A hierarchical cooperative traffic signal control method and device based on direction perception and adaptive strategy fusion, equipment and medium

Abstract

The present application belongs to the technical field of intelligent traffic signal control, and discloses a layered cooperative traffic signal control method, device, equipment and medium based on direction perception and adaptive strategy fusion. The method obtains local traffic state information and direction perception neighborhood traffic state information of each intersection, constructs neighborhood message representation through direction level attention and neighbor level attention, and generates a local control strategy; the local traffic state is subjected to feature extraction, low-dimensional mapping and aggregation to generate global observation information, and a global control guidance strategy is generated; the fusion weight of the local strategy and the global strategy is adaptively adjusted according to the global running performance index change and the traffic pressure dispersion, a fusion control strategy is generated, and the traffic signal phase is controlled. The present application improves the adaptability, cooperativeness, stability and road network traffic efficiency of multi-intersection traffic signal control.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic signal control technology, and in particular to a hierarchical cooperative traffic signal control method, device, equipment and medium based on the fusion of direction perception and adaptive strategy, for adaptive signal control and cooperative optimization of multi-intersection road networks. Background Technology

[0002] With the continuous growth in the number of motor vehicles in cities, coordinated traffic signal control at multiple intersections has become an important technical means to improve road traffic efficiency and alleviate traffic congestion.

[0003] Existing traffic signal control methods mainly include timed control methods, rule-based or traffic pressure-based control methods, and reinforcement learning-based adaptive control methods. Among them, timed control methods rely on pre-set fixed timing schemes, making it difficult to adapt to real-time changes in complex and dynamic traffic environments. While rule-based or traffic pressure-based control methods can make some adaptive adjustments based on local traffic conditions, they usually focus on the control of individual intersections or local areas, lacking effective modeling of the overall road network coordination. This can easily lead to inconsistencies between local optimization and global operational results, and may cause congestion to shift and spread within the road network.

[0004] Traffic signal control methods based on reinforcement learning can improve the system's adaptability to some extent by learning control strategies through interaction with the traffic environment. However, in complex road network scenarios with multiple intersections, existing reinforcement learning methods still have the following problems: On the one hand, in distributed multi-agent control scenarios, relying solely on local traffic state information or simple neighborhood information for decision-making makes it difficult to effectively characterize the differences in traffic state in different inflow directions at adjacent intersections, resulting in insufficient modeling ability of traffic flow spatial propagation patterns, thus limiting the foresight and collaboration of local decisions; on the other hand, centralized reinforcement learning methods experience rapid growth in their state space and action space as the road network scale increases, easily leading to high training complexity, convergence difficulties, and high deployment costs. In addition, in existing hierarchical reinforcement learning traffic signal control methods, the collaborative relationship between the upper-level global strategy and the lower-level local strategy is usually relatively fixed, lacking a mechanism for adaptively fusing the two types of strategies based on the road network operating state, which easily leads to inter-layer strategy conflicts, unstable learning processes, and fluctuations in overall control performance.

[0005] Therefore, how to ensure the rapid response capability of local intersections, make full use of traffic state information from different inflow directions at adjacent intersections, improve the ability to represent the spatial propagation pattern of traffic flow, and combine the overall operation status of the road network to achieve adaptive fusion of local control strategies and global control guidance strategies, thereby realizing stable and efficient multi-intersection traffic signal collaborative control, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the shortcomings of existing multi-intersection traffic signal control methods, such as insufficient characterization of neighborhood traffic state differences, inadequate connection between local control and global coordination, high complexity of centralized modeling, and insufficient adaptability of inter-layer strategy coordination, this invention proposes a hierarchical collaborative traffic signal control method based on the fusion of direction perception and adaptive strategies. This method aims to improve the adaptability, coordination, and operational stability of traffic signal control in complex road network environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy, comprising the following steps: Step 1: Obtain local traffic status information of each intersection in the target road network and traffic status information of the corresponding direction-sensing neighborhood of each intersection; Step 2: Calculate the direction-level attention weights for the state features of each adjacent intersection in different flow directions, and perform weighted aggregation of the state features of the adjacent intersection in different flow directions to obtain the direction aggregation features of the corresponding adjacent intersection; calculate the neighbor-level attention weights for the direction aggregation features of multiple adjacent intersections, and perform weighted aggregation of the direction aggregation features of multiple adjacent intersections to obtain the neighborhood message representation of the current intersection. Step 3: Based on the local traffic state information of each intersection and the corresponding neighborhood message representation, generate the local control strategy distribution of each intersection through the local decision model; Step 4: Extract features, perform low-dimensional mapping and aggregation on the local traffic state information of each intersection obtained in Step 1 to generate global observation information that characterizes the overall operating state of the target road network; based on the global observation information, generate a global control guidance strategy distribution for each intersection through a global coordination model. Step 5: Determine the fusion state based on the changes in global operational performance indicators and traffic pressure dispersion within a preset time window, and determine the fusion weight coefficient between the local control strategy distribution and the global control guidance strategy distribution according to the preset fusion criteria, thereby generating the fusion control strategy distribution for each intersection. Step 6: Determine the traffic signal phase control action according to the fusion control strategy distribution, and apply it to the traffic lights at each intersection for control; update the local decision model and the global coordination model according to the traffic state change feedback to achieve hierarchical collaborative optimization.

