Special scene traffic signal optimization method and system based on PI-GAN and transfer learning

Abstract

The invention discloses a special scene traffic signal optimization method and system based on PI-GAN and transfer learning. A core scheme is divided into three stages of progression: a physical information generative adversarial network (PI-GAN) generates realistic conventional scene data which accords with a traffic flow dynamics law by weighting and fusing adversarial loss and residual loss of an LWR traffic flow conservation equation in a generator loss function; secondly, field adversarial training is adopted to align conventional and extreme scene feature distribution, IDM car-following safety constraints are introduced into strategy optimization, and efficiency and safety are balanced through a penalty mechanism that speed deviation exceeds a threshold value; and finally, a perception-decision-verification three-layer cooperative control architecture is constructed, behaviors violating road capacity or vehicle following safety are intercepted in real time through a physical verification layer, dangerous instructions are compulsively covered, the physical credibility of extreme scene control is remarkably improved, and real data requirements are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to a method and system for coordinated traffic signal control in extreme scenarios based on PI-GAN and transfer learning. Background Technology

[0002] In the field of intelligent traffic signal control, deep learning-based models (such as reinforcement learning) demonstrate good optimization potential in conventional scenarios. However, their application faces fundamental challenges in extreme scenarios such as heavy rain, dense fog, and sudden traffic accidents: traffic data in extreme scenarios (such as vehicle trajectories in heavy rain and congestion distribution after accidents) are naturally scarce and complex, leading to severe undertraining of models. Although Generative Adversarial Networks (GANs) can alleviate the data scarcity problem, their generated results often lack constraints on real traffic flow dynamics (such as traffic flow continuity and car-following models), easily generating data that violates physical common sense (such as maintaining high speed under high-density traffic flow). This results in signal light control strategies (such as green light duration adjustment and phase switching) output by models trained on such data lacking physical credibility, potentially leading to suboptimal or even dangerous decisions (such as sudden light changes causing vehicles to brake abruptly). On the other hand, traditional transfer learning techniques attempt to apply models trained in conventional scenarios to extreme scenarios. However, their core mechanisms typically optimize model generalization solely through data distribution alignment, lacking explicit modeling of the physical nature of traffic flow (such as safety constraints based on traffic flow conservation equations and the Lighthill-Whitham-Richards macroscopic model). This can lead to the transferred strategy ignoring key dynamic constraints in extreme environments. For instance, in pursuing efficiency (such as reducing average waiting time), instructions may violate road capacity limits or trigger micro-vehicle following behaviors (such as rapid acceleration / deceleration) exceeding safety thresholds, resulting in safety hazards and efficiency losses. Therefore, a new method is urgently needed that can embed physical constraints during the extreme scenario data generation phase and integrate physical safety verification during the control strategy transfer and optimization process to overcome data bottlenecks and ensure the effectiveness and safety of control strategies. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for coordinated traffic signal control in extreme scenarios based on physical information generative adversarial networks and transfer learning, thereby solving the problems mentioned in the background art.

