Traffic signal control method for complex traffic network critical congestion suppression

Abstract

The application discloses a traffic signal control method for complex traffic network critical congestion suppression, a joint state representation is formed by a microscopic trajectory tensor and a phase transition risk index, a SM-PTC state space sequence modeling backbone network and a strategy-value joint network are taken as cores, strategy updating is realized by synergistic optimization of an efficiency term, a control cost term and a risk penalty term, and online collaborative control of multiple intersections in a complex road network is realized; wherein the SM-PTC state space sequence modeling backbone network comprises an input mapping unit used for mapping the whole-network microscopic trajectory tensor to a hidden space, a causal time domain filtering unit used for locally filtering time sequence embedding, a task-oriented gated cumulative state updating unit and a multi-time-constant parallel recursion unit. The method effectively solves the technical problems that the prior art is difficult to effectively cope with the complex phase transition behavior of the traffic network due to the lag of critical congestion identification and excessive action switching in the complex network scene.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and autonomous driving technology, and relates to traffic signal control, and more particularly to a traffic signal control method for suppressing critical congestion in complex traffic networks. Background Technology

[0002] Urban traffic congestion has become a core problem restricting the operational efficiency of modern cities. From the perspective of complex systems, as traffic demand increases, the system undergoes a nonlinear phase transition from a "free-flow phase" to a "congested phase." At the critical state, the system is in a metastable state, where even small disturbances can trigger the cascading spread of congestion shock waves, making congestion highly unpredictable. To address this problem, researchers have proposed various signal control methods, but the following technical limitations still exist: First, existing technologies struggle to achieve explicit perception and early intervention of congestion evolution risks. Traditional methods such as fixed timing, sensor-based control, and Max-Pressure are effective under steady-state conditions, but they primarily rely on local statistics for response and lack continuous perception of the spatiotemporal coupling evolution process of the entire network. Some studies have attempted to introduce risk indicators for traffic state identification, but most are reactive control measures implemented after an event occurs. For example, patent document CN118247950A discloses a method for estimating the boundary control activation threshold under sudden traffic events. This method characterizes the critical point of optimal road network efficiency through the extreme points of the macroscopic basic map and introduces a vulnerability index to correct the critical threshold under sudden events. However, this method focuses on determining the timing of boundary control activation and does not dynamically introduce congestion evolution risks as continuous state variables into the signal control closed loop, making it difficult to achieve "prevention"-style early intervention when the system approaches the phase transition critical region.

[0003] Secondly, there is a contradiction between the extraction of long-term time-series dependencies and the efficiency of online deployment in existing sequence modeling methods. Deep reinforcement learning (DRL) has shown great potential in the field of traffic signal control. However, existing methods have significant limitations in long-term time-series dependency modeling. Convolutional neural network-based methods treat traffic flow as a static spatial image, severing the continuous propagation process of congestion waves in the time dimension. Although the Transformer architecture has global attention, its computational complexity increases quadratically with the sequence length and is prone to overfitting to high-frequency random noise. It is unstable when dealing with drastic fluctuations near the phase transition critical point and is difficult to meet the real-time requirements of online deployment of complex road networks. To address the efficiency problem of sequence modeling, Albert Gu and TriDao proposed the Mamba architecture, which can effectively capture long-range dependencies in long sequences while maintaining linear computational complexity, providing a new technical possibility for processing continuous time series such as traffic flow. However, directly applying the original Mamba model to the phase transition control task of complex traffic networks faces the following series of technical difficulties and challenges: (1) Incompatible input data structure. The original Mamba model was designed for general sequence modeling tasks, with input being a regular one-dimensional vector sequence. However, the microscopic observation data of complex traffic networks are three-dimensional tensor structures with multiple entity dimensions. If they are not processed and forcibly flattened to adapt to the Mamba input interface, they will irreversibly destroy the dynamic information between vehicles and lose the key signs of critical congestion. (2) There is a conflict between the state recursion complexity and the stability of PPO reinforcement learning training. The original Mamba uses a complete input-dependent selective state space recursion, and its recursive dynamics are sensitive to changes in parameters and input distribution. PPO reinforcement learning training has the characteristics of non-stationary sample distribution and high reward variance. This environmental characteristic combined with the sensitive recursion mechanism can easily lead to gradient instability, loss oscillation, or even policy collapse during training, increasing the difficulty of model convergence. (3) The complete selective SSM structure is difficult to balance engineering real-time performance and deployment interpretability when facing traffic phase change control tasks. The complete selective scanning mechanism of the original Mamba maintains multiple sets of input-dependent parameter matrices and achieves efficient computation through hardware-aware parallel scanning algorithms. This complete parameterization design aims to capture the complex dependencies in general sequences. However, from the perspective of the physical mechanism of traffic congestion evolution, the congestion phase transition is a "gradual accumulation" process: vehicle queues gradually increase, the average speed of road segments gradually decreases, and local congestion gradually spreads to adjacent road segments. This gradual accumulation dynamic process is mathematically closer to a cumulative integral with input-dependent weights than a complex multi-parameter nonlinear recursive transformation. For traffic phase transition control tasks, the structural complexity of the complete selective scan has a significant over-parameterization problem, which not only brings unnecessary computational overhead but also makes it difficult to meet the engineering constraints of online deployment. (4) Insufficient ability to perceive "weak precursors and strong noise" signals of congestion phase transition.Before a full-blown congestion outbreak, it usually manifests only as weak precursor signals such as small speed fluctuations and slow queue accumulation, with amplitudes significantly smaller than normal traffic fluctuations. The SiLU activation function used in the original Mamba model decays rapidly in the negative region, which can lead to premature decay of such weak signals in multi-layer backpropagation, making it difficult for the model to effectively capture and amplify key early warning information and affecting the early identification of critical states. (5) The single time constant state update mechanism is difficult to take into account the multi-timescale characteristics of congestion evolution. The latest traffic flow dynamics research clearly points out that the phase transition process from free flow to congestion in the traffic system is accompanied by "critical slowing down" (CSD), which means that there are two completely different dynamic timescales in the critical region of the system - the state recovery rate approaches zero (extremely slow), but the variance increases, leading to intensified short-term fluctuations (medium to high frequency). The state recursion of the original Mamba model is essentially an exponential smoothing process with a single time constant. Its decay rate of historical states and absorption rate of current inputs are controlled by the same mechanism, which cannot simultaneously optimize the extraction of two different timescale information. (6) Mamba lacks prior adaptation research in the field of reinforcement learning for traffic signal control. Currently, the application of Mamba models in the transportation field mainly focuses on traffic flow prediction, trajectory prediction, and anomaly detection, but has not yet involved the continuous dynamic decision-making scenario of traffic signal control, nor has it been adapted for task-specific features extraction of congestion phase transitions. Direct transfer faces the challenge of domain adaptation. Although existing research has explored various modifications to the Mamba architecture, it has not provided corresponding technical insights on how to adapt it to traffic signal control.

[0004] Third, existing control objectives are singular, lacking synergistic optimization of control smoothness and risk suppression. The reward functions of existing deep reinforcement learning methods are mostly designed based on the heuristic delay minimization principle, with throughput improvement or delay reduction as the main optimization objectives, lacking physical perception of the critical phenomena of congestion phase transitions. For example, patent document CN114360266A discloses a reinforcement learning-based signal control method for intersections based on connected vehicle detection and state perception. This method treats traffic lights as intelligent agents, uses vehicle update waiting time as the reward function, and only considers a single delay indicator. Patent document CN119252051A discloses a multi-intersection signal collaborative control method based on deep reinforcement learning. This method constructs a comprehensive reward function that includes waiting time, queue length, and the influence of adjacent intersections, optimizing the signal control strategy with the goal of maximizing the comprehensive reward. Although the reward function incorporates multiple heuristic metrics, its design remains at the level of a linear combination of delay and queue length. It does not introduce physical quantities that reflect the risk of congestion evolution, making it difficult for the controller to perceive the state of the system approaching the phase transition critical point. It can only passively respond to congestion that has already occurred rather than actively suppress its evolution.

[0005] In summary, current technologies lack a traffic signal control method capable of simultaneously achieving explicit perception and early intervention of congestion evolution risks, efficient modeling and online deployment of long-term time-dependent processes, and coordinated optimization of traffic efficiency, control smoothness, and risk mitigation under high-disturbance conditions in complex heterogeneous road networks. Therefore, how to construct a collaborative control method for complex heterogeneous road networks, ensuring engineering feasibility and online deployability, and unifying and closed-loop coupling of phase-change risk perception, state-space temporal modeling, multi-time-scale recursion, and action switching constraints, is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0006] In view of the shortcomings and deficiencies of existing technologies, this invention provides a traffic signal control method for suppressing critical congestion in complex traffic networks. It constructs a complete control closed-loop scheme that matches the physical process of progressive congestion accumulation: a joint state representation is formed using microscopic trajectory tensors and phase transition risk indicators; a backbone network modeled by SM-PTC state-space sequences and a policy-value joint network are used as the core; and policy updates are achieved through the collaborative optimization of efficiency, control cost, and risk penalty terms. This enables online collaborative control of signals at multiple intersections in complex road networks, effectively solving the technical problems of lagging critical congestion identification and excessive action switching in existing technologies in complex network scenarios, which makes it difficult to effectively cope with the complex phase transition behavior of traffic networks. This method can improve traffic efficiency, reduce congestion risk and control jitter, and has good engineering feasibility.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A traffic signal control method for suppressing critical congestion in complex traffic networks, comprising the following steps: Step 1. Collect traffic observation data of the controlled traffic network at the current decision moment, and construct the micro-trajectory tensor of the entire network. ; Calculate the phase transition risk index to characterize the risk of congestion evolution. ; Step 2. Convert the microscopic trajectory tensor of the entire network Inputting the SM-PTC state-space sequence to model the backbone network, extracting temporal dependency features, and comparing them with phase transition risk indicators. Feature fusion is performed to form policy input features; the SM-PTC state-space sequence modeling backbone network includes features for... The system includes an input mapping unit mapped to the latent space, a causal temporal filtering unit for local filtering of temporal embeddings, a task-oriented gated accumulation state update unit, and a multi-time constant parallel recursion unit; wherein the multi-time constant parallel recursion unit includes a fast branch gated accumulation module, a slow branch gated accumulation module, an adaptive fusion module, and an output modulation module. Step 3. Feed the policy input features into the policy-value joint network and output the joint signal control action for each controlled intersection; Step 4. Optimize and update the policy-value joint network based on a composite reward function that includes efficiency, control cost, and risk penalty terms, and impose constraints on action switching; Step 5. Repeatedly execute the action delivery, environmental feedback acquisition, and policy update to obtain the trained policy network; Step 6. Based on the trained SM-PTC state-space sequence modeling backbone network and policy network, perform real-time signal control at the controlled intersection.