[0008] The direction-aware neighborhood traffic state information is used to characterize the state information of adjacent intersections in different inflow directions; the different inflow directions include east, west, south, and north; the state information includes queue length, vehicle waiting time, control action of the previous control cycle, current signal phase, and traffic pressure.

[0009] In step 2, the direction-level attention weight is calculated by normalization based on the state characteristics of each inflow direction of the same adjacent intersection; the neighbor-level attention weight is calculated by normalization based on the feature correlation between the current intersection and each adjacent intersection; the neighborhood message representation is obtained by weighted aggregation of the direction aggregation features of each adjacent intersection according to the corresponding neighbor-level attention weight.

[0010] The local control strategy distribution is obtained by the local decision model jointly modeling the local traffic state information of the current intersection and the neighborhood message representation; the local decision model generates the action value of each candidate control action based on the joint features, and obtains the execution probability of each candidate control action through the probability normalization function, thus forming the local control strategy distribution.

[0011] The global observation information is obtained by feature extraction, low-dimensional mapping, and aggregation of local traffic state information at each intersection. The extracted state features include the current signal phase, total queue length, maximum queue length, average waiting time, and traffic congestion level at the intersection. The low-dimensional mapping is used to characterize the congestion propagation features in the target road network.

[0012] The fusion state includes an improved state, a deteriorated state, and a stagnant state; the change in the global operational performance index is the difference between the mean of the global operational performance index in the current time window and the mean of the global operational performance index in the previous time window; the traffic pressure dispersion is the variance of the traffic pressure at each intersection in the current time window; when the change in the global operational performance index is greater than or equal to a first preset threshold and the traffic pressure dispersion is less than or equal to a second preset threshold, it is determined to be an improved state; when the change in the global operational performance index is less than or equal to a third preset threshold, or the traffic pressure dispersion is greater than or equal to a fourth preset threshold, it is determined to be a deteriorated state; when the absolute value of the change in the global operational performance index is less than a fifth preset threshold and the absolute value of the change in traffic pressure dispersion is less than a sixth preset threshold, it is determined to be a stagnant state; the improved state corresponds to increasing the weight of the global control guidance strategy distribution, the deteriorated state corresponds to increasing the weight of the local control strategy distribution, and the stagnant state corresponds to adjusting the fusion weight coefficient; the fusion weight coefficient is limited to a preset range.

[0013] A hierarchical cooperative traffic signal control device based on the fusion of direction perception and adaptive strategy, used to implement the hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy, includes: The traffic status acquisition module is used to acquire local traffic status information of each intersection in the target road network and traffic status information of the direction perception neighborhood corresponding to each intersection. The orientation-aware building block is used to calculate orientation-level attention weights and neighbor-level attention weights, and generate neighborhood message representations. The local decision-making module is used to generate the local control strategy distribution for each intersection based on local traffic state information and neighborhood message representation; The global coordination module is used to generate global observation information based on the local traffic state information of each intersection, and to generate the global control guidance strategy distribution of each intersection based on the global observation information. The fusion control module is used to determine the fusion state based on the changes in global operating performance indicators and traffic pressure dispersion within a preset time window, and to determine the fusion weight coefficient according to the preset fusion criteria, so as to generate the fusion control strategy distribution for each intersection. The execution update module is used to distribute and control traffic lights according to the fusion control strategy to perform phase switching, and update the local decision model and global coordination model based on feedback information.

