Tunnel traffic flow prediction and control method and system

Abstract

The invention belongs to the technical field of traffic control, and discloses a tunnel traffic flow prediction and control method and system. The method comprises the following steps: arranging a sensor network in a tunnel based on a tunnel traffic control scheme of edge calculation, and presetting an initial weight and a redundancy activation rule; collecting and processing data by edge nodes, generating traffic and sensor vectors, and inputting the traffic and sensor vectors into a flow prediction model; through combination of an output degradation mechanism and PPO reinforcement learning, real-time weight is dynamically adjusted, and optimization prediction is fed back; a tunnel three-dimensional model is constructed, an optimal control strategy is generated based on an NSGA-III algorithm, and cooperative control of an induction system, tunnel equipment and signal lamps is achieved; according to the method, the reliability of data acquisition, the accuracy of flow prediction and the robustness of a control strategy in an extreme scene are improved, and a key breakthrough from post-event response to real-time intelligence of tunnel traffic control is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic control, and more particularly to a method and system for predicting and controlling tunnel traffic flow. Background Art

[0002] With the acceleration of urbanization and the continuous growth of traffic volume, tunnels, as key nodes in the transportation network, face challenges in their safe and efficient operation.

[0003] Chinese patent application publication number CN111179601A discloses a method for controlling tunnel traffic operations: Step S1: Collecting traffic operation data within the tunnel and evaluating the warning level of each traffic operation scenario triggered within the tunnel based on the traffic operation data. The traffic operation data includes traffic flow data, vehicle behavior data, event data, and external data; Step S2: Determining the control level and the corresponding traffic operation scenario based on the warning level of each traffic operation scenario; Step S3: Determining the corresponding control plan based on the control level and traffic operation scenario, and performing traffic control according to the control plan. The control plan consists of multiple control measures. This tunnel traffic operation control method can evaluate the traffic operation status under various traffic operation scenarios triggered within the tunnel, allowing managers to promptly grasp the traffic operation warning level within the tunnel. Managers can then formulate targeted control plans based on the warning level to promptly and effectively address operational risks, reduce the probability of accidents, and improve tunnel operation efficiency.

[0004] Although the above methods can meet most scenarios, research and practical application of the above methods and existing technologies have revealed that the above methods and existing technologies have at least the following defects:

[0005] The performance degradation of sensors in extreme environments is not taken into account, and the degradation of sensor output cannot be compensated in real time, resulting in increased traffic flow prediction errors.

[0006] In view of this, the present invention proposes a tunnel traffic flow prediction and control method and system to solve the above problems. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solution: a tunnel traffic flow prediction and control method, comprising the following steps:

[0008] Deploy a sensor network in the tunnel, presetting the initial sensor weights and redundant activation rules for extreme scenarios;

[0009] The sensor network collects and transmits sensor data and traffic data to edge nodes, analyzes and obtains traffic vectors; uses sensor data and initial weights as inputs to the fusion model to obtain sensor vectors;

[0010] The initial weights, sensor vectors, and traffic vectors are used as inputs to the traffic prediction model to obtain traffic prediction values ​​for the future time window.

[0011] Combining sensor vectors, traffic vectors, and traffic prediction values, the sensor output degradation mechanism triggered by edge nodes is used to simulate extreme scenarios, dynamically compensate the sensor output, and adjust the real-time weight based on the compensation results using the PPO reinforcement learning algorithm.

[0012] Based on the real-time weights, feedback is given to the fusion model and traffic prediction model to update the traffic prediction value and build a three-dimensional tunnel model. Then, combined with the real-time weights, the optimal control strategy for controlling the guidance system, tunnel control system, and upstream and downstream traffic lights is generated based on the NSGA-III algorithm.

[0013] Furthermore, the method of simulating extreme scenarios using the sensor output degradation mechanism triggered by the edge node includes:

[0014] When the sensor network detects that the sensor data or traffic data exceeds the preset threshold, the sensor output degradation mechanism is triggered to simulate the scenario and modify the parameters of the sensor output.

