Traffic signal regulation and control system based on dynamic weight fusion and phase collaborative optimization

Abstract

The invention belongs to the field of intelligent traffic systems, particularly relates to a traffic signal regulation and control system based on dynamic weight fusion and phase collaborative optimization, and aims to solve the problem of adaptability difference during traditional fixed matching. A multi-sensor data acquisition module in the system acquires multi-dimensional original traffic data of each road section of a target road in real time and transmits the multi-dimensional original traffic data to a dynamic weight fusion module. The dynamic weight fusion module receives multi-dimensional original traffic data, obtains a traffic state estimation value through the processing of a dynamic weight adaptive fusion algorithm, and transmits the traffic state estimation value to the phase collaborative optimization genetic algorithm module. The phase collaborative optimization genetic algorithm module optimizes proportion, integral and differential coefficients through an improved genetic algorithm based on the traffic state estimated value, outputs an optimal proportion-integral-differential parameter to the proportion-integral-differential control module, generates a timing instruction of a dynamic traffic signal lamp and transmits the timing instruction to the signal output module; and the signal output module receives a timing instruction to realize dynamic regulation and control of the traffic signal lamp.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems, and in particular to a traffic signal control system based on dynamic weight fusion and phase collaborative optimization. Background Technology

[0002] Traffic signal control strategies at urban road intersections are a core factor affecting regional traffic efficiency and safety. Traditional traffic signal control methods can be mainly divided into two categories: fixed timing and simple adaptive control. Fixed timing schemes are preset based on historical statistical data and cannot respond to real-time dynamic changes in traffic flow, easily leading to increased congestion during peak hours and ineffective green lights during off-peak hours. Simple adaptive control (such as inductive control) can adjust the duration of a phase based on the triggering of a single detector (such as a geomagnetic coil), but it lacks deep fusion and global optimization of multi-source heterogeneous traffic data, making it difficult to cope with complex and ever-changing traffic scenarios.

[0003] To improve control accuracy, existing technologies have developed control schemes that combine multi-sensor data fusion with intelligent optimization algorithms. However, at the data fusion level, these schemes often rely on traditional Kalman filtering or a fixed combination of it and neural networks. The former struggles to effectively handle the inherent nonlinearity and non-Gaussian noise of traffic systems; while the latter incorporates neural networks, the weights of each processing unit (e.g., denoising, prediction) are fixed, making it impossible to dynamically adjust the fusion strategy according to different traffic conditions (e.g., congestion, off-peak hours, severe weather), leading to decreased estimation accuracy and insufficient robustness in specific scenarios. At the control parameter optimization level, heuristic algorithms such as genetic algorithms are often used to tune PID parameters, but these commonly suffer from problems such as parameter encoding not considering collaborative constraints between multiple phases, rigid fitness function design unable to respond to sudden changes in traffic flow, and iteration processes not taking real-time requirements into account. These issues easily lead to timing scheme conflicts, slow convergence speeds, and difficulty in achieving fast and accurate online optimization. Summary of the Invention

[0004] The purpose of this invention is to provide a traffic signal control system based on dynamic weight fusion and phase collaborative optimization, which aims to solve the problems of poor adaptability, low control accuracy, slow convergence, and insufficient ability to cope with sudden traffic changes in traditional fixed timing, simple adaptive control and conventional algorithm combination schemes.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a traffic signal control system based on dynamic weight fusion and phase collaborative optimization, comprising: a multi-sensor data acquisition module for deploying various sensors to collect multi-dimensional raw traffic data of each segment of a target road in real time and transmit it to a dynamic weight fusion module; the multi-dimensional raw traffic data includes at least traffic flow, vehicle speed, queue length, and weather conditions; a dynamic weight fusion module for receiving the multi-dimensional raw traffic data, processing it through a dynamic weight adaptive fusion algorithm to obtain traffic state estimates, and transmitting them to a phase collaborative optimization genetic algorithm module; wherein, the dynamic weight adaptive fusion algorithm includes data preprocessing, dynamic weight allocation, nonlinear compensation, noise modeling, and model correction; the dynamic weight fusion module has multiple built-in fusion units, each of which implements nonlinear compensation, noise modeling, model correction, and basic data fusion functions respectively; the dynamic weight allocation dynamically adjusts the weight of each fusion unit based on the real-time traffic state level; and a phase collaborative optimization genetic algorithm module, based on traffic conditions... The system estimates the traffic state and optimizes the proportional-integral-derivative (PID) controller's proportional, integral, and derivative coefficients using an improved genetic algorithm. It outputs the optimal PID parameters to the PID controller and simultaneously generates dynamic traffic light timing commands, transmitting them to the signal output module. The improved genetic algorithm includes a phase-coordinated coding mechanism, a dynamic fitness function, a fitness-enhancing elimination selection mechanism, gene infection and crossover operations, an adaptive mutation operator, and real-time iteration termination constraints. The PID controller receives the optimal PID parameters and, combined with the traffic state estimates, constructs an adaptive PID control model to generate real-time signal control parameters. The signal output module receives the timing commands, converts them into traffic light control signals, and sends them to the traffic light actuator to achieve dynamic traffic light control. Simultaneously, it collects the controlled traffic data in real-time and feeds it back to the multi-sensor data acquisition module.