[0014] An electronic device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, causes the processor to perform the method described thereon.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the method described thereon.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention constructs a hierarchical collaborative control structure that combines the local decision-making layer where the local decision-making model is located with the global coordination layer where the global coordination model is located, enabling each intersection to take into account both local rapid response and overall road network coordination, thereby improving the overall efficiency of traffic signal control at multiple intersections.

[0017] 2. By introducing direction-aware neighborhood traffic state information and adopting a two-layer attention mechanism that combines direction-level attention weights and neighbor-level attention weights, this invention can more precisely characterize the impact of different inflow directions and different adjacent intersections on the current intersection, thereby improving the ability to model the spatial propagation pattern of traffic flow.

[0018] 3. This invention constructs global observation information by extracting features, mapping and aggregating local traffic state information at each intersection, thereby reducing the complexity of global modeling and improving scalability in medium-to-large-scale road networks.

[0019] 4. This invention achieves dynamic coordination between the distribution of local control strategies and the distribution of global control guidance strategies through an adaptive strategy fusion mechanism based on the changes in global operational performance indicators and the dispersion of traffic pressure, thereby improving the adaptability of inter-layer collaboration and the stability of the control process.

[0020] 5. This invention can balance local control refinement, global coordination, and operational stability in complex and dynamic traffic environments, thereby improving the traffic efficiency and collaborative control effect of multi-intersection road networks. Attached Figure Description

[0021] Figure 1 A flowchart of a hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy is provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a hierarchical cooperative traffic signal control framework provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction of direction-aware two-layer attention neighborhood information according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the adaptive strategy fusion mechanism provided in an embodiment of the present invention. Detailed Implementation

[0022] This invention provides a hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategies. It acquires local traffic state information and direction-perception neighborhood traffic state information corresponding to each intersection in a target road network. The direction-perception neighborhood traffic state information is used to characterize the differences in traffic state between adjacent intersections in different inflow directions, thereby enhancing the ability to characterize the spatial propagation patterns of traffic flow.

[0023] Furthermore, this invention calculates directional attention weights for the state features of each adjacent intersection in different inflow directions, and then aggregates the state features of the adjacent intersection in different inflow directions using weighted summaries to obtain the directional aggregated features of the corresponding adjacent intersections. Next, it calculates neighbor-level attention weights for the directional aggregated features of multiple adjacent intersections, and then aggregates the directional aggregated features of multiple adjacent intersections using weighted summaries to obtain the neighborhood message representation of the current intersection. Through this two-layer attention mechanism, the influence of different inflow directions and different adjacent intersections on the control decision of the current intersection can be adaptively identified, thereby improving the ability of local control to utilize key neighborhood information.

[0024] Furthermore, based on the local traffic state information and corresponding neighborhood message representations of each intersection, a local control strategy distribution for each intersection is generated through a local decision model. This local control strategy distribution is used to characterize the execution probability of each candidate traffic signal control action. Compared to control methods that rely solely on local states or simple neighborhood states, this invention can consider both the current intersection state and the impact of neighborhood traffic flow propagation at the local level, improving the refinement and foresight of local decision-making.

[0025] Furthermore, feature extraction, low-dimensional mapping, and aggregation are performed on the local traffic state information of each intersection to generate global observation information representing the overall operational status of the target road network. Based on the global observation information, a global control guidance strategy distribution for each intersection is generated through a global coordination model. The global observation information is represented in a low-dimensional manner to reduce the input dimension of global modeling and improve adaptability to medium-to-large-scale road network scenarios. The global control guidance strategy distribution is used to coordinate and guide the local control of each intersection from the overall road network level.