[0015] Furthermore, the method of performing scenario simulation and parameter correction on sensor output by the output degradation mechanism includes:

[0016] For camera sensors, a light intensity degradation model is built based on the attenuation coefficient, detection distance, and noise to correct the actual light intensity.

[0017] For radar sensors: For laser sensors, the effective detection distance is calculated based on the scattering coefficient, absorption coefficient, transmission power, receiving sensitivity, and target reflectivity; for millimeter wave sensors, the corrected signal-to-noise ratio is calculated based on the original signal-to-noise ratio, combined with rainfall and detection distance.

[0018] For radar sensors, the signal power loss is also calculated based on the reflection coefficient, dynamic delay, and multipath effect.

[0019] For traffic detection sensors, a corrected signal voltage is calculated based on the actual signal voltage, external magnetic field strength, interference frequency, and a preset calibration coefficient, and an updated signal-to-noise ratio is calculated based on the corrected signal voltage.

[0020] According to the above output degradation mechanism, the scenario simulation and parameter correction of sensor output are performed on the data obtained to perform supervised learning training for various sensors.

[0021] Furthermore, the method of obtaining the traffic vector includes:

[0022] The traffic volume is obtained by counting the number of vehicle pulses passing through the geomagnetic coil;

[0023] Traffic detection sensors are used to collect the time difference between vehicles passing through the tunnel to calculate the average speed of vehicles in the corresponding lanes.

[0024] Detect and obtain the vehicle speed distribution of each lane, divide the speed intervals according to the preset speed intervals, and count the proportion of vehicles in each lane in each interval to obtain the vehicle speed distribution;

[0025] Calculate the average vehicle speed and calculate the traffic density based on the traffic flow;

[0026] The actual average time it takes for vehicles to pass through the tunnel is calculated, and the ratio of the actual average time to the free flow time is calculated to obtain the travel time ratio.

[0027] The traffic density, vehicle speed distribution and travel time ratio are spliced ​​together to obtain the traffic vector.

[0028] Furthermore, if the average vehicle speed corresponding to a lane falls below R% of the preset free-flow speed, it is marked as potentially congested. The NSGA-III algorithm is directly triggered to generate an optimal control strategy based on the historical sensor weights, and the weights and optimal control strategy are updated in real time based on updated sensor data; R is a preset threshold.

[0029] Furthermore, the method of obtaining the sensor vector includes:

[0030] Obtain sampling data from each sensor. For camera sensors, obtain the mean square error (MSE) between the sampling data and the processed data. Calculate the peak signal-to-noise ratio (PSNR) based on the MSE and the maximum possible value of the pixel value.

[0031] The sampling data of the camera sensor is used as the input of the camera detection model to obtain the detection confidence of the camera sensor;

[0032] For radar sensors, the number of point clouds per unit area is calculated to obtain the point cloud density. The point cloud density is compared with a preset density threshold. Areas below the preset density threshold are marked as low quality, and the marking results of all areas are counted.

[0033] Through Kalman filtering, a unique ID is assigned to each physical entity (i.e., target) detected by the radar sensor. The occurrence and loss of each target ID in the continuous frame data of the radar sensor is recorded, and the corresponding loss rate is recorded.

[0034] For environmental sensors, calculate the mean of historical data and the deviation rate between real-time data and the mean of historical data;

[0035] The peak signal-to-noise ratio, detection confidence, labeling results, loss rate and deviation rate, as well as the corresponding initial weights are used as inputs of the fusion model to obtain the sensor vector.

[0036] Furthermore, the method for adjusting and obtaining the real-time weight based on the compensation result includes:

[0037] Define the state space: The state space consists of a state vector, which includes environmental parameters, sensor vectors, traffic vectors, traffic flow prediction values, and the deviation between the traffic flow prediction value and the actual traffic flow;

[0038] Define the action space: The action space consists of action vectors, which include weight adjustments of sensor nodes. The weight adjustment of each sensor satisfies the corresponding preset weight adjustment range, and the sum of the weights of all sensors is 1.