[0004] This invention is achieved through the following technical solution: A traffic signal optimization method for special scenarios based on PI-GAN and transfer learning includes the following steps: S1. Construct a physical information generative adversarial network to generate traffic data for typical scenarios and perform adversarial training: The PI-GAN includes a generator and a discriminator. The residual of the traffic flow dynamics equation is introduced into the total loss function of the generator as a physical constraint to generate traffic data in a normal scenario. The generated traffic data in a normal scenario is then subjected to one round of adversarial training to obtain physically compliant traffic data in a normal scenario, which serves as the basis for transfer learning. The traffic flow dynamics equations include the LWR traffic flow conservation equations, and the total loss function of the generator is a weighted sum of adversarial loss and physical loss, where the physical loss is the square mean of the residuals of the LWR traffic flow conservation equations. S2. Based on transfer learning, transfer traffic data from physically compliant, conventional scenarios to extreme scenarios and optimize signal control strategies: Unsupervised domain adaptation is achieved through domain adversarial training. The difference in feature distribution between normal and extreme scenarios is reduced by a domain discriminator, and domain-invariant features are learned. In the optimization of signal control strategy in extreme scenarios, minimizing the average vehicle waiting time is the main objective. At the same time, an IDM car-following model is introduced as a safety constraint. The IDM loss function is used to penalize scenarios where the vehicle speed deviation caused by the strategy adjustment exceeds the safety threshold, resulting in the total control loss with safety constraints. S3. Construct a collaborative control architecture to execute the optimized signal control strategy: The collaborative control architecture includes a perception layer, a decision-making layer, a physical verification and execution layer, and an environmental feedback closed loop. The perception layer performs spatiotemporal alignment, missing value imputation, and feature fusion on real-time sensor data and generated synthetic data, and outputs fused traffic status. Based on the integrated traffic conditions, the decision-making layer outputs initial control commands through a transfer reinforcement learning controller. The physical verification and execution layer verifies whether the initial control command violates the road capacity constraints and IDM car-following constraints. If it violates them, it is forcibly overridden as a safety command; otherwise, the original command is executed. Environmental feedback inputs the post-execution traffic status into the perception layer, forming a closed-loop learning process.

[0005] Furthermore, in S1, the LWR traffic flow conservation equation is: ,in Let v be the traffic flow density, v be the traffic flow speed, t be the time, and x be the spatial location; the physical loss is obtained by calculating the mean squared residual after discretizing the equation by replacing the derivative with the difference quotient.

[0006] Furthermore, in S1, the total loss function of the generator is: , where LG is the adversarial loss, Lphysics is the physical loss, and λp is the physical constraint weight, with a value range of 0.05 to 0.5.

[0007] Furthermore, the objective function for the domain adversarial training is: ,in Let D be the probability distribution of typical scenario data, D be the domain discriminator, and E be the expectation. This represents the probability distribution of data in extreme scenarios. Indicates that the variable follows a distribution , This indicates that the variable x follows a distribution. .

[0008] Furthermore, in S2, the IDM loss function is: , where N is the total number of vehicles, vi is the actual speed of the i-th vehicle, vdesired is the desired safe speed, and the safety threshold is 5.0 m / s.

[0009] Furthermore, in S2, the total control loss is: Where Ltask is the task loss that minimizes the average latency, and Lcontrol is the total control loss. This is the IDM constraint weight, with a value of 0.2.

[0010] Furthermore, in S3, the missing values ​​of the perception layer are filled by the physical compliance synthetic data generated in S1 to complete the missing data caused by sensor failure or extreme environment.

[0011] Furthermore, in S3, the road capacity constraint verification of the physical verification and execution layer includes: calculating the estimated future inflow volume, checking whether it exceeds the maximum capacity of the downstream road segment or intersection, and automatically lowering the capacity limit under extreme weather conditions.

[0012] Furthermore, traffic signal optimization systems for specific scenarios based on PI-GAN and transfer learning include: Physical Information Generative Adversarial Network (PI-GAN) module: In the first stage, it generates traffic data for regular scenarios that conform to the LWR traffic flow conservation equation. After the first round of adversarial training, it adapts and generates extreme scenario data through transfer learning. The generator loss function includes adversarial loss and physical equation residual loss. Transfer learning module: Aligns feature distributions of normal and extreme scenarios through domain adversarial training, and optimizes signal control strategies by combining IDM car-following constraints. The total control loss includes task loss and IDM constraint penalty. The collaborative control module includes a perception submodule, a decision-making submodule, a physical verification and execution submodule, and an environmental feedback submodule. The perception submodule outputs fused traffic status, the decision-making submodule outputs preliminary control commands, the physical verification and execution submodule verifies and executes the commands, and the environmental feedback submodule forms a closed-loop learning mechanism.