[0026] Furthermore, to achieve adaptive coordination between local control strategies and global control guidance strategies, this invention constructs an adaptive strategy fusion mechanism. The fusion state is determined based on the changes in global operational performance indicators and traffic pressure dispersion within a preset time window. Then, according to preset fusion criteria, the fusion weight coefficients between the local control strategy distribution and the global control guidance strategy distribution are adaptively determined, generating the fused control strategy distribution for each intersection. The fusion state includes an improved state, a deteriorated state, and a stagnant state. When the overall operational performance of the target road network improves and the traffic pressure distribution is relatively balanced, the weight of the global control guidance strategy distribution is increased. When the overall operational performance of the target road network deteriorates or the traffic pressure distribution is uneven, the weight of the local control strategy distribution is increased. When the overall operational state of the target road network is relatively stagnant, the fusion weight coefficients are adjusted to enhance the adaptability of the control strategy. Furthermore, the fusion weight coefficients are limited to a preset range to ensure the stability of the control process.

[0027] Furthermore, traffic signal phase control actions are determined based on the distribution of the fusion control strategy and applied to the traffic lights at each intersection for control; the local decision model and the global coordination model are updated based on traffic state change feedback to achieve hierarchical collaborative optimization between the local and global layers.

[0028] This invention also provides a hierarchical cooperative traffic signal control device based on the fusion of direction perception and adaptive strategy, used to implement the aforementioned hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy. The device includes a traffic state acquisition module, a direction perception construction module, a local decision-making module, a global coordination module, a fusion control module, and an execution update module, with each module working together to complete traffic signal control at multiple intersections.

[0029] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the above-described hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy.

[0030] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor performs the above-described hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy.

[0031] The following is in conjunction with the appendix Figures 1 to 4 The specific embodiments of the present invention will be further described below. It should be understood that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0032] In a preferred embodiment, the target road network consists of multiple intersections, denoted as a set: in, This represents the total number of intersections in the target road network. For each intersection... A corresponding local decision-making model is set up to achieve local traffic signal control; simultaneously, a global coordination model is set up to achieve overall coordinated control at the road network level. For any intersection... The set of its adjacent intersections is denoted as .

[0033] At any moment , No. Local observation information of each intersection is denoted as It is composed of local traffic state information and neighborhood message representation, namely: in, Indicates an intersection At any moment Local traffic state information. In a preferred embodiment, the local traffic state information can be represented as: in, Indicates the current signal phase. This indicates the current queue length at the intersection. This represents the average waiting time. This indicates lane occupancy.

[0034] To construct direction-aware neighborhood traffic state information, let the direction set be defined. , representing east, west, south, and north respectively. For any adjacent intersection In direction The state characteristics on are denoted as In a preferred embodiment, It can be represented as: in, Indicates adjacent intersections In direction The length of the queue on the queue. This indicates the waiting time for vehicles in that direction. Indicates adjacent intersections The control action performed in the previous control cycle Indicates adjacent intersections Current signal phase, Indicates adjacent intersections Traffic pressure.

[0035] To enhance the ability to represent differences in traffic conditions from different inflow directions, this invention first calculates direction-level attention weights for the state features of different directions within a single adjacent intersection. For adjacent intersections... In direction Directional attention weights , can be represented as: in, This represents the direction scoring function. This represents its parameters. Based on the directional attention weights, adjacent intersections... The directional state features of the adjacent intersections are obtained by weighted aggregation of the different directional state features. : in, Represents the directional feature mapping function. This indicates its parameters. Using the above method, the importance of different inflow directions within adjacent intersections to the current intersection control decision can be adaptively identified.

[0036] After obtaining the directional aggregation features of each adjacent intersection, this invention further calculates neighbor-level attention weights among multiple adjacent intersections. Let the intersections... The query features are It can be derived from local traffic status information. After feature mapping function To obtain, that is: Adjacent intersections For the current intersection Neighbor-level attention weights It can be represented as: in, Represents the trainable parameter matrix. This represents the feature dimension. Further, the directional aggregation features of multiple adjacent intersections are weighted and aggregated to obtain the current intersection. Neighborhood messages indicate : By combining the aforementioned two-layer attention mechanism of directional attention and neighbor-level attention, the importance of different directional state features can be distinguished first within a single adjacent intersection, and then the importance of different neighbors can be distinguished among multiple adjacent intersections. This improves the ability to model the spatial propagation pattern of traffic flow and enhances the foresight and collaboration of local decision-making.