[0008] As a preferred embodiment of the present invention, step 1 determines the active vehicle set based on the influence radius of the controlled intersection, and collects the micro-motion state of the current active vehicle set at each decision moment, including normalized position, speed, acceleration, heading angle, and lane index; then, the individual micro-feature vectors of the active vehicles are extracted and sorted according to the priority of vehicles closer to the stop line and with lower speed, to obtain the vehicle-feature matrix at a single moment. Subsequently, they are stacked in chronological order to construct the microscopic trajectory tensor of the entire network. .

[0009] As a preferred embodiment of the present invention, the phase transition risk index in step 1 It is obtained by weighted coupling of density and velocity components, and the expression is: ; in, This represents the current cumulative number of vehicles. The saturation threshold The current average speed, As a reference speed threshold, These are the weighting coefficients.

[0010] As a preferred embodiment of the present invention, the input mapping unit in step 2 is used to map the microscopic trajectory tensor of the entire network. Flatten along the time dimension and map it to the latent space through a linear projection layer, then cut it into input branches. Gated branches ; The causal temporal filtering unit employs a causal one-dimensional convolution with a kernel size of 3 to embed the temporal branch, i.e., the input branch. Local filtering is performed, followed by nonlinear activation. The activation function is selected as follows: .

[0011] As a preferred embodiment of the present invention, the task-oriented gated cumulative state update unit in step 2 employs an adaptive gating weight based on input dependency to achieve state recursion, completing the selective writing and state accumulation of key historical evidence. When processing data, it first... Generate input-dependent gating coefficients; where, It is the Sigmoid activation function. Representative to Linear projection is the initial transformation for feature extraction, used to reduce the dimensionality of the data and extract essential features. This indicates that another linear transformation is performed, which is a linear projection over time to capture dynamic evolution information. The low-frequency trend component is the output of the causal time-domain filter unit. This is the gating factor; then through Construct a gated cumulative state update mechanism; among which, The hidden states at adjacent time points; finally, through The accumulated state is modulated at the output; where, The output features of the task-oriented gated cumulative state update unit are provided.

[0012] As a preferred embodiment of the present invention, the fast branch gated accumulation module in step 2 uses the fast branch time constant Output features of the task-oriented gated cumulative state update unit State recursion is performed to track high-frequency intrinsic fluctuations in traffic flow at the signal periodic scale, generating fast branch state. Its gating coefficient Simultaneously dependent on phase transition risk indicators and mid-to-high frequency fluctuation intensity index : ,in The learnable parameters for the fast branch are: feature weights, slow-varying risk weights, fast-varying volatility weights, and a bias term; its state recursion formula is: ; The slow branch gating accumulation module uses the slow branch time constant Output features of the task-oriented gated cumulative state update unit Perform state recursion, preserve the gradual accumulation trend across cycles, and generate slow branch states. Its gating coefficient Relying solely on phase transition risk indicators : ,in , , The learnable parameters for the slow branch are the feature weights, the slow-varying risk weights, and the bias term; its state recursion formula is... ; The adaptive fusion module is used based on phase transition risk indicators. Generate fusion coefficient , , and according to The fast branch state and the slow branch state are merged into the main state. ， ,in , To integrate learnable gating parameters; The output modulation module passes through Output modulation is performed on the accumulated state to obtain the output features related to the current traffic control decision. .

[0013] As a preferred embodiment of the present invention, the expression for the composite reward function in step 4 is: ; in, The revenue item is used to characterize the efficiency of traffic operation. The constraint term characterizes the cost of switching control actions. This is a risk penalty term that characterizes the risk of congestion evolution.

[0014] As a preferred embodiment of the present invention, the control system for implementing the control method includes a road network interaction module, a state evaluation module, an intelligent decision control module, and an online control and feedback module; wherein, the road network interaction module is used to interact with the real road network, and includes a data acquisition module and an action execution module. The data acquisition module is used to collect the microscopic operating status of vehicles and the operating status of intersections in the controlled road network at a preset decision cycle; the action execution module executes signal control actions and drives the online control terminal to proceed according to the preset decision cycle. The state evaluation module is used to construct a network-wide micro-trajectory tensor based on the vehicle micro-operational states collected by the road network interaction module. And calculate the phase transition risk index. Both are used together as the state part of the policy input; The intelligent decision-making and control module is used to extract the temporal dependency features of the micro-trajectory tensor of the entire network based on the SM-PTC state-space sequence backbone network, and to correlate them with the phase transition risk index. The system integrates and outputs joint signal control actions for each controlled intersection through a policy-value joint network, and then distributes these joint signal control actions to the road network interaction module. The intelligent decision control module also includes a model optimization and iteration module, which performs large-scale parallel sampling during offline training, updates the parameters of the policy-value joint network based on historical observation data collected by the road network interaction module, and evaluates and optimizes the model on the validation set according to preset key performance indicators, storing the optimal model and its corresponding version information. The model optimization and iteration module also receives online operation logs from the online control and feedback module to achieve continuous iterative optimization of the model.

[0015] As a further preferred embodiment of the present invention, step 2 obtains the output features related to the current traffic control decision. Subsequently, with macroeconomic phase transition risk indicators The data is then concatenated and subjected to dimensionality reduction via two fully connected layers to output the policy input features.

[0016] As a further preferred embodiment of the present invention, the high-frequency fluctuation intensity index in step 2... Cumulative number of vehicles and average speed Obtained by first-order differential weighting of adjacent decision steps; or by capturing the intensity of mid-to-high frequency fluctuations based on the duration of a single signal cycle at the intersection.

[0017] As a further preferred embodiment of the present invention, step 4 characterizes the benefits of traffic operation efficiency. The expression is: ; in, To complete the current decision-making step in terms of vehicle volume, This represents the average waiting time. To accumulate the number of vehicles on the road network, For efficiency parameters; Constraints representing the cost of switching control actions ;in, This represents the number of intersections where a phase switch occurs within the current decision step. To control cost parameters; Risk penalty item representing the risk of congestion evolution ;in, As a phase transition risk indicator, This is a risk penalty parameter.

[0018] Advantages and beneficial effects of the present invention: (1) This invention addresses the problem that the original Mamba input interface does not match the micro-state tensor of the complex traffic network, making it difficult for fine-grained traffic information to be directly entered into the temporal backbone. By adopting the input link of "first sorting and preprocessing to preserve spatial semantics, then vectorizing and flattening, and finally performing linear projection mapping" for the micro-trajectory tensor of the whole network, a task-oriented input adaptation mechanism is constructed. This achieves the effect of preserving vehicle-level motion information, enhancing the ability to represent key congestion symptoms, and improving the stability of online reinforcement learning training without significantly increasing the complexity of the model.

[0019] (2) Unlike existing methods that rely solely on local aggregated statistics such as queue length and waiting time, this invention dynamically introduces phase transition risk indicators, which reflect the degree to which the system deviates from the free-flow ground state, as state components into the control closed loop. This is deeply integrated with the SM-PTC state-space sequence modeling backbone network to construct a joint state expression mechanism of "microscopic trajectory tensor + phase transition risk indicator," achieving a fundamental leap from "post-event response" to "pre-event prevention" at the control paradigm level. This mechanism overcomes the limitation of existing boundary control methods that only focus on determining the initiation timing. It extracts vehicle microscopic details from the SM-PTC time-series backbone and integrates them with the macroscopic phase transition risk perception of the road network, using them together as a state input strategy-value joint network. This enables the network to effectively filter and retain key evidence of congestion accumulation in high-noise environments, while allowing the controller to proactively perceive congestion evolution trends when the system approaches the phase transition critical region, taking preventative actions in advance to prevent congestion from evolving from local disturbances to network-wide spread. This provides a physically interpretable state basis for proactive intervention under phase transition critical states.

[0020] (3) This invention addresses the specific distribution characteristics of congestion phase change data in complex traffic networks, which are characterized by "weak precursors, strong noise, and short-term sudden fluctuations," as well as the problem that traditional activation functions are insufficient in preserving critical weak signals. By introducing TeLU activation mapping into the gated cumulative state update and output modulation, a nonlinear transformation mechanism more suitable for phase change precursor sensing is constructed. Utilizing its characteristic of "stronger gradient preservation ability" on the negative half-axis, it can better preserve weak congestion precursor signals, reduce noise-induced gated false writes, and enhance the stable transmission capability of key state components in the output modulation and multi-time constant parallel recursive fusion process. This improves the accuracy of early identification of critical congestion, the stability of strategy convergence, and the reliability of online collaborative control.