[0006] The traffic signal control system also includes a server and a three-level feedback module. The server is used to store multi-dimensional raw traffic data, preprocessed data, traffic state estimates, proportional-integral-derivative controller parameters, genetic algorithm optimization process data, historical timing schemes, and a scenario-timing strategy mapping library. The three-level feedback module realizes dynamic adjustment of sensor sampling frequency, optimization of proportional-integral-derivative parameter pool, and rapid matching of timing strategies by constructing a three-level feedback architecture of data layer, parameter layer, and strategy layer.

[0007] The multi-sensor data acquisition module deploys sensors including geomagnetic sensors, video sensors, meteorological sensors, and traffic flow detectors. The sensor deployment covers key sections of the target road, and the sampling frequency of each sensor is dynamically adjusted through data layer feedback to balance data real-time performance and acquisition efficiency.

[0008] The dynamic weight fusion module includes a multilayer perceptron neural network for nonlinear compensation, a convolutional neural network for noise modeling, a recurrent neural network for model correction, and a native Kalman filter framework for basic data fusion.

[0009] The dynamic weight fusion module is specifically configured as follows: Data preprocessing—timestamp synchronization detection, missing data imputation, and scale unification processing are performed on multi-dimensional raw traffic data to obtain standardized data; traffic state levels are classified based on road occupancy in the standardized data, and meteorological data is combined to determine whether there are significant non-Gaussian noise scenarios; Dynamic weight allocation—the weight ratios of the multilayer perceptron neural network, convolutional neural network, recurrent neural network, and Kalman filter native framework are dynamically adjusted according to traffic state levels and noise scenario types; Nonlinear compensation—nonlinear compensation is achieved by learning the nonlinear transfer function of the multilayer perceptron neural network on the multi-dimensional raw traffic data after data preprocessing; Noise modeling—non-Gaussian noise distribution characteristics are learned through autoencoder denoising preprocessing and convolutional neural network, and the noise covariance matrix and observation noise covariance matrix of Kalman filter are dynamically adjusted; Model calibration—the drift law of the traffic system model parameters is learned in real time through recurrent neural network, the state transition matrix and observation matrix of Kalman filter are dynamically updated, and the traffic state estimate including traffic flow density, optimal traffic speed, queue length prediction value, and traffic flow mutation coefficient is output.

[0010] The phase-coordinated optimization genetic algorithm module is specifically configured as follows: The proportional-integral-derivative (PID) controller's proportional coefficient, integral coefficient, and derivative coefficient, along with the green light duration constraints for each phase, are jointly encoded into chromosome individuals to construct an initial population. Phase priority factors are embedded within these chromosome individuals. A traffic flow mutation coefficient is introduced, and the weight coefficients of the fitness function are dynamically adjusted with the optimization objective of minimizing average vehicle waiting time and maximizing traffic efficiency. A fitness-enhancing elimination rule is used for selection, screening high-quality individuals for the next generation. A gene infection crossover method is used for crossover, transmitting gene fragments from the optimal individual to offspring individuals. An adaptive mutation operator is used to dynamically adjust the mutation probability based on the population fitness distribution. The optimal individual from each generation is retained without participating in crossover or mutation operations. When the preset convergence condition is met, the maximum number of iterations is reached, or the fitness value reaches an acceptable threshold, iteration stops and the optimal PID parameters are output to the PID control module. Based on the optimal PID parameters and traffic state estimates, the passage time for each phase of the traffic light is dynamically calculated, timing instructions are generated, and transmitted to the signal output module.

[0011] The proportional-integral-derivative (PI-DE) control module is specifically configured to: receive the optimal PI-DE parameters, combine them with the traffic flow density, optimal speed, queue length prediction, and traffic flow mutation coefficient output by the dynamic weight fusion module, construct an adaptive PI-DE control model, and generate real-time signal control parameters adapted to the current traffic scenario.

[0012] The signal output module uses a wireless communication protocol to transmit control signals and supports real-time data interaction with traffic signal actuators. The signal output module integrates a data feedback unit, which collects traffic flow, vehicle speed and queue length data after regulation in real time, providing data support for the three-level feedback module.

[0013] The three-level feedback module is specifically configured as follows: Data layer feedback—extracting the deviation characteristics between the data after regulation and the fused data, and dynamically adjusting the sensor sampling frequency; Parameter layer feedback—periodically statistically analyzing the optimization effect of proportional-integral-derivative parameters, and constructing a dedicated parameter pool for different traffic scenarios; Strategy layer feedback—based on historical data stored on the server, when it is detected that the current scenario and the historical scenario meet the preset matching degree, calling the historical optimal timing scheme as the initial value for local optimization.

[0014] Each module is connected to the server via communication links to form a closed-loop control system. Specifically, the multi-sensor data acquisition module establishes bidirectional data transmission links with the dynamic weight fusion module and the signal output module; the dynamic weight fusion module establishes a unidirectional data transmission link with the phase collaborative optimization genetic algorithm module; the phase collaborative optimization genetic algorithm module establishes a unidirectional data transmission link with the proportional-integral-derivative control module and the signal output module; the proportional-integral-derivative control module establishes a unidirectional data transmission link with the signal output module; the server establishes bidirectional data transmission links with each module; and the three-level feedback module establishes bidirectional data transmission links with the multi-sensor data acquisition module, the server, and the phase collaborative optimization genetic algorithm module.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application provides a traffic signal control system based on dynamic weight fusion and phase collaborative optimization. By dynamically adapting the weights to the scenario, the system adjusts the weight ratios of multilayer perceptron neural network (MLP, nonlinear compensation), convolutional neural network (CNN, noise modeling), recurrent neural network (RNN, model correction), and Kalman filter (basic fusion) in real time according to the traffic state level (congestion / slow traffic / off-peak) and noise scenario (significant / insignificant non-Gaussian noise). For example, MLP accounts for 60% to enhance nonlinear processing during congestion, and CNN accounts for 50% to enhance noise suppression during rainfall. This solves the problem of the incompatibility between noise robustness and nonlinear adaptation.