[0037] In terms of generating local control strategies, local decision-making models are based on local observation information. Output the action value function of each candidate signal phase action ,in, Indicates an intersection The candidate signal phase action set, This represents the parameters of the local decision-making model. Further, this is achieved through parameters with a temperature coefficient. The softmax function converts action values ​​into local control policy distributions. Its expression is: in, A preset temperature coefficient is used to adjust the exploration level during the training phase. During the execution phase, the control action with the highest probability can be selected based on the local control strategy distribution, or sampling can be performed according to the probability distribution.

[0038] To guide the optimization of local control strategies, this invention constructs a local reward for each intersection. In a preferred embodiment, the local reward is a weighted average of a congestion penalty term, a state improvement term, and a cooperation term, namely: in, , , These are preset weighting coefficients. Further, in a preferred embodiment, each part can be represented as follows: in, and and represent the normalized queue length and average waiting time, respectively. Therefore, the local decision-making model not only focuses on the current congestion level of the intersection itself, but also considers the impact of changes in the state of adjacent intersections on the current control action.

[0039] In terms of constructing global collaborative control information, feature extraction, low-dimensional mapping, and aggregation are performed on the local traffic state information of each intersection to generate global observation information that characterizes the overall operating status of the target road network. For intersections First, through a low-dimensional feature extraction function Obtain its low-dimensional global features : In a preferred embodiment, low-dimensional global features This can be composed of the current signal phase at the intersection, total queue length, maximum queue length, average waiting time, and traffic congestion level. Subsequently, the low-dimensional global features of all intersections are aggregated to obtain global observation information: in, This represents the aggregation function. By employing the aforementioned low-dimensional mapping and aggregation methods, we can avoid the problem of excessively high dimensionality caused by directly concatenating all original local states, and improve the method's adaptability in medium- to large-scale road networks.

[0040] The global coordination model is based on the global observation information. Generate a global control guidance strategy distribution for each intersection. It is used to coordinate and guide local control from the overall road network level.

[0041] In a preferred embodiment, the global coordination model can be constructed using a policy-optimized reinforcement learning model, with the optimization objective being to maximize the global cumulative reward. The global optimization objective can be expressed as: in, These are the parameters for the global coordination model. As a discount factor, This is a global reward.

[0042] To characterize the overall operational performance of the road network, this invention constructs a global reward. In a preferred embodiment, the global reward is a weighted average of the network-wide congestion factor, the network-wide efficiency factor, and the network-wide equilibrium factor, namely: in, , , Preset weighting coefficients; Used to characterize the overall congestion level of the road network Used to characterize the overall traffic efficiency of the road network. This is used to characterize the uniformity of traffic pressure distribution at each intersection. To reduce the impact of differences in flow scales across different traffic phases and enhance training stability, the global reward can be normalized. The normalized global reward is denoted as... Its expression is: in, and They represent the preceding The moving average and moving standard deviation of the global reward over a given historical time step. To prevent smooth terms with a denominator of zero.

[0043] To achieve adaptive coordination between local control strategies and global control guidance strategies, this invention constructs an adaptive strategy fusion mechanism. First, a global operational performance index is defined. In a preferred embodiment, It can be constructed based on a comprehensive analysis of the overall network throughput efficiency, total queue length, and average waiting time, and A larger value indicates a better overall operating effect of the target road network. Let the length be... The average global runtime performance metrics within the time window are: The change in global performance indicators Defined as: To characterize the dispersion of traffic pressure distribution in the road network, the average traffic pressure at intersection i within the current time window is defined as: Traffic pressure dispersion within the current time window It can be represented as: In a preferred embodiment, the fusion state includes an improved state, a deteriorated state, and a stagnant state, and can be determined according to the following rules: when and When, it is judged as an improved state; when or When, it is judged to be a deteriorating state; when and When this occurs, it is determined to be a stagnant state. Among them, , , , , , This is a preset threshold.