[0021] (3) This invention addresses the problems of poor training stability and gradient collapse in the original Mamba complete selective state space recursion in PPO reinforcement learning scenarios by replacing the complete selective SSM recursion with a task-oriented gated cumulative state update mechanism. Experiments show that this modification makes the gradient flow in the PPO training process smoother, effectively alleviating the loss oscillation and policy collapse problems caused by the non-stationary and high variance characteristics of the sample distribution, and enabling the model to converge stably to a high-performance policy.

[0022] (4) This invention addresses the problems of over-parameterization, computational redundancy, and difficulty in meeting engineering deployment requirements of the original Mamba complete selective scanning mechanism in traffic phase change control tasks. By deeply integrating the physical mechanism of progressive accumulation of traffic congestion with state space recursion, the mathematical form of state recursion is reconstructed into a cumulative integral with input-dependent weights. This constructs a historical information accumulation mechanism of "enhancing key evidence and suppressing non-key fluctuations", which enhances the physical interpretability of the model and meets the constraints of real-time performance and engineering deployment in online signal control.

[0023] (5) This invention addresses the core challenge of a single-time-constant state recursion mechanism in handling the multi-timescale evolution of traffic congestion, where long-term trend preservation and short-term fluctuation perception are incompatible. Following the task-oriented gating accumulation unit, a multi-time-constant parallel recursion unit is introduced. The two types of dynamic processes in the critical phase transition of the traffic system are structurally mapped: fast branches correspond to mid-to-high frequency oscillations and sudden disturbances at the signal periodic scale, while slow branches correspond to cross-cycle congestion accumulation and phase drift. This directly transforms the "difference in physical mechanism" into a "difference in state update mechanism," achieving a deep mathematical unification between the model's perception capability in the critical region and the physical mechanism of congestion. Consequently, before the formation of critical congestion, the system can identify effective early warning patterns earlier while maintaining continuous tracking of long-term risk increases, improving the lead time, stability, and interpretability of phase transition perception control, achieving substantial results.

[0024] (6) This invention implements differentiated configuration of the gating guidance signals for fast and slow branches, solving the different guidance information requirements due to the different functional positioning of fast and slow branches. (This also relates to phase transition risk indicators.) With mid-to-high frequency volatility indicators Functional decoupling and separate usage are implemented: the former is used to characterize the slow-changing trend caused by network-level congestion pressure and efficiency degradation, while the latter is used to characterize the fluctuation intensity caused by changes in velocity and queuing at adjacent time points. This differentiated design allows the model to distinguish between "slow-changing evidence that should be retained" and "fast-changing disturbances that need to be guarded against" during the gating and writing phase, and avoids misjudging short-term noise as a long-term deterioration trend. Compared with the common unified risk feature input method, this invention, through semantic separation of indicators and division of labor in recursive paths, enables fast and slow branches to perform their respective functions and complement each other, achieving a simultaneous effect of improving the granularity of risk perception and suppressing false triggers, thereby enhancing the decision robustness and cross-scenario generalization ability in the critical stage.

[0025] (7) The multi-time-constant parallel recursive unit of this invention serves as a core component of the SM-PTC backbone network. It works in deep collaboration with other components in the network to construct a closed-loop collaborative chain of "input adaptation - trend purification - gating accumulation - parallel recursion - policy fusion". The front-end topology-aware input and causal filtering first improve the quality of evidence. The middle section TeLU gating accumulation suppresses noise errors and retains weak precursors. The rear section dual-time-constant parallel recursion and risk-guided fusion complete multi-scale evidence reconstruction. Finally, the policy head achieves stable control output. This chain-like collaboration enables the technical contributions of each module to support and reinforce each other functionally, avoiding the attenuation of benefits caused by "externalizing multi-scale modules and decoupling them from the backbone". It can achieve more consistent training convergence, lower policy jitter, and higher critical congestion suppression reliability without changing the linear time complexity order and deployment interface of the backbone, and has outstanding engineering feasibility.

[0026] (8) This invention constructs a composite reward function consisting of an "efficiency term + control cost term + risk penalty term," breaking through the limitations of existing methods that only aim to improve throughput or reduce delay. It unifies traffic efficiency improvement, action switching cost constraints, and congestion phase transition risk penalties into a single optimization framework. Action switching regularization effectively suppresses excessive intervention in non-critical states, while risk penalties guide the controller to actively avoid high-risk states, enabling the agent to spontaneously maintain control smoothness and actively suppress congestion evolution while pursuing traffic efficiency. Ablation experiments verify the necessity of the three-way synergy: retaining only the efficiency term maintains throughput, but waiting time increases by 51.8%, and the congestion index worsens by 74.8%; simultaneously, removing the risk penalty and switching regularization term causes a complete collapse in system performance. Only when the three are optimized synergistically can the lowest congestion risk be achieved while maintaining high efficiency.

[0027] (9) This invention is the first to apply the Mamba state-space model to the field of traffic signal control and systematically integrates it with the traffic phase change regulation mechanism to construct a collaborative control framework of "physically inspired phase change sensing mechanism + SM-PTC temporal backbone + three-element composite reward". This framework is not a simple superposition of state-space models (such as LSTM) and phase change theory in the prior art, but rather, through task-specific design, it enables the three to produce synergistic effects and constructs a closed-loop collaborative optimization mechanism. Experimental results show that this invention exhibits stable and reproducible performance advantages under complex heterogeneous topology and high load and high disturbance stress tests. It achieves better control effects than the prior art in many core indicators, and the significant advantage in the stress range indicates that its benefits do not come from the randomness of a single scenario, but from the overall suppression capability of the integrated mechanism of "risk perception, temporal coding, collaborative decision-making, and action constraints" on the system-level congestion propagation chain, fully demonstrating the comprehensive engineering value of phase change regulation for complex networks. Attached Figure Description

[0028] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings: Figure 1 A flowchart of a traffic signal control method for suppressing critical congestion in complex traffic networks provided by the present invention; Figure 2 This is a schematic diagram of the backbone network structure for SM-PTC state-space sequence modeling in this invention; Figure 3 The road network topology diagram used for simulation verification provided by this invention; Figure 4 This is a schematic diagram of the control system structure provided by the present invention; Figure 5 This diagram illustrates a performance comparison between the SM-PTC timing backbone and fixed signal control scheme, and the Max-Pressure (maximum pressure control method) of this invention. Figure 6 This is a schematic diagram comparing the performance of the SM-PTC temporal backbone and the deep reinforcement learning method of the present invention; where (a) is the system throughput and (b) is the average travel time.

[0029] Figure 7 The graphs show a comparison of the activation functions TeLU and SiLU and their derivatives; where (a) is a comparison of the activation functions TeLU and SiLU; and (b) is a comparison of the derivatives of TeLU and SiLU. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0031] like Figure 1 As shown, this embodiment provides a traffic signal control method for suppressing critical congestion in complex traffic networks. The method models the traffic signal control problem as a partially observable Markov decision process (POMDP), defining the agent as the central controller, the environment as the SUMO simulation platform or a real road network, and the state as the microscopic trajectory tensor of the entire network. Phase transition risk indicators The action is the joint signal phase decision for each controlled intersection, and the reward is a weighted combination of efficiency gains, switching costs, and risk penalties. Specifically, the method includes the following steps: Step 1. Collect traffic observation data of the controlled traffic network at the current decision moment, and construct the micro-trajectory tensor of the entire network. And calculate the phase transition risk index characterizing the risk of congestion evolution. This generates centralized microscopic observation data; among which, the whole network microscopic trajectory tensor is used to characterize the spatiotemporal evolution sequence of vehicles within the sensing range of each controlled intersection in the road network, and the phase transition risk index is used to quantify the critical risk degree of the system deviating from the free flow ground state; Step 2. Input the entire network's micro-trajectory tensor into the SM-PTC state-space sequence modeling backbone network, extract temporal dependency features, and compare them with the phase transition risk index. Perform feature fusion to form policy input features; Step 3. Input the policy input features into the policy-value joint network and output the joint signal control action of each controlled intersection; wherein, the joint signal control action can be represented as a multi-intersection phase selection vector, each component of which corresponds to the phase decision of the current decision step of a controlled intersection; Step 4. Optimize and update the policy-value joint network based on a composite reward function that includes efficiency, control cost, and risk penalty terms (end-to-end optimization using the PPO algorithm), and impose constraints on action switching; Step 5. Execute actions in a loop, obtain environmental feedback, and update strategies to achieve closed-loop collaborative control of complex traffic networks.

[0032] In this embodiment, the environment can be a traffic simulation platform or an online control terminal, and the feedback data includes reward values, the state at the next moment, and operational performance indicators. During actual operation, the trained SM-PTC state-space sequence modeling backbone network and policy network can be migrated to the online control platform of the controlled intersection and quickly adapted and adjusted to perform real-time signal control of the controlled intersection.

[0033] Furthermore, in this embodiment, the specific process of step 1 is as follows: Step 1.1. Treat all vehicles within the sensing range of all controlled intersections in the controlled road network as a Lagrange particle system and as the set of active vehicles; the set of active vehicles consists of vehicles that meet the following conditions within the controlled area at the current decision time: 1. The vehicle is within the radius of influence of the controlled intersection, or is located in the controlled lane leading to the intersection; 2. The vehicle is in motion, and its speed, position, or queuing status can be stably acquired by the simulator or detector; 3. The vehicle's signal control for the current decision step has observable correlation.

[0034] Preferably, the set of active vehicles can be represented as:

[0035] in, For vehicles Distance to the parking line For the observation radius, For controlled lanes.