[0013] Furthermore, when the physical verification and execution submodule detects that the control command violates the road capacity constraint or IDM car-following constraint, it forcibly locks the current phase and freezes the timing adjustment, while triggering the alarm log.

[0014] The beneficial effects of this invention are as follows: 1. Traditional GANs tend to deviate from real traffic flow dynamics (such as traffic flow conservation) when generating extreme traffic data, leading to the failure of subsequent model training. This invention innovatively couples a Physical Information Neural Network (PINN) with a GAN, explicitly introducing residual constraints from the LWR traffic flow conservation equation into the generator loss function. By dynamically adjusting the physical weights λp (0.05~0.5), the generated data is forced to strictly adhere to the macroscopic physical correlation of "density-speed-flow" while satisfying statistical realism (deceiving the discriminator) (e.g., avoiding distortion scenarios of high-speed driving under high density). For example, in the generation of rainstorm scenarios, the synthetic data can simulate low visibility characteristics while ensuring traffic flow continuity through residual loss. This provides signal optimization models with training samples that are both rich and physically reliable, fundamentally solving the problem of scarce extreme data.

[0015] 2. Traditional transfer learning only aligns data distributions, failing to guarantee the physical rationality of the strategy in extreme scenarios. This invention employs a dual physical constraint mechanism: First, during the transfer phase, while extracting domain-invariant features using domain adversarial training, the IDM car-following model is simultaneously embedded as a safety constraint into the control loss. The weight design with λi=0.2 forces the strategy to penalize dangerous actions where vehicle speed deviation exceeds a threshold (e.g., 5 m / s) when minimizing the average waiting time (L_task). For example, in the transfer control of sudden accidents, while shortening the red light time in the congested direction, the strategy suppresses the sudden speed drop caused by abrupt green light changes due to the IDM loss, thus eliminating the decision blind spot of "efficiency first, safety at the expense of safety" at the mechanistic level.

[0016] 3. Existing signal control systems lack real-time physical verification of control commands, which can easily lead to secondary risks. The collaborative control architecture of this invention adds a physical verification and execution layer, establishing a dual hard-interception mechanism: at the macro level, it dynamically verifies the capacity of downstream road segments (e.g., automatically reducing the capacity limit during heavy rain) and rejects overloaded passage commands; at the micro level, it uses a lightweight IDM model to simulate and predict key vehicle speed deviations, intercepting actions exceeding safety thresholds (e.g., acceleration > 3 m / s²). Once a violation occurs, timing adjustments are immediately frozen, phases are locked, and an alarm is issued. This mechanism transforms traffic flow dynamics into executable safety rules, forming a "decision-verification-execution-feedback" closed loop, ensuring the physical feasibility of control commands at the system level.

[0017] 4. Traditional methods rely on a large amount of real-world data from extreme scenarios, resulting in high costs for model training and hardware deployment. This invention replaces real-world data collection with PI-GAN synthetic data, significantly reducing data acquisition costs; the generated physically compliant data can seamlessly fill in missing values ​​in the perception layer (such as faulty sensor data during heavy rain), improving the system's robustness in abnormal environments and driving the upgrade of intelligent transportation systems from "experience-driven" to "mechanism-data dual-driven". Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto. Example