[0039] Design a reward function: This function is calculated by combining the deviation between the traffic prediction value and the actual traffic flow, the prediction stability within the preset time period, the sensor energy consumption, and the volatility of the sensor fusion weight.

[0040] Simulate extreme tunnel scenarios and compensate for sensor output changes based on sensor output degradation mechanisms;

[0041] Sensor data, environmental data, and historical data are used as inputs to the Actor network, and a multi-layer Transformer+CNN is used as the middle layer to obtain the dynamic weight vector of each sensor.

[0042] The sensor data, environmental data, historical data, and the error between the fused traffic prediction value and the true value are used as the input of the Critic network, and the long-term benefits of the current weight distribution strategy are output through the state value function.

[0043] Select the weight allocation strategy with the highest long-term return as the final adjustment strategy to obtain the corresponding dynamic weight.

[0044] Furthermore, the method for generating the optimal control strategy includes:

[0045] Preset optimization objectives and constraints;

[0046] Preset chromosome encoding scheme: set the gene structure of the chromosome including weight genes and strategy genes;

[0047] Randomly generate real-time weights and control strategy parameters for N groups of sensors. The control strategy parameters include control instructions for the induction system, tunnel control system, and upstream and downstream signal light control instructions. The real-time weights and control strategy parameters of the N groups of sensors are encoded into the genetic structure of the N1 group of chromosomes. Individuals that violate the constraints are eliminated to ensure that the initial population contains all feasible solutions.

[0048] The population is divided into M frontier layers according to the Pareto dominance relationship and arranged in the order of increasing dominance number. The edge node obtains the first L frontier layers and uploads them;

[0049] Based on the target space formed by the first L frontier layers, a uniformly distributed set of reference points is generated in each dimension of the preset optimization target, where the number of reference point sets is E times the number of chromosomes in the first L frontier layers; the position of the reference point set is periodically updated according to the change of the objective function range;

[0050] The parent chromosome parameters are crossed to generate offspring. For weight genes, arithmetic crossover is used to ensure that the sum of weights after crossover is 1; for strategy genes, discrete crossover is used to randomly exchange parent parameter values.

[0051] Random perturbations are applied to offspring through mutation. For weight genes, random perturbations are applied to each weight and then renormalized. For strategy genes, Gaussian noise is added within the parameter range.

[0052] Calculate the Euclidean distance between each offspring individual and the reference point, associate the individual to the reference point with the smallest Euclidean distance, retain the first Q individuals associated with the reference point to enter the offspring, and directly retain the first P individuals on the optimal frontier of the parent generation to enter the offspring; obtain the offspring population;

[0053] Repeat the iteration of the offspring individuals until the preset termination condition is reached.

[0054] Furthermore, the method for obtaining the first L frontier layers includes:

[0055] Step A1, Domination Rule: For two individuals A and B, if A's target values ​​on all preset optimization targets are not lower than B's, and it is better than B on at least one target, then A is said to dominate B, and B's dominated number is increased by 1. Count the dominated numbers of all individuals according to the above domination rule;

[0056] Step A2: extract all individuals in the population that are not dominated by any other individuals, that is, individuals with a domination count of 0, to form the first frontier layer;

[0057] Step A3: Repeat step A1 to extract the dominated numbers of all individuals in the remaining individuals after the first frontier layer, and extract individuals with a dominated number of 0 to form a second frontier layer;

[0058] Step A4: Repeat step A3 to extract the currently undominated individuals one by one to form the next frontier layer until all individuals are assigned to the corresponding frontier layer or the preset maximum number of layers L is reached.