[0036] At each decision moment, the micro motion state of the current active vehicle set is collected, including normalized position, speed, acceleration, heading angle and lane index, a total of 5 features, and the micro feature vector of a single vehicle is:

[0037] in, These represent normalized position, speed, acceleration, heading angle, and lane markings (lane index), respectively.

[0038] Step 1.2. To control computational complexity, a fixed time window length is used at each decision point. Vehicle-level features are extracted from the active vehicle set and encoded according to a unified rule to obtain:

[0039] in, =5 represents the particle feature dimension. This is used to observe the vehicle capacity threshold.

[0040] Specifically, at each decision-making moment, the individual micro-feature vectors of active vehicles are first processed. Extract the data, then sort it according to rules such as "distance to parking line, control relevance, speed," and truncate it to the specified values. The insufficient parts are filled with zeros to obtain the vehicle-feature matrix at a single time step. ; then for continuous At that moment Stacked in chronological order, they form a microscopic trajectory state tensor. .

[0041] Furthermore, to ensure the stability of the input dimension, the following strategy is preferred: a) When the number of active vehicles exceeds When the time comes, vehicles will be prioritized based on their distance from the stop line and their speed, and then truncated. b) When there are missing frames in the window's history, zero-filling, last frame copying, or interpolation filling are used; c) Align frames across time points according to vehicle identifiers to ensure timing consistency.

[0042] In this embodiment, the microscopic trajectory tensor of the entire network is constructed. The sorting and truncation rules used at that time are not only used to control the observation capacity, but also play an important role in preserving the spatial structure information between vehicles.

[0043] Unlike existing spatiotemporal separation modeling methods that introduce graph convolution or are based on the Mamba architecture to process multidimensional traffic data, this embodiment employs a preprocessing strategy of first establishing spatial semantic continuity through sorting, and then organizing it into a tensor structure. This implicitly encodes key spatial information such as relative spatial positions, car-following relationships, and lane interactions between vehicles into the vehicle dimension arrangement order of the tensor. After this sorting, Mid-vehicle dimension The order of the vehicles is not random or solely based on vehicle identification, but rather on a spatial semantic continuity formed by the spatial proximity of the vehicles to the intersection stop lines, their sensitivity to signal control, and their own motion state. After sorting, vehicles at adjacent index positions tend to be physically close and possess similar control relevance and motion characteristics. This design facilitates subsequent adjustments... When vectorized and flattened, vehicle features that were originally spatially adjacent remain locally clustered in the flattened vectors, effectively preserving spatial interactions. Thus, while ensuring a fixed input dimension, controllable computational complexity, and without introducing additional explicit spatial modeling modules, a complete representation of vehicle-level micro-motion information and its spatial interactions is achieved, providing a complete state foundation for the subsequent extraction of congestion evolution features from the SM-PTC time-series backbone.

[0044] Furthermore, in this embodiment, the observed vehicle capacity threshold... The capacity is determined according to capacity estimation rules, rather than by empirically fixed constants. Specifically, the congestion-scale capacity of the controlled area is first estimated: ;

[0045] Then determine the observation budget according to the reduction factor: ;

[0046] in, For the number of controlled lanes, For the effective vehicle length, This is the observation reduction factor, used to balance "sufficiency of observational information" and "affordability of online computation".

[0047] In a preferred embodiment, it is possible to Then we have: ;

[0048] Considering that the risk of phase transition in a transportation system usually occurs before full saturation, the preferred option is... This is used to control the state dimension and online training overhead while ensuring that key congestion signs are preserved, resulting in: ;

[0049] Therefore, this value matches the pre-saturation characteristics of complex road networks before congestion. The determination of the method satisfies both physical interpretability and engineering feasibility, and can be updated across different road networks. , , and Perform adaptive calibration.

[0050] Furthermore, in this embodiment, the phase transition risk index in step 1 is obtained by weighted coupling of the density component and the velocity component, and is defined as:

[0051] in, The number of active vehicles observed at the current decision-making moment. The saturation threshold The average velocity within the same set of observations. As a reference speed threshold, This is a weighting coefficient. This index is determined by the vehicle queue length and average speed at the current decision moment, and is used to quantify the degree to which the system deviates from the free-flow ground state. The density component reflects the degree of queue accumulation, and the velocity component reflects the degree of velocity decrease. Preferably, .

[0052] In this embodiment, The creative contribution lies in dynamically introducing it as a continuous state variable into the deep reinforcement learning control loop, and forming deep synergy with the SM-PTC state-space sequence modeling backbone network and composite reward function, rather than merely using it as a trigger condition for boundary control. Specifically, this is manifested in: 1) Cooperative fusion in state space: With the micro-trajectory tensor of the entire network Together, they constitute the policy input state, enabling the controller to simultaneously obtain complementary information on both "microscopic configuration details" and "macroscopic potential state." This indicator, in the gating cumulative update mechanism of the SM-PTC network, is... The extracted temporal features work together in the state recursion process, enabling the network to effectively filter and retain key evidence of congestion accumulation in high-noise environments, rather than using it as an independent input dimension.

[0053] 2) Physical coupling in the reward function: Directly participate in the compound reward function The risk penalty item enables the intelligent agent to proactively adjust its control strategy based on real-time perception of phase transition risks while pursuing traffic efficiency, thus achieving a fundamental leap from "post-event response" to "pre-event prevention".

[0054] 3) Consistency of parameter calibration: Weight parameters Risk penalty weights in the reward function Offline pairing calibration is adopted to form a unified parameter framework for "risk perception - temporal modeling - action constraints", ensuring that risk signals maintain physical consistency across the three levels of state perception, feature extraction and policy optimization.

[0055] Furthermore, in this embodiment, The preferred method is to determine the process according to the procedure of "observation consistency - threshold calibration - weight calibration".

[0056] 1) Observational Consistency: To ensure consistency between state input and risk description, the micro-trajectory tensor is used. The same set of observed vehicles (set of active vehicles) Calculate the risk item, where:

[0057] in, The set of observation vehicles participating in state construction at the current moment. Let be the speed of the i-th vehicle.

[0058] 2) Threshold calibration: (1) The critical accumulation threshold for road network is determined by MFD (Macroscopic Fundamental Diagram) analysis and is calibrated by the critical accumulation points under historical low-load and high-load scenarios. It is used to characterize the accumulation threshold for "transition from stable operation to severe congestion". (2) Based on the determination of the starting speed of significant slow traffic in urban traffic, the speed boundary point from free flow to significant slow traffic in urban roads is calibrated to characterize the speed degradation trigger threshold.

[0059] (3) Weight calibration: To balance the risk contributions of "cumulative near-saturation" and "significant rate decrease," the method is determined through paired validation set grid search or normalized fitting, with the preferred method satisfying the following conditions. .

[0060] In a preferred embodiment, a saturation level accumulation threshold determined according to the critical accumulation analysis method based on the MFD model can be used. Established based on the reference speed at which severe urban traffic congestion begins. and take Under the above parameters, when the accumulated vehicle volume approaches the saturation threshold or the average speed is below the reference threshold, This will simultaneously increase and trigger stronger risk suppression signals, thereby guiding strategies to take intervention actions in advance.

[0061] Furthermore, such as Figure 2 As shown in this embodiment, the SM-PTC state-space sequence backbone network in step 2 is a task-oriented modification of its core formula to address the insufficient adaptability of the original Mamba structure to phase transition control tasks in complex transportation networks. Specifically, this includes: Step 2.1. Input mapping unit: Input the entire network micro-trajectory tensor Flattened along the time dimension to fit the traffic micro-state input, by merging linear projection layers. Mapping to the latent space:

[0062]

[0063] in, For the input branch, For gated branches, For vectorization operators, It is a linear function.

[0064] In this embodiment, the input branch With gated branches They are the same; both are flattened microscopic trajectory tensors.

[0065] Considering that the phase transition sensing and control task of complex traffic networks focuses on how the micro-level vehicle states gradually evolve into network congestion, the single-moment observation and input is a "vehicle-level multidimensional micro-state tensor," rather than the original Mamba token sequence oriented towards general sequence embedding input. This makes it difficult to directly match the input interface of the original sequence model and easily leads to the loss of critical congestion precursor information during discretization. In this embodiment, by analyzing the entire network's micro-trajectory tensor... The sorting preprocessing, combined with the input modification of first vectorizing and flattening, and then linear projection mapping, ensures that elements that were originally adjacent in the vehicle dimension (corresponding to spatially nearby vehicle features) still maintain local clustering in the flattened long vector. The spatial interaction relationship is not destroyed, but is completely preserved in a compressed form. In addition, this input adaptation chain of "first sorting to preserve spatial semantics, then vectorizing and flattening, and finally linear projection to adapt to the interface" achieves compatibility between the three-dimensional traffic tensor and the Mamba one-dimensional sequence input paradigm while fully preserving the vehicle-level micro-motion information and its spatial interaction relationship. It has the advantages of strong input adaptability, high fidelity of micro-information, controllable parameters and computational overhead, and good compatibility with online reinforcement learning parallel training, thus helping to improve the stability of critical congestion identification and control decisions.

[0066] Step 2.2. Causal Temporal Filtering Unit: A causal one-dimensional convolution with a kernel size of 3 is used to locally filter the temporal embedding to suppress short-term random fluctuations and preserve the congestion evolution trend. The activation function is selected as follows: :

[0067] in, Indicates from step 1 to step 2. The historical input sequence of the step, =[ . For a one-dimensional convolution, where subscript This indicates that only one-dimensional convolution is taken at time t. The output of the convolutional layer; this convolutional layer acts as a learnable low-pass filter, designed to extract low-frequency trend components reflecting the propagation and evolution of congestion waves from high-frequency vehicle micro-motion data. The use of causal convolution ensures that feature extraction is based solely on historical information, meeting the real-time requirements of online signal control.