[0020] See Figure 1 : A traffic signal optimization method for special scenarios based on PI-GAN and transfer learning includes the following steps: S1. Construct a Physical Information Generative Adversarial Network (PI-GAN) to generate traffic data for typical scenarios: PI-GAN includes a generator and a discriminator. The residual of the traffic flow dynamics equation is introduced into the total loss function of the generator as a physical constraint to generate traffic data in a normal scene. Then, an adversarial training is performed on the generated traffic data in a normal scene to obtain physically compliant traffic data in a normal scene, which serves as the basis for transfer learning. The traffic flow dynamics equations include the LWR traffic flow conservation equations, and the total loss function of the generator is a weighted sum of adversarial loss and physical loss, where the physical loss is the square mean of the residuals of the LWR traffic flow conservation equations. S2. Based on transfer learning, transfer the model from normal scenarios to extreme scenarios and optimize the signal control strategy: Unsupervised domain adaptation is achieved through domain adversarial training. The difference in feature distribution between normal and extreme scenarios is reduced by a domain discriminator, and domain-invariant features are learned. In the optimization of signal control strategy in extreme scenarios, minimizing the average vehicle waiting time is the main objective. At the same time, an IDM car-following model is introduced as a safety constraint. The IDM loss function is used to penalize scenarios where the vehicle speed deviation caused by the strategy adjustment exceeds the safety threshold, resulting in the total control loss with safety constraints. S3. Construct a collaborative control architecture to execute the optimized signal control strategy: The collaborative control architecture includes a perception layer, a decision-making layer, a physical verification and execution layer, and an environmental feedback closed loop. The perception layer performs spatiotemporal alignment, missing value imputation, and feature fusion on real-time sensor data and generated synthetic data, and outputs fused traffic status. Based on the integrated traffic conditions, the decision-making layer outputs initial control commands through a transfer reinforcement learning controller. The physical verification and execution layer verifies whether the initial control command violates the road capacity constraints and IDM car-following constraints. If it violates them, it is forcibly overridden as a safety command; otherwise, the original command is executed. Environmental feedback inputs the post-execution traffic status into the perception layer, forming a closed-loop learning process.

[0021] Furthermore, in S1, the LWR traffic flow conservation equation is: ,in Let v be the traffic flow density, v be the traffic flow speed, t be the time, and x be the spatial location; the physical loss is obtained by calculating the mean squared residual after discretizing the equation by replacing the derivative with the difference quotient.

[0022] The LWR traffic flow conservation equation describes the core physical law of the macroscopic evolution of traffic flow. Its essence is that traffic flow neither arises from nothing nor disappears out of thin air; that is, it is the rate of change of traffic flow density over time. ) and traffic flow ( Rate of change with space ( It is necessary to satisfy conservation relationships to ensure the continuity of traffic flow in the spatiotemporal dimensions. In the calculation, the difference quotient is used to replace the micro quotient, such as... replace( ), replace( Discretizing the continuous equation and then calculating the mean squared residual as the physical loss can accurately quantify the degree of deviation between the generated data and the probability of the law. In this process, the extreme scene data generated by PI-GAN not only closely resembles the real data in terms of statistical characteristics, but also strictly follows the intrinsic relationship between "density-velocity-flow", avoiding situations that violate physical common sense, such as maintaining high speed under high density.

[0023] Furthermore, in S1, the total loss function of the generator is: , where LG is the total loss, Ladv is the adversarial loss, Lphys is the physical loss, and λp is the physical constraint weight, with a value range of 0.05 to 0.5.

[0024] This loss function is the core of achieving the dual objective of optimizing data realism and physical compliance. The adversarial loss Ladv drives the generator to produce samples that can deceive the discriminator, ensuring that the data is consistent with the real scene in terms of statistical distribution. The physical loss Lphys forces the data to conform to the laws of traffic flow dynamics. The range of values ​​for λp provides a flexible adjustment mechanism: when physical laws dominate in extreme scenarios (such as highway traffic flow), a value of 0.3-0.5 can be used to strengthen physical constraints; when the scenario relies more on detailed features (such as complex traffic flow at intersections), a value of 0.05-0.2 can be used to focus on realism. The example value of 0.1 is the choice to balance the two, so that the generated data can simulate the special features of scenarios such as rainstorms and accidents, without violating the essential laws of traffic flow.

[0025] Furthermore, the objective function for the domain adversarial training is: ,in Let D be the probability distribution of typical scenario data, D be the domain discriminator, and E be the expectation. This represents the probability distribution of data in extreme scenarios. This indicates that the variable x follows a distribution. , This indicates that the variable x follows a distribution. .