[0016] 2. This application provides a traffic signal control system based on dynamic weight fusion and phase collaborative optimization. Through multiple innovations in the phase collaborative optimization genetic algorithm, it significantly improves optimization efficiency and adaptability. Furthermore, through a three-layer feedback architecture of data layer, parameter layer, and strategy layer, it achieves dynamic optimization across the entire link, significantly enhancing system robustness. Attached Figure Description

[0017] Figure 1 This is a structural diagram of a traffic signal control system based on dynamic weight fusion and phase collaborative optimization provided in an embodiment of this application; Figure 2 This is a schematic diagram of a sensor setup provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 This application provides a traffic signal control system based on dynamic weight fusion and phase collaborative optimization, comprising: The multi-sensor data acquisition module 1 is used to deploy various sensors to collect multi-dimensional raw traffic data of each segment of the target road in real time and transmit it to the dynamic weight fusion module. The multi-dimensional raw traffic data includes at least traffic flow, vehicle speed, queue length, and weather conditions.

[0020] An exemplary multi-sensor data acquisition module 1 deploys sensors including a geomagnetic sensor, a video sensor, a meteorological sensor, and a traffic flow detector. The sensors cover key sections of the target road. The sampling frequency of each sensor is dynamically adjusted via data layer feedback to balance data real-time performance and acquisition efficiency. The geomagnetic sensor detects vehicle passage and acquires traffic flow data; the video sensor supplements the data with pedestrian flow, vehicle type, and queue length information; the meteorological sensor collects meteorological data such as rainfall and visibility; and the traffic flow detector monitors traffic flow at each approach lane in real time.

[0021] As one possible implementation, geomagnetic sensors are installed 5-15m before the lane stop line, video sensors are deployed on traffic light poles at intersections, meteorological sensors are installed in the median strip or green belt at intersections, and traffic flow detectors are deployed at each approach lane. The sampling frequency of each sensor is initially set to 0.5-1 second / time, which can be increased to 0.3 seconds / time under congested conditions and decreased to 1 second / time under off-peak conditions to ensure a balance between data real-time performance and acquisition efficiency.

[0022] The dynamic weight fusion module 2 receives multi-dimensional raw traffic data, processes it using a dynamic weight adaptive fusion algorithm to obtain traffic state estimates, and then transmits these estimates to the phase collaborative optimization genetic algorithm module. The dynamic weight adaptive fusion algorithm includes data preprocessing, dynamic weight allocation, nonlinear compensation, noise modeling, and model calibration. The dynamic weight fusion module 2 incorporates multiple fusion units, each corresponding to nonlinear compensation, noise modeling, model calibration, and basic data fusion functions. Dynamic weight allocation dynamically adjusts the weights of each fusion unit based on the real-time traffic state level.

[0023] For example, each fusion unit of the dynamic weight fusion module 2 includes a multilayer perceptron (MLP) neural network for nonlinear compensation, a convolutional (CNN) neural network for noise modeling, a recurrent (RNN) neural network for model correction, and a Kalman filter native framework for basic data fusion.

[0024] The dynamic weight fusion module 2 is specifically configured with the following five steps: Data preprocessing: Multi-dimensional raw traffic data undergoes timestamp synchronization detection, missing data imputation, and scale unification to obtain standardized data. Traffic state levels are categorized based on road occupancy rates in the standardized data, and meteorological data is used to determine the presence of significant non-Gaussian noise scenarios. Timestamp synchronization detection calculates the interval between adjacent timestamps in the multi-dimensional raw traffic data. When the interval exceeds a preset threshold, missing data is imputed using linear interpolation. The imputed data is then scaled using a min-max normalization method to obtain standardized data. Based on road occupancy rates in the standardized data, a road occupancy rate greater than 60% is classified as congested, between 30% and 60% as slow-moving, and less than 30% as off-peak. Meteorological data is also used to determine the presence of significant non-Gaussian noise scenarios, such as rainfall, snowfall, or sensor obstruction.

[0025] Dynamic weight allocation: Based on traffic condition levels and noise scene types, the weight proportions of the multilayer perceptron neural network (MLP), convolutional neural network (CNN), recurrent neural network (RNN), and the Kalman filter native framework are dynamically adjusted. For example, in congested conditions, the weight proportion of the MLP neural network is adjusted to 60%, the CNN network to 20%, the RNN network to 15%, and the Kalman filter native framework to 5%. In scenarios with significant non-Gaussian noise, the weight proportion of the MLP neural network is adjusted to 30%, the CNN network to 50%, the RNN network to 15%, and the Kalman filter native framework to 5%. In off-peak conditions, the weight proportion of the MLP neural network is adjusted to 20%, the CNN network to 10%, the RNN network to 10%, and the Kalman filter native framework to 50%.