[0068] Step 2.3. Task-oriented gated cumulative state update unit: Implement state recursion based on adaptive gating weights dependent on input.

[0069] To address the core challenge of significantly reducing the recursive complexity of Mamba's state recursion while preserving its core input-dependent selectivity capability, making its mathematical structure compatible with the physical mechanism of progressive traffic congestion, and simultaneously improving the stability and convergence efficiency of PPO reinforcement learning training, we first attempted to stabilize training by introducing regularization techniques such as gradient pruning and strong weight decay while maintaining the complete selective SSM structure. The results showed that while gradient pruning suppressed gradient explosion to some extent, it also pruned weak early warning signals of congestion formation, leading to a significant lag in critical congestion identification. Subsequently, we tried replacing it with a simpler state-space model (such as S4D). While computational efficiency improved, the fixed state transition parameters of S4D resulted in the loss of Mamba's core "input-dependent selectivity" capability. The model could not dynamically adjust the retention or forgetting weights of historical information based on the urgency of the current traffic situation, leading to a significant decrease in the ability to identify congestion phase transitions in complex traffic scenarios.

[0070] Based on multiple experiments, a task-oriented gated cumulative state update mechanism was finally determined, starting from the physical mechanism of the "gradual accumulation" of traffic congestion (which is mathematically closer to a cumulative integral with input-dependent weights). The core calculation process of this unit is as follows: Step 2.3.1. Through Input dependency gating coefficients are generated to characterize the importance of the current traffic state to the writing of historical states; among them, It is the Sigmoid activation function. Representative to Linear projection is the initial transformation for feature extraction, used to reduce the dimensionality of the data and extract essential features. This indicates that another linear transformation is performed, which is a linear projection over time to capture dynamic evolution information.

[0071] This gating coefficient preserves input-dependent selectivity: it filters out historical evidence that truly pushes the system close to congestion thresholds, such as slow-moving diffusion, queue accumulation, and speed collapse—key congestion signs. In this embodiment, it utilizes... Instead of generating complete selective SSM input dependency parameters, it can significantly reduce structural complexity while retaining the selectivity of input dependencies, thus enabling effective screening of congestion symptoms.

[0072] Step 2.3.2. Through A gated cumulative state update mechanism is constructed to replace the original Mamba full selective state space recursion, thereby reducing training complexity and enhancing the stability of reinforcement learning training. Let the initial hidden state be... .

[0073] While a complete selective scan offers greater expressive power, it presents greater optimization challenges in online PPO training scenarios. Reinforcement learning inherently exhibits high variance and non-stationary sample distributions, making overly complex state recursion detrimental to stable training. Therefore, this embodiment utilizes "gated accumulated historical evidence" instead of "complete selective state space recursion." This is a core task-oriented modification centered around traffic phase transition control, preserving congestion accumulation capabilities while being more suitable for large-scale parallel PPO rollout and stable training, thus addressing the issue of needing to retain and gradually accumulate congestion evidence over a long period.

[0074] Step 2.3.3. Through The accumulated state is modulated at the output.

[0075] The lightweight gated modulation output method reduces readout complexity and improves the specificity of the output representation. Compared with more complex readout mechanisms, this structure is lighter and more stable during training.

[0076] This implementation retains the original Mamba causal modeling, input dependency adjustment, and linear time processing ideas. It reconstructs the original Mamba structure in a task-oriented manner around the phase transition control task of complex traffic networks. This significantly reduces the complexity of recursive transformations, effectively alleviates gradient instability and policy collapse problems in PPO training, and is highly consistent with the physical mechanism of gradual accumulation of traffic congestion in mathematical model, thus having better physical interpretability. At the same time, it provides a high-quality information foundation for subsequent parallel recursive units with multiple time constants.

[0077] Furthermore, the SM-PTC state-space sequence modeling backbone network, after the simplified gated cumulative state update unit, also includes a multi-time-constant parallel recursion unit.

[0078] This embodiment addresses the challenge that a single-time-constant state update mechanism cannot adequately account for the multi-timescale characteristics of congestion evolution. Firstly, it considers the issue from the perspective of filtering theory, specifically regarding the gating coefficient... A moving average filter is introduced at the input to suppress the interference of high-frequency fluctuations on long-term memory. However, the moving average is essentially a low-pass filter, which, while attenuating high-frequency noise, inevitably attenuates congestion precursor signals that are also in the relatively high-frequency range (such as the first significant drop in speed, the first incomplete dissipation of queues within a certain period, and other key short-term events), exhibiting the inherent limitation that "signal and noise in the same frequency band cannot be separated." Subsequently, the time constant is parameterized, letting... It can learn and allow the network to autonomously adjust the strength of memory. However, it was found that under the high variance and non-stationary distribution conditions of PPO reinforcement learning, the learnable time constant faces optimization difficulties: the time constant parameter is located in the exponential decay term, and its gradient decays or increases exponentially in high-order recursion, which can easily lead to gradient vanishing or gradient explosion; at the same time, the alternating optimization of policy update and value function update in PPO further exacerbates the instability of time constant parameter convergence.

[0079] Dynamics theory from state-space models shows that explicitly encoding structural priors with separated time scales is more stable and controllable than implicitly learning time constants from high-noise data. Further analysis of the critical phase transition dynamics of traffic congestion reveals that the evolution of congestion from free flow to saturated flow is a gradual cumulative process, mathematically described as a slow drift of the system state in phase space, superimposed with high-frequency oscillations at the signal periodic scale. During the critical transition of congestion, statistical indicators such as system variance and hysteresis autocorrelation exhibit different warning characteristics, with variance (rate of change) possessing independent warning value in bifurcation-induced phase transitions. This finding, from a nonlinear dynamics perspective, suggests that the macroscopic trend (level indicator) and mid-to-high frequency fluctuations (rate of change indicator) of the traffic system provide complementary and non-substitutable warning information at the critical state.

[0080] Inspired by this physical understanding, this invention starts with the recursive structure of the state-space model and constructs a fast and slow dual-time-constant parallel gated recursion and adaptive fusion mechanism to simultaneously characterize the intrinsic fluctuations in medium and high frequencies and the slow congestion accumulation trend. Specifically, the multi-time-constant parallel recursion unit includes: 1) Fast Tributary Gated Accumulation Module. In this implementation, a fast tributary is defined as follows: used to capture inherent fluctuations in medium- and high-frequency traffic flow—queue changes within a signal cycle, short-term speed oscillations, etc. These fluctuations are potential precursor signals for phase transitions in critical states and need to be detected promptly.

[0081] With fast branch time constant Output features of the task-oriented gated cumulative state update unit State recursion is performed to track high-frequency intrinsic fluctuations in traffic flow at the signal periodic scale, generating fast branch state. Its gating coefficient Simultaneously dependent on the phase transition risk index and mid-to-high frequency fluctuation intensity index : ,in The learnable parameters for the fast branch are: feature weights, slow-varying risk weights, fast-varying volatility weights, and a bias term; its state recursion formula is: .

[0082] Because fast branches not only need to sense fluctuations, but also need to know the risk conditions the system is in; under the same fluctuation amplitude, high The control implications of different operating conditions are different, therefore their gating coefficients are retained. Operating condition bias can be provided. Furthermore, to avoid the fast branch being excessively pulled by the slow-changing risk term, a "fluctuation-dominated, risk-corrected" constraint is applied to the fast branch gating during the calibration phase: under normalized input conditions, it is preferable to... The main operating conditions are established at that time; and for Small-range initial values ​​and regularization constraints are used to ensure that fast branches maintain mid-to-high frequency sensitivity and do not degenerate into slow trend channels; 2) Slow Branch Gated Accumulation Module. In this implementation, slow branches are defined as follows: used to preserve slow-accumulating congestion trends—net growth in queues over multiple cycles, continuous decline in average speed, etc. These trends are the core basis for judging whether the system is approaching a critical point and need to be preserved for a long time without being eroded by high-frequency fluctuations.

[0083] With slow branch time constant Output features of the task-oriented gated cumulative state update unit Perform state recursion, preserve the gradual accumulation trend across cycles, and generate slow branch states. Its gating coefficient Relying solely on the aforementioned phase transition risk index : ， in , , The learnable parameters for the slow branch are the feature weights, the slow-varying risk weights, and the bias term; its state recursion formula is... .

[0084] In the design of the gating guidance signal, to meet the different state update requirements of the fast and slow branches, a phase transition risk index is used. The degree to which the quantified system deviates from the free-flow ground state is used to reflect the slow-accumulating congestion trend on slow tributaries. This ensures that fast tributaries can capture the inherent fluctuations in traffic flow at the signal cycle scale, without relying on... pass Implicit learning (under the high variance conditions of PPO training, implicit learning faces convergence difficulties in multi-timescale features) innovatively designed a high-frequency fluctuation intensity index. As an explicit scalar fluctuation intensity indicator, direct injection into the gating provides a stable and physically interpretable guiding signal channel for characterizing queue fluctuations and short-term velocity oscillations within the signal period. During the traffic congestion phase transition, the system may be in a "high" state. Low (Continuously approaching the critical point but changing slowly) or "low" high The various combinations of states, such as "not yet close to saturation but fluctuating wildly," correspond to different stages of congestion evolution. Injecting both into the gating of fast and slow branches is precisely to capture the complexity of this multi-stage evolution.

[0085] Furthermore, in this embodiment, Based on the current cumulative number of vehicles and average speed The first-order difference weighted summation of adjacent decision steps yields: The parameters involved are determined according to the following logic: (1) This is a range clipping function that takes the input as an interval. Limited to Inside: when Output 0 when Output 1 in the first case, and output 1 in the other cases. This operation itself can suppress the destructive effects of outliers on gating and ensure... With gate function The input range is matched.

