increases synchronously to 0.7; after 50 epochs of training, the second weight coefficient is fixed at 0.3, and the first weight coefficient is fixed at 0.7. The learning rate scheduler automatically adjusts the weight ratio of the two items in the joint loss function according to the current epoch number, ensuring that the model prioritizes learning the data completion task, and gradually strengthens the optimization intensity of the hidden danger identification task after the completion error is stable.

[0119] The present application further proposes that when the output traffic hidden danger risk level is a medium risk hidden danger or a high risk hidden danger, a pre-trained attribution analysis model is called; the attribution analysis model analyzes the deep fusion feature vector, calculates the contribution score of the traffic flow index sequence, the driving behavior feature sequence and the road environment parameter sequence to the current hidden danger risk level; the dominant factor causing the hidden danger is determined according to the contribution score, and the traffic hidden danger risk level and the dominant factor are pushed to the traffic control center together.

[0120] Among them, the attribution analysis model adopts the gradient back propagation method, and quantifies the contribution of different indicators by calculating the gradient amplitude of each index feature in the deep fusion feature vector to the hidden danger classification output. In specific implementation, the contribution score is mapped to the interval of 0 to 1 through normalization processing, and the index with a score higher than a preset threshold is determined as the dominant factor. For example, when the contribution score of the crosswind speed is 0.85, the contribution score of the sudden deceleration frequency is 0.72, and the contribution scores of other indicators are all lower than 0.5, the system identifies the crosswind speed and the sudden deceleration behavior as the dominant factors.

[0121] Specifically, the attribution analysis model is automatically triggered when the hidden danger risk level reaches medium-high risk, and the input is the deep fusion feature vector extracted by the shared feature encoder. The model calculates the sensitivity of the hidden danger classification probability to the input features through layer-by-layer back propagation, generating a contribution heat map of each original data sequence. After smoothing and threshold filtering, the heat map selects key indicators that have a significant impact on the current risk level. Further, the system associates the dominant factors with the risk level, and when pushing to the traffic control center, the specific indicator values and time stamps are attached, such as "high risk hidden danger: crosswind speed 8.5 m / s (contribution 0.85), sudden deceleration frequency 12 times per minute (contribution 0.72)". This process makes the hidden danger warning information interpretable, providing direct basis for traffic control decisions.

[0122] As a preferred embodiment, the scheme of the present application is implemented as follows: when the traffic hazard risk level output by the hazard identification branch is a medium risk hazard or a high risk hazard, the system automatically activates the pre-trained attribution analysis model deployed on the edge computing node. The model is constructed based on gradient-based attribution algorithm, and the gradient sensitivity of traffic flow index sequence, driving behavior feature sequence and road environment parameter sequence on deep fusion feature vector is calculated through back propagation. Specifically, first, the deep fusion feature vector is input into the input layer of the attribution analysis model, and is mapped to the corresponding latent space of each index through three independent fully connected layers respectively. The integral gradient algorithm is used to calculate the path integral value along the feature dimension of each index, and the contribution score of each sequence to the current hazard risk level is generated. When the contribution score of the number of sudden deceleration of the driving behavior feature sequence exceeds the preset threshold, the system determines that it is the dominant factor, and a visual analysis report containing the risk level, the dominant factor label and the contribution distribution is transmitted in real time to the emergency command platform of the traffic control center through the 5G communication module.

[0123] The present application further proposes a multi-index dynamic fusion traffic hazard identification completion system, which comprises a processor and a memory, the memory and the processor are connected, the memory is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the memory to realize the multi-index dynamic fusion traffic hazard identification completion method.

[0124] Among them, the processor adopts a multi-core heterogeneous architecture, including a graphics processing unit and a central processing unit, the graphics processing unit is configured to perform spatial graph convolution operation and timing modeling of the gating cycle unit, and the central processing unit is configured to run the self-attention mechanism and the fully connected classification network. The memory is divided into a cache area and a persistent storage area, the cache area preloads the road network topology structure and the preset hazard label data, and the persistent storage area stores the weight matrix of the cross-index attention module and the contribution calculation algorithm of the attribution analysis model. The processor and the memory establish a data transmission channel through the PCIe bus, the channel bandwidth is not less than 16 GT / s, and the low delay transmission of the completed space-time data tensor between the shared feature encoder and the hazard identification branch is ensured.