[0059] A tunnel traffic flow prediction and control system, implementing the tunnel traffic flow prediction and control method, comprises:

[0060] Network construction module: Deploys a sensor network in the tunnel, presetting the initial sensor weights and redundant activation rules for extreme scenarios;

[0061] Data acquisition module: The sensor network collects and transmits sensor data and traffic data to edge nodes, analyzes and obtains traffic vectors; uses sensor data and initial weights as input to the fusion model to obtain sensor vectors;

[0062] Traffic prediction module: takes the initial weight, sensor vector and traffic vector as inputs of the traffic prediction model to obtain the traffic prediction value of the future time window;

[0063] Compensation Analysis Module: This module combines sensor vectors, traffic vectors, and traffic flow predictions, uses the sensor output degradation mechanism triggered by edge nodes to simulate extreme scenarios, dynamically compensates sensor outputs, and uses the PPO reinforcement learning algorithm to adjust the compensation results to obtain real-time weights.

[0064] Intelligent control module: Based on real-time weights, it provides feedback to the fusion model and traffic prediction model, updates the traffic prediction value, builds a three-dimensional tunnel model, and then generates the optimal control strategy for the guidance system, tunnel control system, and upstream and downstream traffic lights based on the NSGA-III algorithm in combination with the real-time weights.

[0065] The technical effects and advantages of the tunnel traffic flow prediction and control method and system of the present invention are as follows:

[0066] The present invention forms a core technical advantage that is missing from traditional solutions by constructing a sensor output degradation mechanism at the edge node and combining PPO reinforcement learning to dynamically compensate for sensor performance degradation in extreme scenarios: in extreme scenarios such as dense fog, heavy rain, and accidents, sensor anomalies can be identified in real time, and the accuracy of multi-source data fusion can be improved by dynamically adjusting the fusion weights, thereby reducing traffic prediction errors; the edge computing architecture supports localized real-time processing, which can shorten the response time in extreme scenarios and ensure that reliable control strategies can still be quickly generated when sensor performance degrades; the present invention improves the reliability of data collection in extreme scenarios, the accuracy of traffic prediction, and the robustness of control strategies through the "model simulation-weight compensation-edge decision-making" closed loop, achieving a key breakthrough in tunnel traffic control from ex post response to real-time intelligent response. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Schematic diagram of the tunnel traffic flow prediction and control method of the present invention;

[0068] Figure 2 This is the architecture diagram of the tunnel traffic flow prediction and control system of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] Example 1

[0071] See also Figure 1 As shown, this embodiment provides a tunnel traffic flow prediction and control method, comprising the following steps:

[0072] A sensor network is deployed in the tunnel, and the initial weights of the sensors and redundant activation rules for extreme scenarios are preset. For example, the lidar + millimeter-wave radar combination is activated in heavy rain (the activation weights of the lidar + millimeter-wave radar are increased accordingly, such as activation weight = 1.5*initial weight), the thermal imaging camera + radar combination is activated in dense fog, and redundant sensor clusters are activated in accidents. A sensor network is deployed in the tunnel and the initial weights of the sensors and redundant activation rules for extreme scenarios are preset. Multi-dimensional data such as traffic flow and environment can be collected in real time through multimodal sensors (such as lidar, cameras, geomagnetic coils, etc.). The initial weights ensure the efficient fusion of multi-source data in conventional scenarios, and the redundant activation rules dynamically activate the backup sensor cluster in extreme scenarios such as heavy rain, dense fog, and accidents to ensure the integrity and reliability of data collection. The two provide stable and high-precision input data for subsequent traffic prediction models, supporting edge nodes to dynamically adjust fusion weights based on PPO reinforcement learning to improve traffic prediction accuracy; at the same time, they provide real-time and reliable sensor status information for the generation of intelligent induction control strategies, combined with the NSGA-III algorithm to optimize control parameters such as traffic light timing and lane allocation, achieving rapid response and global optimal control in extreme scenarios, and ultimately enhancing the prediction accuracy, emergency robustness and traffic efficiency of the tunnel traffic system.