[0086] (2) Fluctuation Coupling Weighting Coefficient Controlling speed fluctuations and cumulative fluctuations separately The contribution intensity. To ensure that physical quantities of different dimensions have comparable scales before entering the gating, Pick , Pick This calibration allows for the maximum reasonable fluctuation (rate of change reaching) The cumulative change reached )hour, The theoretical contribution value is approximately After coupling, it remains in the middle of [0,1] to avoid premature gating saturation.

[0087] (3) Involved , Phase transition risk indicators Using the same set of observed vehicles ensures consistency in statistical methods between fluctuation intensity and risk level; Preferably, if the fluctuation intensity is captured based on the duration of a single signal cycle at the intersection, the standard deviation within a sliding window can be used to define it. : ;in, and Speed ​​and vehicle accumulation in the past Standard deviation within a sliding window for each decision step, window length The preferred value is 3-5 (corresponding to approximately 45-75 seconds based on a 15-second decision cycle, covering the typical timescale of queue formation and dissipation within a single signal cycle). Compared to the first-order difference version, the standard deviation version corresponds to the average power of the signal within the window in the frequency domain, and its focus on fluctuation scale can be determined by the window length. Both explicit control schemes are Effective implementation methods: The first-order difference version is lighter and has better real-time performance; the standard deviation version is more stable and has a clearer physical meaning, and can be selectively used in different engineering constraint scenarios.

[0088] In this embodiment, the values ​​of the fast and slow time constants are determined based on the correspondence between the traffic decision-making cycle and the physical time scale of congestion evolution, and are calibrated as follows: (1) Fast branch time constant Taking a larger value allows the fast branch to assign a higher update weight to recent inputs, thereby sensitively tracking queue fluctuations and short-term speed oscillations within the signal period. The preferred value range is ; (2) Slow branch time constant Taking a smaller value slows down the decay of the historical state of the slow branch, thus preserving the net accumulation of queues and the continuous drift trend of speed across cycles in the long term. The preferred value range is ; (3) Time scale separation constraint: satisfies This explicitly encodes the physical prior that "fast branches are more sensitive to mid-to-high frequency disturbances, while slow branches retain more evidence accumulated over a long period of time."

[0089] Analysis of the duality between traffic phase transition physics and gating mechanisms reveals that the core characteristic of critical slowing is that, near the phase transition point, the system's local recovery rate approaches zero (extremely slow dynamics), while increased variance leads to intensified short-term fluctuations (medium-to-high frequency dynamics). This dynamic characteristic of "coexistence of fast and slow" forms a dual mapping mathematically with the parallel recursion of the fast and slow branches—the time constant of the slow branch (…). Small, slow decay) corresponds to the slow recovery process of critical slowing, and the short time constant of the fast branch ( Large volume and rapid decay correspond to the need for rapid response after the fluctuations intensify. Explicitly encoding this physical prior into the model structure ensures that the model's behavior pattern is consistent with the physical mechanism of traffic congestion evolution in mathematical structure, thereby enhancing the model's perception ability and policy robustness in the critical region.

[0090] In a preferred embodiment, take At this time, the fast branch road is about Decision cycles (in) Decision cycle calculation The slow branch responds significantly to state changes within approximately [timescale], matching the queuing formation timescale within a single signal cycle; 1 decision cycle (approximately) A single effective state update can cover the net accumulation process of multiple signal cycles. This configuration structurally ensures the effectiveness of multi-timescale sensing.

[0091] 3) Adaptive fusion module: used to integrate the phase transition risk index Generate fusion coefficient , , and according to The fast branch state and the slow branch state are merged into the main state. ， ,in , To integrate the learnable parameters of the gating, so that When it rises Decrease.

[0092] Due to the gating of the fast and slow branches Fine-grained input selectivity has been completed (including...) (Perception of fluctuations) Only based on phase transition risk The decision is made regarding whether to rely more on the fast or slow branch. This design makes the physical causal chain of the fusion behavior clear and interpretable, avoiding the introduction of... Additional variables cause feature redundancy and optimization difficulties. Furthermore, the gated cumulative recursive formula... Essentially, it is a residual update with learnable gating, which is similar in mathematical structure to the GRU-style gating that replaces typical residual connections with gating update operations to improve gradient flow. Experiments have shown that it has good gradient flow properties in PPO training.

[0093] 4) Output modulation module: via The cumulative state is modulated to obtain the output features related to the current traffic control decision.

[0094] In this embodiment, each learnable parameter and gating fusion coefficient is preferably initialized first, and then optimized end-to-end through backpropagation of the composite reward function during PPO training.

[0095] In this embodiment, based on the task-oriented reconstruction of the original Mamba structure, the Mamba architecture is further modified, and the "historical evidence accumulation" stage is deepened and expanded into a fast and slow dual-path parallel recursive structure, making... It can simultaneously capture the intrinsic fluctuations in the mid-to-high frequency range of the signal at the periodic scale and the slow accumulation trend across the periodic scale, and use phase transition risk indicators. Guided fusion coefficient Achieving a dynamic balance between the contributions of the two branches. Thus, the SM-PTC backbone network has formed a complete and continuous processing chain, encompassing whole-network micro-trajectory input projection, local trend filtering, input-dependent gating generation, fast and slow dual-path historical evidence accumulation, risk-guided adaptive fusion, and output correlation modulation. This makes it more suitable for complex traffic networks with high disturbances, partial observability, and online reinforcement learning control scenarios.

[0096] Furthermore, in this embodiment, the activation function TeLU used by the gated cumulative state update unit and the multi-time constant parallel recursion unit in step 2 is defined as: ;

[0097] This implementation chooses TeLU to replace SiLU in the original Mamba dataset. This choice addresses issues such as "weak precursor overload, noise-induced false triggering, difficulty in balancing fast and slow scales with a single time constant, and large training jitter in PPO" after multiple rounds of modification attempts. It is a theoretically grounded, non-obvious function selection adapted to the specific distribution characteristics of "weak precursors and strong noise" in traffic phase change data. Specifically, under the same training budget and the same road network task, the following comparative attempts were conducted: (1) Directly use the original Mamba (retain SiLU and single-path recursion). The results show that in the weak signal stage before the formation of critical congestion, the value estimation fluctuates significantly, the strategy update is prone to "short-term improvement followed by regression", and the gating write is more sensitive to instantaneous noise, making it difficult to reliably capture congestion precursors in advance.

[0098] (2) Simply performing routine activation replacements (such as continuing to use SiLU family variants or adopting SwiGLU type gating) does not directly transfer the experience in language / visual supervision tasks to the benefit of this scenario; under high variance PPO closed loop, weak precursors remain unstable and have limited improvement on the contribution allocation of fast and slow branches in parallel recursion with multiple time constants.

[0099] (3) Although learning-based activation or stronger smooth activation schemes can suppress noise in local stages, problems such as "compression of precursor details and delayed response" and "increased coupling of training parameters leading to decreased convergence consistency" occur, failing to simultaneously meet the requirements of interpretability, real-time performance and stability.

[0100] After all the above approaches failed to achieve the desired results, it was finally decided to introduce TeLU in synergistically at both the pre-gating trend purification and output modulation stages, and to design it in conjunction with a multi-time-constant parallel recursive unit. Its technical adaptability is reflected in: (1) To address the contradiction of "weak precursors and strong noise", TeLU exhibits slower gradient decay and faster convergence in the negative region (see Figure 7 This can improve the retention of critical precursors in the backpropagation link and reduce the risk of early evidence being prematurely erased.

[0101] (2) To address the problem of “noise easily penetrating gating”, TeLU helps to suppress excessive writing driven by noise in the gating generation related path, making state accumulation more focused on the real congestion evolution trend and reducing false triggering.

[0102] (3) To address the problem that "a single time constant is difficult to take into account both fast and slow information", TeLU and fast-branch-slow-branch parallel recursion form a synergy in the multi-time-constant parallel recursion unit: the fast branch remains sensitive to mid-to-high frequency fluctuation evidence, while the slow branch remains stable to cross-cycle cumulative evidence, and the fusion state Multiscale distortion and mutual masking are suppressed.

[0103] (4) Regarding the issue of "unstable online training", TeLU is generally smooth and differentiable (see Figure 7 The replacement does not change the main topology and recursive interface, making it a low-intrusion modification; compared to SiLU, TeLU has both better positive activation strength and gradient propagation efficiency (see...). Figure 7 Under the conditions of high variance and non-stationary distribution in PPO online training, the policy update jitter is reduced and the convergence process is more continuous, and the improvement in stability significantly exceeds expectations.

[0104] In summary, this implementation method forms a three-stage gain chain through "pre-gated trend purification activation, post-gated output modulation activation, and parallel recursive collaboration of multiple time constants": the pre-stage improves the quality of input evidence, the mid-stage improves the stability of state updates and output modulation, and the post-stage improves the reliability of fast and slow scale information separation and fusion. These three stages collectively enhance the training stability, feature robustness, and control reliability in complex traffic network phase transition control tasks, achieving earlier identification of congestion signs, fewer noise-induced false triggers, and more robust cross-cycle congestion suppression techniques.

[0105] Furthermore, in this embodiment, the feature fusion method in step 2 involves combining the output features related to the current traffic control decision. Macroeconomic phase transition risk indicators After concatenation, the data is reduced in dimensionality by a two-layer fully connected network to output the policy input features, which are then used by the subsequent Actor-Critic network.