[0044] set up The fusion weight coefficient of the global control guidance strategy distribution in the fusion control strategy distribution is then the intersection At any moment Distribution of fusion control strategies It can be represented as: in, The range of values ​​is limited to a preset interval. Internally, when the fusion state is in an improved state, the weight of the global control guidance strategy distribution is increased; when the fusion state is in a deteriorated state, the weight of the local control strategy distribution is increased; when the fusion state is in a stagnant state, the fusion weight coefficients are adjusted by perturbation. In a preferred embodiment, the fusion weight coefficients can be updated as follows: Improve the state deteriorating state stagnation in, , To preset the adjustment step size, For the preset disturbance amount, This represents the amplitude limiting function.

[0045] In a preferred embodiment, the overall optimization objective of the present invention can be expressed as: in, This is a balance coefficient between local and global objectives. Indicates the first The cumulative local reward of the local decision-making model at each intersection is expressed as: Regarding model updates, the local decision-making model and the global coordination model are updated separately. In a preferred embodiment, the local decision-making model is trained using a value function-based reinforcement learning model, whose loss function can be expressed as: Among them, the target value It can be represented as: in, This represents the target network parameters.

[0046] In a preferred embodiment, the global coordination model is trained using a policy optimization learning model, and its parameter updates can be performed based on policy gradient backpropagation using the advantage function and the pruning objective function. Preferably, its pruning objective function can be expressed as: in, Represents the dominance function. Indicates the cutting factor. This represents the ratio of the probability of the new strategy to the probability of the old strategy.

[0047] Regarding the system execution flow, at each environmental interaction moment The system executes the following process: First, road network sensors and computing nodes acquire local traffic state information and direction-aware neighborhood traffic state information for each intersection; second, neighborhood message representations for each intersection are constructed according to direction-level attention weights and neighbor-level attention weights, and local control strategy distributions are generated by the local decision model, while global control guidance strategy distributions are generated by the global coordination model; then, the fusion state is determined based on the changes in global operational performance indicators and traffic pressure dispersion within a preset time window, and the fusion weight coefficients are updated; subsequently, the fusion control strategy distribution for each intersection is generated based on the fusion weight coefficients, and signal phase control actions are determined accordingly; finally, environmental feedback is obtained, and the parameters of the local decision model are adjusted accordingly. and global coordination model parameters Iterative updates will be performed.

[0048] Regarding the implementation environment, in a preferred embodiment, the intersection employs an eight-phase signal control method, the simulation environment is constructed using the SUMO platform, the road network scale is a two-dimensional grid topology with 12 or 16 intersections, the action control cycle is set to 10 seconds, and the total simulation time for a single model training and testing cycle is set to 3600 seconds. Those skilled in the art will understand that the intersection phase control method, action control cycle, and road network scale can be adjusted according to the actual application scenario; this invention is not limited to the specific parameter settings described above.

[0049] The above description is only a preferred embodiment of the present invention. All modifications, equivalent substitutions or improvements made by those skilled in the art without departing from the concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy, characterized in that, Includes the following steps: Step 1: Obtain local traffic status information of each intersection in the target road network and traffic status information of the corresponding direction-sensing neighborhood of each intersection; Step 2: Calculate the direction-level attention weights for the state features of each adjacent intersection in different flow directions, and perform weighted aggregation of the state features of the adjacent intersection in different flow directions to obtain the direction aggregation features of the corresponding adjacent intersection; calculate the neighbor-level attention weights for the direction aggregation features of multiple adjacent intersections, and perform weighted aggregation of the direction aggregation features of multiple adjacent intersections to obtain the neighborhood message representation of the current intersection. Step 3: Based on the local traffic state information of each intersection and the corresponding neighborhood message representation, generate the local control strategy distribution of each intersection through the local decision model; Step 4: Extract features, perform low-dimensional mapping and aggregation on the local traffic state information of each intersection obtained in Step 1 to generate global observation information that characterizes the overall operating state of the target road network; based on the global observation information, generate a global control guidance strategy distribution for each intersection through a global coordination model. Step 5: Determine the fusion state based on the changes in global operational performance indicators and traffic pressure dispersion within a preset time window, and determine the fusion weight coefficient between the local control strategy distribution and the global control guidance strategy distribution according to the preset fusion criteria, thereby generating the fusion control strategy distribution for each intersection. Step 6: Determine the traffic signal phase control action according to the fusion control strategy distribution, and apply it to the traffic lights at each intersection for control; update the local decision model and the global coordination model according to the traffic state change feedback to achieve hierarchical collaborative optimization.