[0073] The sensor network collects and transmits sensor data and traffic data to edge nodes, where it analyzes and obtains traffic vectors. The sensor data and initial weights are used as inputs to the fusion model to obtain sensor vectors. The sensor network collects and transmits sensor data and traffic data to edge nodes, where they are analyzed and processed to generate traffic vectors reflecting real-time traffic status, providing intuitive traffic flow feature input for the traffic prediction model. Simultaneously, the raw sensor data and preset initial weights are input into the fusion model, where sensor vectors are generated through weighted fusion and noise filtering. This effectively integrates multi-source sensor information and suppresses the performance degradation of a single sensor in complex environments. Together, these two provide multi-dimensional, high-precision input data for the traffic prediction module, enabling edge nodes to accurately predict future time window traffic flow based on models such as LSTM / Transformer. Furthermore, these two components provide real-time, reliable traffic status information for the generation of intelligent induction control strategies, enabling the NSGA-III algorithm to dynamically optimize control parameters such as signal timing, lane allocation, and induction information dissemination. This completes a closed-loop process from data collection to prediction and decision-making, significantly improving the real-time and intelligent level of tunnel traffic control.

[0074] Methods for obtaining traffic vectors include:

[0075] The traffic volume is obtained by counting the number of vehicle pulses passing through the geomagnetic coil;

[0076] Traffic detection sensors are used to collect the time difference between vehicles passing through the tunnel to calculate the average speed of vehicles in the corresponding lanes.

[0077] The detection obtains the vehicle speed distribution of each lane, divides the speed interval into intervals according to the preset speed interval, and calculates the proportion of vehicles in each interval in each lane to obtain the vehicle speed distribution. If the average vehicle speed of the corresponding lane is lower than the preset free flow speed (the highest average speed at which vehicles can travel stably under ideal traffic conditions, such as no traffic congestion, no traffic accidents, and good weather) by R% (R is a preset threshold, such as 50%), based on traffic flow theories such as the Greenshields model and car-following theory, as well as empirical settings derived from actual observation data, the lane is marked as potentially congested.

[0078] Calculate the average vehicle speed and calculate the traffic density based on the traffic flow;

[0079] The actual average time it takes for vehicles to pass through the tunnel is calculated, and the ratio of the actual average time to the free flow time is calculated to obtain the travel time ratio.

[0080] The traffic density, vehicle speed distribution and travel time ratio are spliced ​​together to obtain the traffic vector.

[0081] The training methods for the fusion model include:

[0082] S groups of fusion training data are collected in advance, and the fusion training data includes sensor data, initial weights and sensor vectors.

[0083] The sensor data and initial weights are used as the input of the fusion model, and the sensor vector is used as the output of the fusion model. The goal is to minimize the error between the output sensor vector and the actual sensor vector. The network parameters of the fusion model are optimized through a nature-inspired optimization algorithm to obtain the network parameters that minimize the error between the sensor vector output by the fusion model and the actual sensor vector. The fusion model constructed with the corresponding network parameters is used as the trained fusion model.

[0084] Methods for obtaining sensor vectors include:

[0085] Obtain sampling data from each sensor. For camera sensors, obtain the mean square error (MSE) between the sampling data and the processed data. Calculate the peak signal-to-noise ratio (PSNR) based on the MSE and the maximum possible value of the pixel value, such as 255.

[0086] The sampling data of the camera sensor is used as the input of the camera detection model to obtain the detection confidence of the camera sensor;

[0087] The training methods for the camera detection model include:

[0088] K groups of detection training data are collected in advance. The detection training data includes camera sensor sampling data and detection confidence.

[0089] The sampling data of the camera sensor is used as the input of the camera detection model, and the detection confidence is used as the output of the camera detection model. The goal is to minimize the error between the output detection confidence and the actual detection confidence. The network parameters of the camera detection model are optimized through the nature-inspired optimization algorithm to obtain the network parameters that minimize the error between the detection confidence output by the camera detection model and the actual detection confidence. The camera detection model constructed with the corresponding network parameters is used as the trained camera detection model.

[0090] For radar sensors, the number of point clouds per unit area is calculated to obtain the point cloud density. The point cloud density is compared with a preset density threshold. Areas below the preset density threshold are marked as low quality, and the marking results of all areas are counted.