[0106] In this embodiment, the number of controlled intersections is set to be... , No. The set of selectable phases for each intersection is: In step 3, the joint signal control action is the Cartesian product of the phase sets of each controlled intersection:

[0107] In the decision-making step Output joint action: ;

[0108] in, For the number of controlled intersections, each component This represents the signal phase selected at intersection d in the current decision step. This joint action form can explicitly model cross-intersection coupling relationships, improving network-level coordinated control capabilities.

[0109] Considering that complex heterogeneous road networks simultaneously face three coupled problems under high load conditions—decreased traffic efficiency, increased control jitter, and accelerated risk accumulation—step 4 proposes a physics-inspired composite reward optimization approach and designs a unified reward function:

[0110] in, The revenue item is used to characterize the efficiency of traffic operation. The constraint term characterizes the cost of switching control actions. This is a risk penalty term that characterizes the risk of congestion evolution.

[0111] Furthermore, considering that the primary goal of traffic control is to "improve discharge capacity and reduce system congestion," and addressing the issue that using only a single throughput indicator might lead to localized worsening of time consumption or inventory buildup, this embodiment designs an efficiency term comprised of throughput, waiting time, and accumulated inventory:

[0112] The indicators are set according to the following logic: (1) Considering that the direct engineering goal of signal control is to improve the discharge capacity per unit time, it is proposed to complete the vehicle volume in the current decision step. As a primary indicator of positive returns, it is used to characterize immediate emission efficiency; (2) Considering that pursuing throughput alone may lead to phase assignment bias and increase vehicle delay, it is proposed to reduce the average system waiting time. Incorporate negative constraints to constrain the delay cost of the strategy; (3) Considering that the congestion phase transition is usually accompanied by a continuous increase in the number of vehicles in the network, a method for calculating the cumulative number of vehicles in the road network is proposed. Explicit penalties are used to prevent the formation of high stock levels in advance; By modeling the efficiency terms described above, a multi-dimensional efficiency evaluation mechanism for complex heterogeneous road networks is formed. This avoids the local imbalance problem caused by a single throughput-oriented approach and provides a stable revenue basis for subsequent cost control and risk mitigation.

[0113] In this embodiment, the efficiency term parameter Preferably, the parameters are determined through offline pairing scenario calibration. Specifically, under fixed observation interface, action interface, and disturbance protocol, the parameters are subjected to gridded search or segmented scaling based on the combination of traffic intensity and disturbance level scenarios, so that... Positive incentives and The negative constraint is at a comparable level in the early stages of training and ensures that throughput improvement does not come at the cost of waiting and significant deterioration of inventory.

[0114] In a preferred embodiment, the efficiency parameter may be: ;

[0115] This allows for a coordinated balance between "emission capacity enhancement, waiting constraints, and inventory suppression" under high-load disturbance conditions.

[0116] Furthermore, considering that high-frequency phase switching in multi-port joint control can cause execution jitter and increase control costs, the control cost item in step 4 includes a switching count penalty, where the switching count is:

[0117] in, This represents the number of intersections where a phase switch occurs within the current decision step.

[0118] To address the issue of frequent short-period transitions in the control sequence, a control action switching cost term was designed:

[0119] Regularization is applied to the number of action switching operations performed at each decision step, thereby reducing control jitter and improving phase switching smoothness.

[0120] In response to the evolutionary characteristic of congestion phase transition—"risk escalates first, congestion manifests later"—a risk feedforward penalty mechanism is proposed, and a risk penalty term for congestion evolution risk is designed:

[0121] in, As a phase transition risk indicator, the higher the risk, the greater the penalty, guiding intelligent agents to actively avoid high-risk states and delay the occurrence of congestion phase transitions; Through the above reward structure, synergistic optimization of efficiency improvement, smooth operation and risk suppression can be achieved under the same goal, thereby enhancing the control stability and deployability of complex transportation networks under high load disturbance conditions.

[0122] In this embodiment, the cost parameter is controlled. With risk penalty parameters The preferred method is to use an offline pairing calibration process. The calibration objective is to suppress high-frequency switching and enhance the ability to intervene early during the critical risk escalation phase without introducing overly conservative controls.

[0123] In a preferred embodiment, the following may be taken: ;

[0124] and with risk indicator parameters By combining these parameters, a consistent calibration scheme is formed, encompassing "efficiency terms, switching constraints, and risk penalty terms." Therefore, the specific parameters of the reward function can be written as follows:

[0125] Under the above parameter set, the reward function can suppress motion jitter and reduce the accumulation of phase transition risk while ensuring improved throughput.

[0126] like Figure 4 As shown, the present invention also provides a control system for implementing the control method described above. This system includes a road network interaction module, a state assessment module, an intelligent decision-making control module, and an online control and feedback module. In actual operation, the preferred model is deployed at the online control terminal, forming a complete "state acquisition" system. Status assessment Strategy Generation Action execution Log return "Model iteration" closed-loop control system.

[0127] Specifically, the road network interaction module interacts with the simulation environment or the real road network, and includes a data acquisition module and an action execution module. The data acquisition module collects the micro-operational status of vehicles and intersections within the controlled road network at preset decision cycles. The action execution module executes signal control actions and drives the traffic simulation environment or online control terminal to proceed according to the preset decision cycle. State assessment module: Based on the microscopic operating status of vehicles collected by the road network interaction module, constructs the microscopic trajectory tensor of the entire network. And calculate the phase transition risk index. Both are used together as the state part of the policy input; Intelligent decision control module: used to extract the temporal dependency features of the micro-trajectory tensor of the entire network based on the SM-PTC state-space sequence backbone network, and to correlate them with the phase transition risk index. The system integrates and outputs joint signal control actions for each controlled intersection through a policy-value joint network, and then distributes these joint signal control actions to the road network interaction module. The intelligent decision control module also includes a model optimization and iteration module, which performs large-scale parallel sampling during offline training, updates the parameters of the policy-value joint network based on historical observation data collected by the road network interaction module, and evaluates and optimizes the model on a validation set according to preset key performance indicators, storing the optimized model and its corresponding version information. The model optimization and iteration module also receives online operation logs from the online control and feedback module, performs incremental learning or fine-tuning of the optimized model, and achieves continuous iterative optimization of the model. Online control and feedback module: Deploy the optimal model at the online control terminal, collect detector data in real time, call the intelligent decision control module to perform strategy reasoning, send the generated signal control actions to the signal controller for execution, and record the online operation log (including status data, action data, reward data and system operation status), and send the online operation log back to the model optimization and iteration module at a preset period to form a closed-loop feedback.

[0128] This invention uses simulation experiments to verify the beneficial effects of the invention: This embodiment is based on the SUMO simulation platform, and the training and evaluation interface uses the standardized SUMO-RL. The main experimental road network adopts the DLR-TS Bologna Joined Network (heterogeneous topology, including standard intersections, roundabouts, and irregular multi-way intersections). The number of directly controlled signalized intersections is... The controller only operates on signal intersections and does not directly control structures without signals.

[0129] The training server is configured with dual Intel Xeon Gold 6148 CPUs (2.40 GHz, 40 physical cores / 80 threads), 128 GB of memory, and an NVIDIA GeForce RTX 4090 GPU. The number of parallel sampling environments is 72, and the main training interaction steps are approximately... The software environment uses the Ubuntu 18.04 operating system, and the deep reinforcement learning framework is implemented based on PyTorch 1.10.

[0130] (1) Experimental stress test scenario setup: To evaluate the control performance of the method of this invention under different traffic loads and disturbance intensities, a 5×3 stress test matrix was constructed (combining 5 flow levels with 3 disturbance levels) to cover the complete dynamic process from free flow to severe oversaturation. Each scenario was run independently with 5 random seeds. Specific parameters are shown in Table 1.

[0131] Table 1. Parameter Configuration Table for 5×3 Pressure Test Scenario Matrix

[0132] (2) Definition of evaluation indicators: This embodiment sets two types of evaluation indicators: basic operation indicators and transition-sensitive congestion indicators.

[0133] Basic operational indicators include: Total system throughput ( ( ): The total number of vehicles completing their journeys per unit of time, reflecting the overall traffic capacity of the road network; Average travel time ( ): The average time taken for all vehicles to complete their journey from departure to completion, measuring the efficiency of the road network; Average waiting time ( : The average time all vehicles wait at traffic lights, reflecting the impact of signal control on delays.

[0134] To quantify the degree of congestion evolution and system stability, two transition-sensitive congestion indicators are introduced: Composite Congestion Index (CCI) The overall congestion level of the road network is comprehensively characterized and defined as:

[0135] in, These represent the measured average travel time and average waiting time, respectively. This is the corresponding value under low load reference conditions. The larger the value, the more severely the system deviates from its stable state.

[0136] Sensitivity proxy indicators ( ): Reflecting the amplification effect of congestion susceptibility and instability, defined as:

[0137] This indicator adds waiting time weights to the overall congestion level to assess the risk of system instability under critical conditions.

[0138] All indicators were evaluated by collecting data every 15 seconds in a simulation environment (consistent with the signal decision cycle), and the average value was taken as the final evaluation result after multiple random seed experiments.

[0139] (3) Experimental setup and results analysis: The baseline was a fixed signal timing, serving as the benchmark for frame-level comparison. The engineering benchmark methods were Max-Pressure and PressLight. Results are detailed in Table 2. Figure 5 .

[0140] Table 2 Performance Comparison of Mainstream Engineering Benchmark Methods

[0141] The results in Table 2 verify the effectiveness of the overall framework of this invention compared to traditional non-learning methods. Combined with the significant advantages in throughput and travel time, this invention demonstrates that it achieves an effective balance between overall efficiency and risk mitigation.