2. The hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy according to claim 1, characterized in that, The direction-aware neighborhood traffic state information is used to characterize the state information of adjacent intersections in different inflow directions; the different inflow directions include east, west, south, and north; the state information includes queue length, vehicle waiting time, control action of the previous control cycle, current signal phase, and traffic pressure.

3. The hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy according to claim 1, characterized in that, In step 2, the direction-level attention weight is calculated by normalization based on the state characteristics of each inflow direction of the same adjacent intersection; the neighbor-level attention weight is calculated by normalization based on the feature correlation between the current intersection and each adjacent intersection; the neighborhood message representation is obtained by weighted aggregation of the direction aggregation features of each adjacent intersection according to the corresponding neighbor-level attention weight.

4. The hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy according to claim 1, characterized in that, The local control strategy distribution is obtained by the local decision model jointly modeling the local traffic state information of the current intersection and the neighborhood message representation; the local decision model generates the action value of each candidate control action based on the joint features, and obtains the execution probability of each candidate control action through the probability normalization function, thus forming the local control strategy distribution.

5. The hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy according to claim 1, characterized in that, The global observation information is obtained by feature extraction, low-dimensional mapping, and aggregation of local traffic state information at each intersection. The extracted state features include the current signal phase, total queue length, maximum queue length, average waiting time, and traffic congestion level at the intersection. The low-dimensional mapping is used to characterize the congestion propagation features in the target road network.

6. The hierarchical cooperative traffic signal control method based on the fusion of direction perception and adaptive strategy according to claim 1, characterized in that, The fusion state includes an improved state, a deteriorated state, and a stagnant state; the change in the global operational performance index is the difference between the mean of the global operational performance index in the current time window and the mean of the global operational performance index in the previous time window; the traffic pressure dispersion is the variance of the traffic pressure at each intersection in the current time window; when the change in the global operational performance index is greater than or equal to a first preset threshold and the traffic pressure dispersion is less than or equal to a second preset threshold, it is determined to be an improved state; When the change in the global operational performance index is less than or equal to the third preset threshold, or the traffic pressure dispersion is greater than or equal to the fourth preset threshold, it is determined to be a deteriorating state; when the absolute value of the change in the global operational performance index is less than the fifth preset threshold and the absolute value of the change in traffic pressure dispersion is less than the sixth preset threshold, it is determined to be a stagnant state; the improved state corresponds to increasing the weight of the global control guidance strategy distribution, the deteriorated state corresponds to increasing the weight of the local control strategy distribution, and the stagnant state corresponds to adjusting the fusion weight coefficient; the fusion weight coefficient is limited to a preset range.

7. A hierarchical cooperative traffic signal control device based on the fusion of direction perception and adaptive strategy, characterized in that, The hierarchical cooperative traffic signal control method based on direction perception and adaptive strategy fusion as described in any one of claims 1 to 6 includes: The traffic status acquisition module is used to acquire local traffic status information of each intersection in the target road network and traffic status information of the direction perception neighborhood corresponding to each intersection. The orientation-aware building block is used to calculate orientation-level attention weights and neighbor-level attention weights, and generate neighborhood message representations. The local decision-making module is used to generate the local control strategy distribution for each intersection based on local traffic state information and neighborhood message representation; The global coordination module is used to generate global observation information based on the local traffic state information of each intersection, and to generate the global control guidance strategy distribution of each intersection based on the global observation information. The fusion control module is used to determine the fusion state based on the changes in global operating performance indicators and traffic pressure dispersion within a preset time window, and to determine the fusion weight coefficient according to the preset fusion criteria, so as to generate the fusion control strategy distribution for each intersection. The execution update module is used to distribute and control traffic lights according to the fusion control strategy to perform phase switching, and update the local decision model and global coordination model based on feedback information.

8. An electronic device, characterized in that, The device includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the method according to any one of claims 1 to 6.

Images