[0026] The objective function is key to achieving cross-scenario knowledge transfer. Its core logic is to force the model to learn "domain-invariant features" through "adversarial" interaction. The domain discriminator D's task is to distinguish whether the input data comes from a normal scenario or an extreme scenario. The objective function drives the discriminator D to minimize the classification error, thus more accurately distinguishing the domains. Meanwhile, the feature extractor counteracts the domain discriminator by maximizing the objective function, gradually weakening scene-specific features in the data (such as the difference in lighting between sunny and rainy days) and strengthening essential features (such as the correlation between queue length and green light duration). Ultimately, the model trained in normal scenarios can work effectively in extreme scenarios with the help of these invariant features. For example, the signal control strategy for rainy days can be modulated based on the queuing conditions in normal scenarios without relying on the data from rainy days.

[0027] Furthermore, in S2, the IDM loss function is: , where N is the total number of vehicles, vi is the actual speed of the i-th vehicle, Vdesired is the desired safe speed, and the safety threshold is 5.0 m / s.

[0028] The IDM loss function is a core constraint for ensuring traffic safety at the micro level. It is based on the "Intelligent Driver Model (IDM)," which describes the relationship between safe speed and distance when following another vehicle (e.g., following vehicles should decelerate synchronously to maintain a safe distance when the vehicle in front decelerates). The loss function calculates the deviation between each vehicle's actual speed vi and the desired safe speed Vdesired, and only applies a penalty for the deviation exceeding 5 m / s (e.g., a penalty of 3 when the speed drops suddenly by 8 m / s). This forces the signal control strategy to avoid dangerous behaviors such as "sudden acceleration and sudden braking." For example, if a sudden change in the traffic light causes the vehicle speed at the intersection to drop suddenly from 15 m / s to 5 m / s (a deviation of 10 m / s), the loss will increase significantly, forcing the strategy to modulate the green light to gradually shorten, allowing vehicles to decelerate smoothly, which conforms to the safe following behavior described by the IDM.

[0029] Furthermore, in S2, the total control loss is: L task To minimize the task loss due to average latency, L control To control overall losses, This is the IDM constraint weight, with a value of 0.2.

[0030] Total control loss is achieved by merging two objectives to balance control requirements under extreme scenarios, L task Focusing on traffic efficiency, with minimizing average vehicle wait times as the core objective (e.g., reducing queues by optimizing green light duration); L IDM The focus is on safety penalties for changes in dangerous speeds. This weighting avoids both the pursuit of efficiency alone (such as excessively compressing red light time leading to vehicles rushing through) and overprotection (such as inefficiency caused by the green light process). For example, in the event of a sudden accident, the strategy will prioritize shortening the green light time in the congested direction to reduce waiting time (lowering L). task However, if this action causes a vehicle speed deviation of 6 m / s (L... IDM =1), the overall control loss will increase, forcing the strategy to be fine-tuned to ensure that the optimal balance between efficiency and safety is found.

[0031] Furthermore, in S3, the missing values ​​of the perception layer are filled by the physical compliance synthetic data generated in S1 to complete the missing data caused by sensor failure or extreme environment.

[0032] In extreme scenarios, sensor malfunctions may lead to the loss of critical data (such as traffic density on a certain road segment). However, the synthetic data generated by PI-GAN, due to the LWR equation, can fill in missing values ​​and avoid physical contradictions caused by traditional interpolation methods. For example, when the sensor at the east entrance of an intersection fails during a rainstorm, the perception layer can call the traffic flow-density data of that entrance generated by PI-GAN (derived based on real-time data of the west entrance and the LWR conservation relationship). The completed data can reflect the traffic flow characteristics during rainstorms and also conform to the physical law that "the traffic flow at the entrance is positively correlated with the queue length", providing unbiased input for the decision-making layer.

[0033] Furthermore, in S3, the road capacity constraint verification of the physical verification and execution layer includes: calculating the estimated future inflow volume, checking whether it exceeds the maximum capacity of the downstream road segment or intersection, and automatically lowering the capacity limit under extreme weather conditions.