[0026] Nonlinear compensation: Nonlinear compensation is achieved by learning the nonlinear transfer function of the preprocessed multi-dimensional raw traffic data through a multilayer perceptron neural network (MLP). The MLP replaces the Taylor expansion linearization step of the traditional extended Kalman filter, avoiding the accumulation of linearization errors. For example, the MLP neural network has three hidden layers, uses ReLU as the activation function, and the number of hidden layer nodes is 2-3 times the input dimension.

[0027] Noise Modeling: Denoising preprocessing using an autoencoder and learning non-Gaussian noise distribution characteristics via a convolutional neural network dynamically adjusts the noise covariance matrix Q and the observation noise covariance matrix R of the Kalman filter. Referring to the technical manuals of geomagnetic sensors, video sensors, etc., their inherent noise levels (such as measurement error range and signal fluctuation threshold) are extracted. This error is quantized and used as the initial element of the R matrix (reflecting the noise intensity in the observation stage). Based on traffic engineering principles, the natural fluctuations in traffic flow (such as small changes in traffic flow during off-peak hours) have an inherent range. This fluctuation amplitude is converted into process noise intensity and used as the initial value of the Q matrix (reflecting the random changes within the traffic system itself). Initially, both the Q and R matrices are diagonal matrices. The diagonal elements correspond to the noise intensity of each traffic data dimension (such as flow rate, vehicle speed, and queue length), while the off-diagonal elements are set to 0 (assuming each dimension's noise is independent). Subsequently, noise features learned through the CNN are used to adjust the off-diagonal elements of the noise covariance matrix Q and the observation noise covariance matrix R in real time based on these learned noise features.

[0028] Model calibration: A recurrent neural network (RNN) is used to learn the drift patterns of the traffic system's model parameters in real time, dynamically updating the state transition matrix A and observation matrix H of the Kalman filter. The output includes traffic flow density, optimal speed, predicted queue length, and traffic flow mutation coefficients, providing a traffic state estimate. For example, the time step of the RNN is set to 5, and the iteration step size is consistent with the sensor sampling frequency.

[0029] Phase-cooperative optimization genetic algorithm module 3, based on traffic state estimates, optimizes the proportional, integral, and derivative coefficients of the proportional-integral-derivative (PID) controller contained in the proportional-integral-derivative (PID) control module 4 using an improved genetic algorithm. It outputs the optimal PID parameters to the PID control module 4 and simultaneously generates dynamic traffic light timing commands, transmitting them to the signal output module 5. The improved genetic algorithm includes a phase-cooperative coding mechanism, a dynamic fitness function, a fitness-enhancing elimination selection mechanism, gene infection and crossover operations, an adaptive mutation operator, and real-time iteration termination constraints.

[0030] For example, the phase-coordinated optimization genetic algorithm module 3 is specifically configured as follows: the proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd of the proportional-integral-derivative (PID) controller, along with the green light duration constraints for each phase, are jointly encoded into chromosome individuals to construct an initial population of 30-50. Kp values ​​range from [0.1, 5.0], Ki values ​​range from [0.01, 1.0], and Kd values ​​range from [0.05, 2.0]. Each chromosome individual embeds a phase priority factor, including the lower limit of the green light duration for the main traffic direction and the phase switching interval constraint. A traffic flow mutation coefficient γ (γ = current queue length / historical average queue length) is introduced, and the weight coefficients of the fitness function are dynamically adjusted with the optimization objective of minimizing the average vehicle waiting time and maximizing traffic efficiency. When γ≤1.5 (no mutation), the fitness function F=1 / (α・T+β・L), where α=0.6 and β=0.4; when γ>1.5 (mutation), the fitness function F=1 / (α・T+β・L), where α=0.3 and β=0.7; where T is the average vehicle waiting time and L is the average queue length, and the fitness value is positively correlated with the control effect.

[0031] A fitness-enhancing elimination algorithm is used for selection. Building upon roulette wheel selection, a forced elimination mechanism for low-fitness individuals is added to select superior individuals for the next generation. A gene infection crossover method is used for crossover, with a gene fragment from the optimal individual serving as the "infectious gene," which is passed to offspring during crossover to ensure a steady increase in average fitness during evolution. The crossover probability is set at 0.6-0.8. An adaptive mutation operator is employed to dynamically adjust the mutation probability based on the population fitness distribution (initial mutation probability set at 0.01-0.05) to prevent premature convergence. Preservation... Each generation of the optimal individual does not require crossover or mutation operations; when the preset convergence condition is met, the maximum number of iterations (50-100 times) is reached, or the fitness value reaches an acceptable threshold (average vehicle waiting time ≤ 30 seconds), the iteration stops and the optimal proportional-integral-derivative parameters (proportional coefficient Kp, integral coefficient Ki, derivative coefficient Kd) are output to the proportional-integral-derivative control module; based on the optimal proportional-integral-derivative parameters and traffic state estimates, the green light duration, red light duration, and yellow light duration of each phase of the traffic light are dynamically calculated, timing instructions are generated and transmitted to the signal output module 5.

[0032] In some embodiments, the maximum number of generations of the phase-cooperative optimization genetic algorithm is set to 50-100, the initial crossover probability is set to 0.7, and the initial mutation probability is set to 0.03; when the population fitness variance is less than a preset threshold, it is determined to be converged.

[0033] The proportional-integral-derivative (PI-DI) control module 4 is used to receive the optimal PI-DI parameters, combine them with traffic state estimates to construct an adaptive PI-DI control model, and generate real-time signal control parameters.