[0091] Through Kalman filtering, a unique ID is assigned to each physical entity detected by each radar sensor, that is, the detected target. The occurrence and loss of each target ID in the continuous frame data of the radar sensor are recorded, and the corresponding loss rate is recorded;

[0092] For environmental sensors, calculate the mean of historical data and the deviation rate between real-time data and the mean of historical data;

[0093] The peak signal-to-noise ratio, detection confidence, labeling results, loss rate and deviation rate, as well as the corresponding initial weights are used as inputs of the fusion model to obtain the sensor vector.

[0094] The initial weights, sensor vectors, and traffic vectors are used as inputs to the traffic prediction model to obtain traffic prediction values ​​for the future time window.

[0095] The training methods for the traffic prediction model include:

[0096] Y groups of detection training data are collected in advance. The detection training data include initial weights, sensor vectors and traffic vectors, as well as traffic prediction values ​​in future time windows.

[0097] The initial weights, sensor vectors and traffic vectors are used as the input of the traffic prediction model, and the traffic prediction value of the future time window is used as the output of the traffic prediction model. The goal is to minimize the error between the traffic prediction value output in the future time window and the actual traffic value. The network parameters of the traffic prediction model are optimized through a nature-inspired optimization algorithm to obtain the network parameters that minimize the error between the detection confidence output by the traffic prediction model and the actual detection confidence. The traffic prediction model constructed with the corresponding network parameters is used as the trained traffic prediction model.

[0098] By combining sensor vectors, traffic vectors, and traffic flow predictions, the system simulates extreme scenarios using sensor output degradation mechanisms triggered by edge nodes, dynamically compensates for sensor output, and adjusts real-time weights based on the compensation results using the PPO reinforcement learning algorithm. These steps provide real-time compensation for sensor output degradation in complex environments. This process significantly improves the accuracy and robustness of multi-source data fusion, providing more reliable input for traffic flow prediction models. Furthermore, the real-time weights, as quantitative representations of sensor states, directly contribute to the generation of intelligent induction control strategies, helping the NSGA-III algorithm dynamically balance sensor reliability and control strategy efficiency in multi-objective optimization. This allows for precise control of induction systems, tunnel equipment, and traffic lights. This shortens emergency response times, especially in accidents or inclement weather, effectively improving tunnel traffic safety and efficiency.

[0099] Methods for dynamically adjusting fusion weights include:

[0100] Define the state space: The state space consists of a state vector, which includes environmental parameters, sensor vectors, traffic vectors, traffic flow prediction values, and the deviation between the traffic flow prediction value and the actual traffic flow;

[0101] Define the action space: The action space consists of action vectors, which include weight adjustments of sensor nodes. The weight adjustment of each sensor satisfies the corresponding preset weight adjustment range, and the sum of the weights of all sensors is 1.

[0102] Design a reward function: This function is calculated by combining the deviation between the predicted traffic volume and the actual traffic volume, the stability of the prediction within a preset time period (e.g., the deviation does not exceed the preset deviation threshold for 10 consecutive minutes), sensor energy consumption (e.g., the total sensor power consumption decreases), and the volatility of the sensor fusion weight (e.g., if the sensor fusion weight is too low, resulting in missed detections).

[0103] Simulate extreme tunnel scenarios and compensate for sensor output changes based on sensor output degradation mechanisms;

[0104] Methods for compensating for variations in sensor output include:

[0105] Obtain tunnel BIM model data and simulate the tunnel scene, including pavement material, lighting system, tunnel wall texture, number of lanes, speed limit signs, entrance and exit coordinates, vehicle type distribution (such as the ratio of passenger cars to trucks), initial traffic flow density (such as peak / off-peak period parameters), and sensor installation location (such as the height and angle of the lidar installation on the tunnel roof / sidewall).

[0106] Preset multimodal extreme scenario data; for example, dense fog scenarios, such as setting the environmental visibility to 5-20 meters and the air humidity to >90%, simulate the infrared radiation characteristics of thermal imaging cameras; accident scenarios, such as triggering a vehicle collision event at a specified location, generate dynamic interference such as the accident vehicle stagnating, debris scattering, exhaust emissions (affecting sensor signals), and simultaneously simulate the sudden change of traffic flow caused by the intervention of rescue vehicles.