[0034] Road capacity constraint verification is crucial to preventing secondary congestion caused by policy implementation. It ensures the feasibility of control commands by dynamically adapting to the road network's carrying capacity. The verification process involves: first, estimating the number of vehicles entering the road within a certain timeframe based on the current green light duration and traffic speed; second, comparing this number with the maximum capacity of the area, and deeming the command invalid if it exceeds this limit. In extreme weather conditions, the system automatically lowers the capacity limit due to increased braking distance caused by slippery roads, thus reducing actual traffic capacity. For example, during heavy rain, if a section of road has a normal capacity of 500 vehicles per hour, and verification reveals that the estimated number of vehicles entering the road reaches 550, the system will determine it as invalid and refuse to execute it, avoiding a vicious cycle of vehicle congestion and secondary congestion caused by overloading.

[0035] Furthermore, traffic signal optimization systems for specific scenarios based on PI-GAN and transfer learning include: Physical Information Generative Adversarial Network (PI-GAN) module: In the first stage, it generates traffic data for regular scenarios that conform to the LWR traffic flow conservation equation. After the first round of adversarial training, it adapts and generates extreme scenario data through transfer learning. The generator loss function includes adversarial loss and physical equation residual loss. Transfer learning module: Aligns feature distributions of normal and extreme scenarios through domain adversarial training, and optimizes signal control strategies by combining IDM car-following constraints. The total control loss includes task loss and IDM constraint penalty. The collaborative control module includes a perception submodule, a decision-making submodule, a physical verification and execution submodule, and an environmental feedback submodule. The perception submodule outputs fused traffic status, the decision-making submodule outputs preliminary control commands, the physical verification and execution submodule verifies and executes the commands, and the environmental feedback submodule forms a closed-loop learning mechanism.

[0036] Furthermore, when the physical verification and execution submodule detects that the control command violates the road capacity constraint or IDM car-following constraint, it forcibly locks the current phase and freezes the timing adjustment, while triggering the alarm log.

[0037] When a control command may trigger a risk (such as a sudden drop in vehicle speed of 7 m / s due to violation of IDM constraints, or an overload of downstream road sections due to violation of capacity constraints), the submodule will immediately perform three operations: First, lock the current phase (e.g., keep the green light unchanged) to prevent phase switching from exacerbating the danger; second, freeze timing adjustments (e.g., suspend green light shortening commands) to terminate the risk source; and third, trigger the alarm log to record the violation type (e.g., "IDM constraint violation"), location (e.g., "eastern approach to an intersection"), and key parameters (e.g., speed deviation of 7 m / s). For example, in a sudden accident scenario, if the decision-making level commands an excessively long green light, causing downstream road sections to be overloaded by 50%, the submodule will lock the current phase and issue an alarm, while the system will re-optimize its strategy to prevent further chaos at the accident scene.