[0034] For example, the proportional-integral-derivative control module 4 is specifically configured to: receive the optimal proportional-integral-derivative parameters, combine them with the traffic flow density, optimal traffic speed, queue length prediction value and traffic flow mutation coefficient output by the dynamic weight fusion module, construct an adaptive proportional-integral-derivative control model, and generate real-time signal control parameters adapted to the current traffic scenario.

[0035] The timing instructions generated by the phase-coordinated optimization genetic algorithm module 3 focus on the macro-level allocation of travel time (such as the green light duration range for each phase). Real-time signal control parameters further clarify details such as signal light switching timing, duration fine-tuning step size, and phase transition rhythm, ensuring the accurate implementation of timing instructions. By combining traffic flow density, queue length prediction, and traffic flow mutation coefficient from traffic state estimates, control parameters are adjusted in real time (such as accelerating the green light duration adjustment response speed when traffic flow changes abruptly), enabling signal light control to closely follow the dynamic fluctuations of traffic flow and avoiding a disconnect between the timing scheme and actual traffic conditions.

[0036] The signal output module 5 is used to receive timing instructions, convert the timing instructions into traffic light control signals and send them to the traffic light actuator to realize dynamic control of traffic lights. At the same time, it collects traffic data after control in real time and feeds it back to the multi-sensor data acquisition module 1.

[0037] For example, signal output module 5 converts timing instructions into traffic light control signals and sends them to the traffic light actuator to achieve dynamic control of traffic lights. A more specific traffic light timing scheme is as follows: When traffic is congested (occupancy rate > 60%), the green light duration for the main traffic direction is extended to 50-60 seconds, and for the secondary direction it is shortened to 25-35 seconds. When traffic is slow (30% ≤ occupancy rate ≤ 60%), the traffic light cycle length is set to 100-120 seconds, and the green light duration for each direction is allocated according to the traffic flow ratio. During off-peak hours (occupancy rate <30%), the traffic light cycle length is set to 80-100 seconds, and the green light duration for each direction is allocated according to the traffic flow ratio. During rainy weather, the yellow light duration is extended to 5 seconds.

[0038] For example, the traffic signal control system also includes a server 6 and a three-level feedback module 7. The server is used to store multi-dimensional raw traffic data, preprocessed data, traffic state estimates, proportional-integral-derivative (PID) controller parameters, genetic algorithm optimization process data, historical timing schemes, and a scenario-timing strategy mapping library. The three-level feedback module, through the construction of a three-level feedback architecture consisting of a data layer, a parameter layer, and a strategy layer, enables dynamic adjustment of sensor sampling frequency, optimization of the PLD parameter pool, and rapid matching of timing strategies.

[0039] As one possible implementation, the signal output module 5 uses a wireless communication protocol to transmit control signals and supports real-time data interaction with the traffic signal actuator. The signal output module 5 integrates a data feedback unit, which collects traffic flow, vehicle speed and queue length data after regulation in real time, providing data support for the three-level feedback module.

[0040] Each module is connected to the server 6 via communication links (5G wireless or wired communication) to form a closed-loop control system. Specifically, the multi-sensor data acquisition module 1 establishes bidirectional data transmission links with the dynamic weight fusion module 2 and the signal output module 5, respectively. The dynamic weight fusion module 2 establishes a unidirectional data transmission link with the phase collaborative optimization genetic algorithm module 3, which in turn establishes unidirectional data transmission links with the proportional-integral-derivative control module 4 and the signal output module 5, respectively. The proportional-integral-derivative control module 4 also establishes a unidirectional data transmission link with the signal output module 5. The server 6 establishes bidirectional data transmission links with each module. The three-level feedback module 7 establishes bidirectional data transmission links with the multi-sensor data acquisition module 1, the server 6, and the phase collaborative optimization genetic algorithm module 3, respectively.

[0041] The generated traffic light timing scheme is then adjusted in real time by combining the parameter pool optimization results from the three-level feedback module 7 with the strategy mapping library matching results. Simultaneously, the signal output module sends control commands to the traffic lights to execute dynamic timing. The data feedback unit collects real-time data on traffic flow, vehicle speed, and queue length after the adjustment. This data is then used by the three-level feedback module 7 to adjust the sampling frequency of each sensor, update the parameter pool, and optimize the strategy mapping library, providing a basis for the next round of parameter optimization and timing adjustment, thus forming a continuous closed-loop optimization.

[0042] As one possible implementation, the above-mentioned traffic light timing scheme modification also includes adjustments to time period characteristics such as holidays and morning and evening peak hours. During peak hours, the traffic light cycle length is extended to 120-150 seconds, and during holidays, the green light duration ratio of each direction is dynamically adjusted according to traffic flow fluctuations. The scenario division dimensions of the scenario-timing strategy mapping library include traffic status level, time period type, weather conditions, and traffic flow change.

[0043] For example, the three-level feedback module is specifically configured as follows: Data layer feedback: Extracting deviation characteristics between the regulated data and the fused data to dynamically adjust the sensor sampling frequency. The core of data layer feedback is to dynamically adapt the sensor sampling frequency by comparing the deviation characteristics between the regulated data and the fused data. The regulated data is the real-time traffic operation data (including measured indicators such as traffic flow, vehicle speed, and queue length) collected by the signal output module after executing timing commands; the fused data is the traffic state estimate output by the dynamic weight fusion module (including accurate reference data such as traffic flow density and predicted queue length). The system extracts deviation characteristics between the two types of data, such as the deviation rate between the predicted and actual queue length values, and the difference between the estimated and measured traffic speed values. If the deviation is large, it indicates that the current sampling data is insufficient to support precise regulation or that the system is in a complex traffic state such as congestion, and the sensor sampling frequency will be increased to ensure data real-time performance. If the deviation is small or the system is in a low-peak state, the sampling frequency will be reduced to reduce system energy consumption, ultimately achieving a dynamic balance between data real-time performance and acquisition efficiency.