[0125] Specifically, when the system is running, the code instructions stored in the memory are loaded into the instruction register of the processor, driving the graphics processing unit to perform a spatial graph convolution operation on the input multi-dimensional traffic data sequence, and after generating a spatially fused feature, a cross-indicator attention module is called to perform dynamic weight allocation. The central processing unit synchronously receives the weighted multi-indicator feature, captures the propagation delay effect in the time dimension through the gated recurrent unit, and outputs the initial completed data tensor. After the spatiotemporal data tensor is completed, it is temporarily stored in the cache area and then input into the shared feature encoder for deep feature extraction. The deep fusion feature vector is distributed to the hazard identification branch and the data reconstruction branch through the bus. The graphics processing unit calculates the reconstruction loss in parallel, and the central processing unit synchronously calculates the cross-entropy loss. The joint loss function updates the model parameters through the back propagation algorithm. When the hazard risk level exceeds the threshold, the processor calls the attribution analysis model in the persistent storage area to calculate the contribution score of each indicator and pushes the result to the traffic control center. Through the cooperative scheduling of the processor and the memory, the hardware architecture realizes end-to-end processing of data completion and hazard identification, effectively eliminates the data delay between independent processing links, and improves the response speed and accuracy of hazard early warning.

[0126] As a preferred embodiment, the scheme of the application is implemented as follows: a multi-indicator dynamic fusion traffic hazard identification and completion system includes a processor and a memory. When the program instructions stored in the memory are executed by the processor, first, multi-dimensional traffic data of a target monitoring area are obtained from a traffic management data platform, a vehicle terminal and a weather service system, wherein the data include missing traffic flow indicators, driving behavior characteristics and road environment parameter sequences. The processor performs interpolation on the missing data through a spatiotemporal dynamic graph completion network, which includes a spatial graph convolution layer and a gated recurrent unit, and a cross-indicator attention module is introduced between the two to dynamically allocate fusion weights of different indicators. The completed spatiotemporal data tensor is input into a joint optimization model, which includes a shared feature encoder, a hazard identification branch and a data reconstruction branch. The shared feature encoder uses stacked self-attention layers to extract deep fusion features, the hazard identification branch generates risk level classification results through a fully connected layer and a Softmax function, and the data reconstruction branch reconstructs spatiotemporal data through a symmetric decoder. The processor synchronously calculates the cross-entropy loss of the classification results and the preset label, and the mean square error loss of the reconstructed data and the completed data, and realizes joint optimization through dynamic adjustment of the weight coefficient. When the risk level reaches a medium-high risk, the processor calls an attribution analysis model to calculate the contribution of each indicator, and pushes the result to the traffic control center.

[0127] In conclusion, the application fills in the network through the space-time dynamic graph, combines the space graph convolution layer of the road network topology, the gating recurrent unit modeling the timing rules and the cross-index attention module, accurately fills in the missing data, fully fuses the multi-index dynamic correlation and considers the propagation delay and the space-time evolution law; the shared feature encoder is used to extract deep fusion features, link the hidden danger identification branch, dynamically fuse the traffic flow, driving behavior and environmental parameters, capture the dynamic formation path of hidden dangers; the joint loss function is used to update the parameters of the filling-in network and the identification model synchronously, form a closed loop optimization, avoid the accumulation of filling-in errors, and significantly improve the accuracy and reliability of hidden danger identification in complex traffic environment.

[0128] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0129] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives and variations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and supplementing traffic hazards through dynamic fusion of multiple indicators, characterized in that, The method includes: Acquire a multi-dimensional traffic data sequence containing missing values ​​within the target monitoring area. The multi-dimensional traffic data sequence includes a traffic flow index sequence, a driving behavior characteristic sequence, and a road environment parameter sequence. The missing values ​​in the multi-dimensional traffic data sequence are imputed by the spatiotemporal dynamic graph completion network to generate a completed spatiotemporal data tensor. The spatiotemporal dynamic graph completion network establishes a spatial graph convolutional layer based on the road network topology to capture the spatial correlation of traffic indicators and couples it with a gated recurrent unit suitable for processing time series data to model the temporal evolution law and propagation delay between multiple indicators. The completed spatiotemporal data tensor is input into a joint optimization model consisting of a shared feature encoder, a hazard identification branch, and a data reconstruction branch; The deep fusion feature vector of the completed spatiotemporal data tensor is extracted using the shared feature encoder. The deep fusion feature vector is input into the hazard identification branch to generate the traffic hazard risk level of the target monitoring area; The deep fusion feature vector is input into the data reconstruction branch to generate a reconstructed spatiotemporal data tensor. The identification loss between the traffic hazard risk level output by the hazard identification branch and the preset hazard label, as well as the reconstruction loss between the reconstructed spatiotemporal data tensor and the completed spatiotemporal data tensor, are calculated. A joint loss function is constructed based on the weighted sum of the identification loss and the reconstruction loss, and the joint loss function is backpropagated to synchronously update the model parameters of the spatiotemporal dynamic graph completion network, the shared feature encoder, the hazard identification branch, and the data reconstruction branch.

2. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 1, characterized in that, The acquisition of multi-dimensional traffic data sequences containing missing values ​​within the target monitoring area includes: Obtain a traffic flow index sequence associated with the target monitoring area from the traffic management data platform. The traffic flow index sequence includes average vehicle speed, lane occupancy rate, and traffic flow. The driving behavior feature sequence is obtained from the vehicle terminal or roadside unit. The driving behavior feature sequence includes the number of rapid accelerations, the number of rapid decelerations, the frequency of abnormal lane changes, and the standard deviation of following distance. The road environmental parameter sequence is obtained from the meteorological service system or environmental sensors. The road environmental parameter sequence includes visibility, rainfall intensity, road surface slip coefficient and crosswind speed.

3. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 1, characterized in that, The spatiotemporal dynamic graph completion network also includes a cross-index attention module, which is located between the spatial graph convolutional layer and the gated recurrent unit. The cross-index attention module is used to calculate the dynamic correlation scores between each pair of the traffic flow index sequence, driving behavior feature sequence and road environment parameter sequence, and to assign fusion weights to the features of different indicators based on the dynamic correlation scores, so as to enhance the representation of hazard-sensitive indicators.

4. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 3, characterized in that, The spatiotemporal dynamic graph completion network imputes missing values ​​in the multi-dimensional traffic data sequence to generate a completed spatiotemporal data tensor, including: Align the timestamps of each indicator in the multi-dimensional traffic data sequence, and mask the missing values ​​to generate the input data matrix and the corresponding missing mask matrix. The input data matrix and the adjacency matrix defined by the road network topology are input into the spatial graph convolutional layer to generate spatial fusion features; The spatial fusion features are input into the cross-index attention module to generate weighted multi-index features; The weighted multi-index features are input into the gated loop unit to capture temporal dependencies and output an initial completed data tensor containing interpolated values. The initial completed data tensor is multiplied element-wise with the missing mask matrix, retaining only the interpolated values ​​at the original missing positions to generate the final completed spatiotemporal data tensor.

5. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 1, characterized in that, The shared feature encoder consists of stacked self-attention layers and feedforward neural network layers. The stacked self-attention layers and feedforward neural network layers are used to perform deep nonlinear transformations on the completed spatiotemporal data tensor to capture complex high-order feature patterns before the occurrence of potential hazards.

6. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 1, characterized in that, The hazard identification branch is a fully connected classification network. The fully connected classification network receives the deep fusion feature vector and outputs a probability distribution corresponding to four preset levels: "no hazard", "low-risk hazard", "medium-risk hazard" and "high-risk hazard" through a Softmax activation function. The level with the highest probability is taken as the traffic hazard risk level.

7. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 1, characterized in that, The data reconstruction branch is a decoder network that is symmetrical to the shared feature encoder structure. The decoder network receives the deep fusion feature vector and upsamples it to the same dimension as the completed spatiotemporal data tensor to generate the reconstructed spatiotemporal data tensor.

8. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 1, characterized in that, The construction of a joint loss function based on the weighted sum of the discrimination loss and the reconstruction loss includes: The discrimination loss is calculated using the cross-entropy loss function; The reconstruction loss is calculated using the mean square error function; Set a first weighting coefficient and a second weighting coefficient, multiply the discrimination loss by the first weighting coefficient, multiply the reconstruction loss by the second weighting coefficient, and then sum them to obtain the joint loss function; In the early stages of model training, the second weight coefficient is increased to prioritize ensuring the accuracy of data completion. In the later stages of model training, the first weight coefficient is increased to focus on improving the accuracy of hazard identification.

9. The method for identifying and supplementing traffic hazards by dynamic fusion of multiple indicators according to claim 6, characterized in that, The method further includes: When the output traffic hazard risk level is "medium risk hazard" or "high risk hazard", a pre-trained attribution analysis model is invoked. The attribution analysis model analyzes the deep fusion feature vector and calculates the contribution scores of the traffic flow index sequence, driving behavior feature sequence, and road environment parameter sequence to the current hidden danger risk level. The dominant factors causing the potential hazards are determined based on the contribution score, and the risk level of the traffic hazard and the dominant factors are pushed to the traffic control center.

10. A traffic hazard identification and completion system with multi-indicator dynamic fusion, characterized in that, The method includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the multi-indicator dynamic fusion method for identifying and supplementing traffic hazards as described in any one of claims 1-9.

Images