[0107] For each sensor, an output degradation mechanism under extreme scenarios is established. Based on the output degradation mechanism, degradation simulation is performed on different types of sensors, and the sensors are trained according to the updated parameters obtained from the degradation simulation.

[0108] Methods for simulating degradation of different types of sensors include:

[0109] For camera sensors, such as cameras in tunnels, a light intensity degradation model is built based on the attenuation coefficient, detection distance, and noise. The attenuation coefficient is positively correlated with PM2.5 concentration, and the noise term is negatively correlated with the signal-to-noise ratio.

[0110] For radar sensors: For laser sensors, such as lidar, the effective detection distance is calculated based on the scattering coefficient, absorption coefficient, transmit power, receive sensitivity, and target reflectivity; for millimeter wave sensors, such as millimeter wave radar, the corrected signal-to-noise ratio is calculated based on the original signal-to-noise ratio, combined with rainfall and detection distance.

[0111] For radar sensors, the signal power loss is also calculated based on the reflection coefficient, dynamic delay, and multipath effect.

[0112] For traffic detection sensors, such as geomagnetic sensors and piezoelectric sensors, a corrected signal voltage is calculated based on the actual signal voltage, external magnetic field strength, interference frequency, and preset calibration coefficients. An updated signal-to-noise ratio is then calculated based on the corrected signal voltage.

[0113] The sensors are trained using the updated parameters of various sensors obtained by simulating the sensor output and correcting the parameters according to the output degradation mechanism.

[0114] A high-precision simulation scenario is constructed with reference to the tunnel BIM model. Multimodal extreme scenario data such as dense fog and accidents are preset, and physical degradation models are established for different sensors such as cameras and radars. This allows for real-time simulation of the sensor performance degradation process under extreme conditions at the edge node. Training the edge-side PPO reinforcement learning algorithm based on the updated sensor parameters from the degradation simulation enables the system to learn the sensor output patterns in extreme scenarios in advance, and then dynamically adjust the fusion weights during actual operation to ensure the accuracy of multi-source data fusion. The recognition rate of abnormal sensor data in dense fog and accident scenarios remains high. This process provides the traffic prediction model with input data that is closer to real extreme environments, reducing prediction errors in emergency scenarios. At the same time, it provides a reliable sensor status basis for the generation of intelligent induction control strategies, helping the NSGA-III algorithm to quickly respond to sudden changes in traffic flow, shorten emergency response times in extreme scenarios, and significantly improve the safety and traffic efficiency of the tunnel transportation system.

[0115] Sensor data, environmental data, and historical data are used as inputs to the Actor network, and a multi-layer Transformer+CNN is used as the middle layer to obtain the dynamic weight vector of each sensor.

[0116] The sensor data, environmental data, historical data, and the error between the fused traffic prediction value and the true value are used as the input of the Critic network, and the long-term benefits of the current weight distribution strategy are output through the state value function.

[0117] Select the weight allocation strategy with the highest long-term return as the final adjustment strategy to obtain the corresponding dynamic weight.

[0118] Based on real-time weights, the fusion model and traffic flow prediction model are fed back to update traffic flow predictions and construct a three-dimensional tunnel model. The NSGA-III algorithm, combined with the real-time weights, generates optimal control strategies for the guidance system, tunnel control system, and upstream and downstream traffic lights. This feedback mechanism, based on real-time weights, dynamically optimizes the fusion accuracy of multi-source sensor data, reduces the error of the updated traffic flow predictions in extreme scenarios, and provides more accurate input for short-term trend analysis of tunnel traffic flow. The three-dimensional tunnel digital twin model constructed with real-time weights maps vehicle trajectories, sensor status, and equipment operation in real time, providing the NSGA-III algorithm with visual spatial and state constraints (such as lane conflict risk assessment and equipment power threshold verification). Based on this, the NSGA-III algorithm generates an optimal control strategy through multi-objective optimization (prediction error, response time, and system energy consumption). This strategy enables coordinated control of the guidance system, tunnel equipment, and upstream and downstream traffic lights, improving traffic efficiency in conventional scenarios and shortening emergency response time in extreme scenarios. This completes a closed-loop process of "data fusion optimization → accurate prediction → intelligent decision-making," significantly enhancing the real-time, robust, and global optimality of tunnel traffic control.