[0044] Parameter layer feedback: Regularly analyzes the optimization effects of proportional-integral-derivative parameters to build dedicated parameter pools for different traffic scenarios. Parameter layer feedback focuses on the continuous optimization and scenario-based accumulation of proportional-integral-derivative parameters, building dedicated parameter pools through regular statistical analysis of parameter optimization effects. The system reviews the control effect of proportional-integral-derivative parameters in the current traffic scenario at fixed intervals (e.g., hourly), with core judgment criteria including indicators such as average vehicle waiting time, intersection traffic efficiency, and queue length changes. For different traffic scenarios such as congestion, off-peak hours, and rainfall, it selects the parameter ranges that consistently perform optimally (e.g., the optimal range of proportional coefficient values ​​in congestion scenarios) and categorizes and stores them in scenario-specific parameter pools. Subsequent optimizations do not require a global search from scratch; instead, local iterations are performed directly within the corresponding scenario's parameter pool, significantly shortening the parameter optimization cycle and improving control response speed.

[0045] Strategy layer feedback: Based on historical data stored on the server, when the current scenario and a historical scenario meet a preset matching degree, the optimal timing scheme from the past is used as the initial value for local optimization. Strategy layer feedback relies on historical data stored on the server to achieve rapid adaptation and optimization of timing strategies through scenario matching. The server records massive amounts of historical traffic scenarios, covering multi-dimensional features such as traffic state level, time period type, weather conditions, and sudden changes in traffic flow, along with corresponding optimal timing schemes, forming a scenario-timing strategy mapping library. When the system detects the current traffic scenario, such as morning rush hour + congestion + no rain, it compares its features with historical scenarios in the mapping library. If the matching degree reaches a preset threshold (e.g., ≥85%), the optimal timing scheme of that historical scenario is directly used as the initial value, and then local adjustments are made based on subtle differences in the current traffic state to generate the final timing scheme. This approach avoids repetitive full-process optimization, ensuring the adaptability of the timing strategy while significantly improving the control response efficiency in complex scenarios.

[0046] Reference Figure 2 This application also provides a specific embodiment, the implementation process of which is as follows: A multi-sensor data acquisition module is deployed at an urban intersection. Geomagnetic sensors are installed at 5m, 10m, and 15m before the stop line, one per lane; four sets of video sensors are deployed on the traffic light poles in the four directions of the intersection; one set of meteorological sensors is installed in the central median of the intersection; and four sets of traffic flow detectors are deployed corresponding to each approach lane. A dynamic weight fusion module, a phase collaborative optimization genetic algorithm module, and a proportional-integral-derivative control module are deployed inside the intersection traffic control box. The server is deployed in the urban traffic management center, and data interaction between the modules is achieved through 5G communication.

[0047] The parameter settings for the dynamic weight fusion module are as follows: the input dimension of the MLP neural network is 4, corresponding to traffic flow, vehicle speed, queue length, and weather data, respectively; the number of hidden layer nodes is set to 8, 12, and 8, respectively; the input dimension of the autoencoder is 4, and the encoding dimension is 2; the time step of the RNN neural network is set to 5. The dynamic weight allocation thresholds are as follows: congested state (occupancy > 60%), slow-moving state (30% ≤ occupancy ≤ 60%), and off-peak state (occupancy < 30%).

[0048] The parameter settings for the phase-cooperative optimization genetic algorithm module are as follows: the initial population size is set to 40, the maximum number of iterations is set to 80, the crossover probability is set to 0.7, the initial mutation probability is set to 0.03, the fitness function is α=0.6 and β=0.4 when there is no mutation, and α=0.3 and β=0.7 when there is a mutation; the threshold for the traffic flow mutation coefficient γ is 1.5; the acceptable fitness threshold corresponds to an average vehicle waiting time ≤30 seconds.

[0049] The parameter ranges of the proportional-integral-derivative control module are as follows: Kp∈[0.1,5.0], Ki∈[0.01,1.0], Kd∈[0.05,2.0]; the scenario-specific parameter pool is initially divided into three scenarios: congestion, off-peak, and rainfall.

[0050] The timing parameter thresholds are set as follows: road occupancy congestion threshold 60%, off-peak threshold 30%, yellow light base duration 3 seconds, extended to 5 seconds in rainy weather; traffic light cycle duration 120-150 seconds during peak hours, 80-100 seconds during off-peak hours, and 100-120 seconds during slow traffic periods.

[0051] The parameters for the three-level feedback module are set as follows: data layer feedback sampling frequency adjustment step size is 0.1 seconds / time; parameter layer feedback update cycle is 1 hour; strategy layer feedback scenario matching threshold is 85%.