[0142] Comparison of shared interfaces in deep reinforcement learning: Under the same PPO framework, only the feature extraction backbone was replaced, including LSTM (Long Short-Term Memory Network), Wang et al. (cooperative grouping reinforcement learning traffic signal control method), and Transformer. The results are detailed in Table 3. Figure 6 .

[0143] Table 3 Performance Comparison of Deep Reinforcement Learning Methods

[0144] The results show that, under the matching PPO framework, the throughput of this invention (664.80 veh) is improved by approximately 8.7%, 10.7%, and 12.0% compared with LSTM (Long Short-Term Memory Network), Wang et al. (Cooperative Grouping Reinforcement Learning Traffic Signal Control Method), and Transformer, respectively; and the average travel time (5.44 min) is reduced by approximately 8.1%, 9.8%, and 13.4%, respectively. and All indicators are significantly better, demonstrating the advantages of the SM-PTC backbone in long-term dependency extraction and its effective suppression of long-term congestion evolution.

[0145] Ablation experiments: To verify the synergistic effect of the overall technical solution of this invention, four sets of ablation experiments were designed under extreme load scenarios (λ=2.0, high perturbation): Efficiency Only: Remove risk penalty items and switching constraints, retain only efficiency items. ; Remove risk penalty items: only Remove from the reward function; Remove toggle regular expression: Only remove toggle constraints .

[0146] Table 4 Ablation Experiment Results

[0147] As shown in Table 4, while the efficiency term alone can maintain approximately the same throughput, the waiting time and congestion agent volume deteriorate significantly; performance drops significantly after removing the risk term; and the system degrades most severely after removing the switching regularization term. These results indicate that the technical effectiveness of this invention stems from a synergistic mechanism of "efficiency term + control cost term + risk penalty term," rather than being driven by a single efficiency objective.

[0148] The experimental results above show that the key to this invention lies in the overall collaborative mechanism. By coupling the SM-PTC state-space sequence modeling backbone network with the phase transition risk index through dual pathways, combining joint action switching constraints and complex network closed-loop optimization into the same control framework, a comprehensive technical effect improvement that can be repeatedly verified under high-voltage heterogeneous road networks is achieved, which has significant engineering practical value and a high degree of creativity.

[0149] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the traffic signal control method for suppressing critical congestion in complex traffic networks as described above.

[0150] The present invention also provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the traffic signal control method for suppressing critical congestion in complex traffic networks as described above.

[0151] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0152] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0153] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A traffic signal control method for suppressing critical congestion in complex traffic networks, characterized in that, The method includes the following steps: Step 1. Collect traffic observation data of the controlled traffic network at the current decision moment, and construct the micro-trajectory tensor of the entire network. ; Calculate the phase transition risk index to characterize the risk of congestion evolution. ; Step 2. Convert the microscopic trajectory tensor of the entire network Inputting the SM-PTC state-space sequence to model the backbone network, extracting temporal dependency features, and comparing them with phase transition risk indicators. Feature fusion is performed to form policy input features; the SM-PTC state-space sequence modeling backbone network includes features for... The system includes an input mapping unit mapped to the latent space, a causal temporal filtering unit for local filtering of temporal embeddings, a task-oriented gated accumulation state update unit, and a multi-time constant parallel recursion unit; wherein the multi-time constant parallel recursion unit includes a fast branch gated accumulation module, a slow branch gated accumulation module, an adaptive fusion module, and an output modulation module. Step 3. Feed the policy input features into the policy-value joint network and output the joint signal control action for each controlled intersection; Step 4. Optimize and update the policy-value joint network based on a composite reward function that includes efficiency, control cost, and risk penalty terms, and impose constraints on action switching; Step 5. Repeatedly execute the action delivery, environmental feedback acquisition, and policy update to obtain the trained policy network; Step 6. Based on the trained SM-PTC state-space sequence modeling backbone network and policy network, perform real-time signal control at the controlled intersection.

2. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 1, characterized in that, Step 1 determines the active vehicle set based on the controlled intersection's influence radius. At each decision time, the micro-motion state of the current active vehicle set is collected, including normalized position, speed, acceleration, heading angle, and lane index. Then, the individual micro-feature vectors of the active vehicles are extracted and sorted according to priority based on the vehicles being closer to the stop line and having lower speeds, resulting in a single-time vehicle-feature matrix. Subsequently, they are stacked in chronological order to construct the microscopic trajectory tensor of the entire network. .

3. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 2, characterized in that, The phase transition risk index mentioned in step 1 It is obtained by weighted coupling of density and velocity components, and the expression is: ； in, This represents the current cumulative number of vehicles. The saturation threshold The current average speed, As a reference speed threshold, These are the weighting coefficients.

4. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 3, characterized in that, The input mapping unit in step 2 is used to map the microscopic trajectory tensor of the entire network. Flatten along the time dimension and map it to the latent space through a linear projection layer, then cut it into input branches. Gated branches ; The causal temporal filtering unit employs a causal one-dimensional convolution with a kernel size of 3 to embed the temporal branch, i.e., the input branch. Local filtering is performed, followed by nonlinear activation. The activation function is selected as follows: .

5. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 4, characterized in that, The task-oriented gated cumulative state update unit described in step 2 uses an adaptive gating weight based on input dependency to achieve state recursion, completing the selective writing and state accumulation of key historical evidence. When processing data, it first... Generate input-dependent gating coefficients; where, It is the Sigmoid activation function. Representative to Linear projection is the initial transformation for feature extraction, used to reduce the dimensionality of the data and extract essential features. This indicates that another linear transformation is performed, which is a linear projection over time to capture dynamic evolution information. The low-frequency trend component is the output of the causal time-domain filter unit. This is the gating factor; then through Construct a gated cumulative state update mechanism; among which, The hidden states at adjacent time points; finally, through The accumulated state is modulated at the output; where, The output features of the task-oriented gated cumulative state update unit are provided.

6. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 5, characterized in that, The fast branch gated accumulation module described in step 2 uses the fast branch time constant Output features of the task-oriented gated cumulative state update unit State recursion is performed to track high-frequency intrinsic fluctuations in traffic flow at the signal periodic scale, generating fast branch state. Its gating coefficient Simultaneously dependent on phase transition risk indicators and mid-to-high frequency fluctuation intensity index : ,in The learnable parameters for the fast branch are feature weights, slow-varying risk weights, fast-varying volatility weights, and bias terms. Its state recursion formula is: ; The slow branch gating accumulation module uses the slow branch time constant Output features of the task-oriented gated cumulative state update unit Perform state recursion, preserve the gradual accumulation trend across cycles, and generate slow branch states. Its gating coefficient Relying solely on phase transition risk indicators : ,in , , These are the learnable parameters for the slow branch, namely feature weights, slow-varying risk weights, and bias terms; Its state recursion formula is: ; The adaptive fusion module is used based on phase transition risk indicators. Generate fusion coefficient , , and according to The fast branch state and the slow branch state are merged into the main state. ， ,in , All are fusion-gated learnable parameters; The output modulation module passes through Output modulation is performed on the accumulated state to obtain the output features related to the current traffic control decision. .

7. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 6, characterized in that, The expression for the composite reward function in step 4 is: ； in, The revenue item is used to characterize the efficiency of traffic operation. The constraint term characterizes the cost of switching control actions. To characterize the risk penalty term representing the risk of congestion evolution, In the decision-making step The combined action of the output.

8. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 7, characterized in that, The control system for implementing the control method includes a road network interaction module, a state evaluation module, an intelligent decision control module, and an online control and feedback module. The road network interaction module interacts with the real road network and includes a data acquisition module and an action execution module. The data acquisition module collects the microscopic operating status of vehicles and intersections within the controlled road network at preset decision cycles. The action execution module executes signal control actions and drives the online control terminal to proceed according to the preset decision cycle. The state evaluation module is used to construct a network-wide micro-trajectory tensor based on the vehicle micro-operational states collected by the road network interaction module. And calculate the phase transition risk index. Both are used together as the state part of the policy input; The intelligent decision-making and control module is used to extract the temporal dependency features of the micro-trajectory tensor of the entire network based on the SM-PTC state-space sequence backbone network, and to correlate them with the phase transition risk index. The system integrates and outputs joint signal control actions for each controlled intersection through a policy-value joint network, and then distributes these joint signal control actions to the road network interaction module. The intelligent decision control module also includes a model optimization and iteration module, which performs large-scale parallel sampling during offline training, updates the parameters of the policy-value joint network based on historical observation data collected by the road network interaction module, and evaluates and optimizes the model on the validation set according to preset key performance indicators, storing the optimal model and its corresponding version information. The model optimization and iteration module also receives online operation logs from the online control and feedback module to achieve continuous iterative optimization of the model.

9. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 8, characterized in that, Step 2 obtains the output features relevant to the current traffic control decision. Subsequently, with phase transition risk indicators The data is concatenated and then subjected to dimensionality reduction via two fully connected layers to output the policy input features. High-frequency fluctuation intensity index in step 2 Based on the current cumulative number of vehicles and current average speed Obtained by first-order differential weighting in adjacent decision steps; or by capturing the intensity of mid-to-high frequency fluctuations based on the duration of a single signal cycle at the intersection.

10. The traffic signal control method for suppressing critical congestion in complex traffic networks according to claim 9, characterized in that, Step 4: Revenue Items Characterizing Traffic Operation Efficiency The expression is: ； in, To complete the current decision-making step in terms of vehicle volume, This represents the average waiting time. To accumulate the number of vehicles on the road network, For efficiency parameters; Constraints representing the cost of switching control actions ;in, This represents the number of intersections where a phase switch occurs within the current decision step. To control cost parameters; Risk penalty item representing the risk of congestion evolution ;in, As a phase transition risk indicator, This is a risk penalty parameter.

Images