[0038] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A traffic signal optimization method for special scenarios based on PI-GAN and transfer learning, characterized in that, Includes the following steps: S1. Construct a physical information generative adversarial network to generate traffic data for typical scenarios and perform adversarial training: PI-GAN includes a generator and a discriminator. The residual of the traffic flow dynamics equation is introduced into the total loss function of the generator as a physical constraint to generate traffic data in a normal scene. Then, an adversarial training is performed on the generated traffic data in a normal scene to obtain physically compliant traffic data in a normal scene, which serves as the basis for transfer learning. The traffic flow dynamics equations include the LWR traffic flow conservation equations, and the total loss function of the generator is a weighted sum of adversarial loss and physical loss, where the physical loss is the square mean of the residuals of the LWR traffic flow conservation equations. S2. Based on transfer learning, transfer traffic data from physically compliant, conventional scenarios to extreme scenarios and optimize signal control strategies: Unsupervised domain adaptation is achieved by employing domain adversarial training, which reduces the difference in feature distribution between normal and extreme scenarios through a domain discriminator, and learns domain-invariant features. In the optimization of signal control strategy in extreme scenarios, the main task objective is to minimize the average vehicle waiting time. At the same time, the IDM car-following model is introduced as a safety constraint. The IDM loss function is used to penalize scenarios where the vehicle speed deviation caused by the strategy adjustment exceeds the safety threshold, thus obtaining the total control loss with safety constraints. S3. Construct a collaborative control architecture to execute the optimized signal control strategy: The collaborative control architecture includes a perception layer, a decision-making layer, a physical verification and execution layer, and an environmental feedback closed loop. The perception layer performs spatiotemporal alignment, missing value imputation, and feature fusion on real-time sensor data and generated synthetic data, and outputs fused traffic status. Based on the integrated traffic conditions, the decision-making layer outputs initial control commands through a transfer reinforcement learning controller. The physical verification and execution layer verifies whether the initial control command violates the road capacity constraints and IDM car-following constraints. If it violates them, it is forcibly overridden as a safety command; otherwise, the original command is executed. Environmental feedback inputs the post-execution traffic status into the perception layer, forming a closed-loop learning process.

2. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S1, the LWR traffic flow conservation equation is: ,in Let v be the traffic flow density, v be the traffic flow speed, t be the time, and x be the spatial location; the physical loss is obtained by calculating the mean squared residual after discretizing the equation by replacing the derivative with the difference quotient.

3. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S1, the total loss function of the generator is: , where LG is the adversarial loss, Lphys is the physical loss, and λp is the physical constraint weight, with a value range of 0.05 to 0.

5.

4. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S2, the objective function of the domain adversarial training is: ,in Let D be the probability distribution of typical scenario data, D be the domain discriminator, and E be the expectation. This represents the probability distribution of data in extreme scenarios. Indicates that the variable follows a distribution , This indicates that the variable x follows a distribution. .

5. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S2, the IDM loss function is: Where N is the total number of vehicles, vi is the actual speed of the i-th vehicle, and v desired The desired safe speed is set at a safety threshold of 5.0 m / s.

6. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S2, the total control loss is: Where Ltask is the task loss that minimizes the average latency, and Lcontrol is the total control loss. This is the IDM constraint weight, with a value of 0.

2.

7. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S3, the missing values ​​of the perception layer are filled by using the physical compliance synthetic data generated in S1 to complete the missing data caused by sensor failure or extreme environment.

8. The traffic signal optimization method for special scenarios based on PI-GAN and transfer learning according to claim 1, characterized in that, In S3, the road capacity constraint verification of the physical verification and execution layer includes: calculating the estimated future inbound volume, checking whether it exceeds the maximum capacity of the downstream road segment or intersection, and automatically lowering the capacity limit under extreme weather conditions.

9. A traffic signal optimization system for special scenarios based on PI-GAN and transfer learning, characterized in that, include: Physical information generative adversarial network module: In the first stage, it generates traffic data for regular scenarios that conform to the LWR traffic flow conservation equation. After the first round of adversarial training, it adapts and generates extreme scenario data through transfer learning. The generator loss function includes adversarial loss and physical equation residual loss. Transfer learning module: Aligns feature distributions of normal and extreme scenarios through domain adversarial training, and optimizes signal control strategies by combining IDM car-following constraints. The total control loss includes task loss and IDM constraint penalty. The collaborative control module includes a perception submodule, a decision-making submodule, a physical verification and execution submodule, and an environmental feedback submodule. The perception submodule outputs fused traffic status, the decision-making submodule outputs preliminary control commands, the physical verification and execution submodule verifies and executes the commands, and the environmental feedback submodule forms a closed-loop learning mechanism.

10. The traffic signal optimization system for special scenarios based on PI-GAN and transfer learning according to claim 9, characterized in that, When the physical verification and execution submodule detects that the control command violates the road capacity constraint or IDM car-following constraint, it forcibly locks the current phase and freezes the timing adjustment, while triggering the alarm log.

Images