[0052] Sensors collect multi-dimensional raw traffic data at an initial frequency of 0.5 seconds per acquisition and transmit it to the dynamic weight fusion module. The dynamic weight fusion module first detects the data timestamps. When a sensor's data is missing (timestamp interval greater than 30 seconds), the missing value is filled using linear interpolation. After performing min-max normalization on all data, it determines that the eastbound approach road occupancy rate is 65% (congested state), there is no rainfall (non-Gaussian noise is not significant), and dynamically allocates weights: MLP 60%, CNN 20%, RNN 15%, and the Kalman filter native framework 5%. Through MLP nonlinearity compensation, autoencoder denoising, CNN noise modeling, and RNN model correction, it outputs traffic flow density, optimal traffic speed, predicted queue length, and traffic flow mutation coefficient γ=1.2 (no mutation).

[0053] Proportional-Integral-Derivative Parameter Phase Co-optimization: The phase co-optimization genetic algorithm module, based on the fused traffic state data, uses phase co-encoding to jointly encode Kp, Ki, Kd with the lower limit of east-west green light duration (≥40 seconds) and phase switching interval (≥3 seconds) to construct an initial population. Based on γ=1.2, a mutation-free fitness function is used. After 35 generations of evolution, the fitness value reaches an acceptable threshold (average vehicle waiting time of 28 seconds), and 10 rapid fine-tuning iterations are initiated to output the optimal PID parameters (Kp=3.2, Ki=0.42, Kd=0.95). During the optimization process, a congestion scenario-specific parameter pool (Kp∈[2.0-3.5], Ki∈[0.3-0.5], Kd∈[0.8-1.0]) is called to shorten the search time.

[0054] Dynamic Timing and Execution: The proportional-integral-derivative (PI-DE) control module generates traffic light control parameters based on optimal parameters and real-time traffic conditions. During morning rush hour congestion, the green light duration for eastbound traffic is adjusted from 40 seconds to 58 seconds, for westbound traffic to 52 seconds, and for northbound and southbound traffic to remain at 35 seconds each. The timing scheme is matched through strategy layer feedback; the current scenario matches historical morning rush hour congestion scenarios with a 92% match rate. Local optimization is performed using the historically optimal timing scheme as the initial value to ultimately determine the timing parameters. Timing commands are transmitted to the signal output module via 5G communication. The signal output module converts the timing commands into traffic light control signals and sends them to the traffic signal actuator. The traffic signal actuator performs dynamic adjustments and collects actual traffic data every minute, feeding it back to the three-level feedback module.

[0055] The three-level feedback execution is as follows: The data layer extracts the deviation features between the actual queue length and the predicted value (deviation rate 8%), and maintains a sampling frequency of 0.5 seconds / time; the parameter layer counts the optimization effect within 1 hour and updates Kp∈[2.8-3.4], Ki∈[0.38-0.45], and Kd∈[0.9-1.0] to the congestion scenario-specific parameter pool; the strategy layer optimizes the timing strategy mapping relationship of the morning peak congestion scenario based on the current control effect (eastbound queue length reduced by 48%).

[0056] Through the application of the system provided in this application embodiment, the average queue length of vehicles at the target intersection during peak hours is reduced by 48%, the road space occupancy rate is reduced by 38%, the traffic efficiency is increased by 45%, the vehicle idling energy consumption is reduced by 35%, and the response speed in sudden traffic flow scenarios is increased by 2 times, which significantly improves traffic congestion and enhances the system's operating energy efficiency and robustness.

[0057] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0058] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A traffic signal control system based on dynamic weight fusion and phase collaborative optimization, characterized in that, include: The multi-sensor data acquisition module is used to deploy multiple sensors to collect multi-dimensional raw traffic data of each section of the target road in real time and transmit it to the dynamic weight fusion module; Multidimensional raw traffic data should include at least traffic flow, vehicle speed, queue length, and weather conditions; The dynamic weight fusion module receives the multi-dimensional raw traffic data, processes it using a dynamic weight adaptive fusion algorithm to obtain traffic state estimates, and transmits them to the phase collaborative optimization genetic algorithm module. The dynamic weight adaptive fusion algorithm includes data preprocessing, dynamic weight allocation, nonlinear compensation, noise modeling, and model correction. The dynamic weight fusion module incorporates multiple fusion units, each corresponding to nonlinear compensation, noise modeling, model correction, and basic data fusion functions. The dynamic weight allocation dynamically adjusts the weight of each fusion unit based on the real-time traffic state level. The phase-cooperative optimization genetic algorithm module, based on the traffic state estimate, optimizes the proportional coefficient, integral coefficient, and derivative coefficient of the proportional-integral-derivative (PID) controller included in the proportional-integral-derivative (PID) control module using an improved genetic algorithm, outputs the optimal PID parameters to the PID control module, and simultaneously generates dynamic traffic light timing commands and transmits them to the signal output module. The improved genetic algorithm includes a phase-cooperative coding mechanism, a dynamic fitness function, a fitness enhancement elimination selection mechanism, gene infection crossover operation, an adaptive mutation operator, and a real-time iteration termination constraint. The proportional-integral-derivative (PI-DI) control module is used to receive the optimal PI-DI parameters, combine them with the traffic state estimate to construct an adaptive PI-DI control model, and generate real-time signal control parameters. The signal output module is used to receive the timing command, convert the timing command into a traffic light control signal and send it to the traffic light actuator to realize dynamic control of the traffic light. At the same time, it collects traffic data after control in real time and feeds it back to the multi-sensor data acquisition module.

2. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 1, characterized in that: It also includes a server and a three-level feedback module; The server is used to store multi-dimensional raw traffic data, preprocessed data, traffic state estimates, proportional-integral-derivative controller parameters, genetic algorithm optimization process data, historical timing schemes, and a scenario-timing strategy mapping library. The three-level feedback module constructs a three-level feedback architecture consisting of a data layer, a parameter layer, and a strategy layer, enabling dynamic adjustment of the sensor sampling frequency, optimization of the proportional-integral-derivative parameter pool, and rapid matching of timing strategies.

3. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 1, characterized in that: The multi-sensor data acquisition module deploys sensors including geomagnetic sensors, video sensors, meteorological sensors, and traffic flow detectors. The sensors are deployed to cover key sections of the target road. The sampling frequency of each sensor is dynamically adjusted through data layer feedback to balance data real-time performance and acquisition efficiency.

4. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 1, characterized in that, The dynamic weight fusion module includes a multilayer perceptron neural network for nonlinear compensation, a convolutional neural network for noise modeling, a recurrent neural network for model correction, and a native Kalman filter framework for basic data fusion.

5. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 4, characterized in that, The dynamic weight fusion module is specifically configured as follows: data preprocessing—timestamp synchronization detection, missing data filling and scale unification processing are performed on multi-dimensional raw traffic data to obtain standardized data; traffic status levels are divided based on the road occupancy rate in the standardized data, and meteorological data is combined to determine whether there are significant non-Gaussian noise scenarios; dynamic weight allocation—the weight ratio of the multilayer perceptron neural network, convolutional neural network, recurrent neural network and Kalman filter native framework is dynamically adjusted according to the traffic status level and noise scenario type. Nonlinear compensation—Nonlinear compensation is achieved by learning the nonlinear transfer function of multidimensional raw traffic data after data preprocessing through a multilayer perceptron neural network; Noise modeling—Non-Gaussian noise distribution characteristics are learned through autoencoder denoising preprocessing and convolutional neural network, and the noise covariance matrix and observation noise covariance matrix of Kalman filter are dynamically adjusted. Model calibration—Learns the drift patterns of traffic system model parameters in real time through recurrent neural networks, dynamically updates the state transition matrix and observation matrix of Kalman filter, and outputs traffic state estimates including traffic flow density, optimal speed, queue length predictions, and traffic flow mutation coefficients.

6. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 1, characterized in that, The phase-coordinated optimization genetic algorithm module is specifically configured as follows: The proportional-integral-derivative (PID) controller's proportional coefficient, integral coefficient, and derivative coefficient, along with the green light duration constraints for each phase, are jointly encoded into chromosome individuals to construct an initial population. Phase priority factors are embedded within these chromosome individuals. A traffic flow mutation coefficient is introduced, and the weight coefficients of the fitness function are dynamically adjusted with the optimization objective of minimizing average vehicle waiting time and maximizing traffic efficiency. A fitness-enhancing elimination rule is used for selection, screening high-quality individuals for the next generation. A gene infection crossover method is used for crossover, transmitting gene fragments from the optimal individual to offspring individuals. An adaptive mutation operator is used to dynamically adjust the mutation probability based on the population fitness distribution. The optimal individual from each generation is retained without participating in crossover or mutation operations. When the preset convergence condition is met, the maximum number of iterations is reached, or the fitness value reaches an acceptable threshold, iteration stops and the optimal PID parameters are output to the PID control module. Based on the optimal PID parameters and traffic state estimates, the passage time for each phase of the traffic light is dynamically calculated, timing instructions are generated, and transmitted to the signal output module.

7. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 1, characterized in that, The proportional-integral-derivative (PI-DE) control module is specifically configured to: receive the optimal PI-DE parameters, combine them with the traffic flow density, optimal traffic speed, queue length prediction, and traffic flow mutation coefficient output by the dynamic weight fusion module, construct an adaptive PI-DE control model, and generate real-time signal control parameters adapted to the current traffic scenario.

8. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 1, characterized in that, The signal output module uses a wireless communication protocol to transmit control signals and supports real-time data interaction with the traffic signal light actuator. The signal output module integrates a data feedback unit, which collects traffic flow, vehicle speed and queue length data after regulation in real time, providing data support for the three-level feedback module.

9. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 2, characterized in that, The three-level feedback module is specifically configured as follows: data layer feedback—extracting the deviation characteristics between the data after regulation and the fused data, and dynamically adjusting the sensor sampling frequency; parameter layer feedback—periodically statistically analyzing the optimization effect of proportional-integral-derivative parameters, and constructing a dedicated parameter pool for different traffic scenarios; Policy layer feedback—Based on historical data stored on the server, when it is detected that the current scene and the historical scene meet the preset matching degree, the historical best timing scheme is called as the initial value for local optimization.

10. The traffic signal control system based on dynamic weight fusion and phase collaborative optimization according to claim 2, characterized in that, Each module is connected to the server via communication links to form a closed-loop control system. Specifically, the multi-sensor data acquisition module establishes bidirectional data transmission links with the dynamic weight fusion module and the signal output module; the dynamic weight fusion module establishes a unidirectional data transmission link with the phase collaborative optimization genetic algorithm module; the phase collaborative optimization genetic algorithm module establishes a unidirectional data transmission link with the proportional-integral-derivative control module and the signal output module; the proportional-integral-derivative control module establishes a unidirectional data transmission link with the signal output module; the server establishes bidirectional data transmission links with each module; and the three-level feedback module establishes bidirectional data transmission links with the multi-sensor data acquisition module, the server, and the phase collaborative optimization genetic algorithm module.

Images