[0119] Methods for generating optimal control strategies include:

[0120] Preset optimization objectives and constraints, such as optimization objectives including: minimizing prediction error, minimizing emergency response time and minimizing system energy consumption; constraints including: communication delay is lower than the preset delay threshold, device power is lower than the preset power threshold and the optimization strategy shall not lead to lane conflict or speeding risks.

[0121] Preset chromosome encoding scheme: set the gene structure of the chromosome including weight genes and strategy genes;

[0122] The real-time weights and control strategy parameters of N groups of sensors are randomly generated. The control strategy parameters include control instructions for the induction system (such as the update frequency of the induction screen), control instructions for the tunnel control system (such as the ventilation system startup threshold), and control instructions for upstream and downstream signal lights (such as the signal light cycle). The real-time weights and control strategy parameters of the N groups of sensors are encoded into the genetic structure of N1 groups of chromosomes. Individuals that violate the constraints are eliminated to ensure that the initial population contains all feasible solutions. Among them, N and N1 are numerically equal.

[0123] The population is divided into M frontier layers according to the Pareto dominance relationship and arranged in the order of increasing dominance number. The edge node obtains the first L frontier layers and uploads them;

[0124] Methods for obtaining L frontier layers include:

[0125] Step A1, Domination Rule: For two individuals A and B, if A's target values ​​on all preset optimization targets are not lower than B's, and it is better than B on at least one target, then A is said to dominate B, and B's dominated number is increased by 1. Count the dominated numbers of all individuals according to the above domination rule;

[0126] Step A2: extract all individuals in the population that are not dominated by any other individuals, that is, individuals with a domination count of 0, to form the first frontier layer;

[0127] Step A3: Repeat step A1 to extract the dominated numbers of all individuals in the remaining individuals after the first frontier layer, and extract individuals with a dominated number of 0 to form a second frontier layer;

[0128] Step A4: Repeat step A3 to extract the currently undominated individuals one by one to form the next frontier layer until all individuals are assigned to the corresponding frontier layer or the preset maximum number of layers L is reached.

[0129] Based on the target space formed by the first L frontier layers, a uniformly distributed reference point set is generated in each dimension of the preset optimization target, where the number of reference point sets is 1.5 times the number of chromosomes in the first L frontier layers; the position of the reference point set is periodically updated according to the change of the objective function range;

[0130] The parent chromosome parameters are crossed to generate offspring. For weight genes, arithmetic crossover is used to ensure that the sum of weights after crossover is 1; for strategy genes, discrete crossover is used to randomly exchange parent parameter values.

[0131] Random perturbations are applied to offspring through mutation. For weight genes, random perturbations are applied to each weight and then renormalized. For strategy genes, Gaussian noise is added within the parameter range.

[0132] Calculate the Euclidean distance between each offspring individual and the reference point, associate the individual to the reference point with the smallest Euclidean distance, retain the first Q individuals associated with the reference point to enter the offspring, and directly retain the first P individuals on the optimal frontier of the parent generation to enter the offspring; obtain the offspring population;

[0133] Repeat the iteration of the offspring individuals until the preset termination condition is reached (such as the volatility of the objective function is less than 5% for 10 consecutive generations, or the preset maximum number of iterations is reached).

[0134] Example 2

[0135] This embodiment provides a quantum-photon hybrid computing architecture applied to Embodiment 1, including:

[0136] Deploy a quantum annealer (such as D-Wave) in the cloud for offline training of the PPO policy network and the NSGA-III multi-objective optimization model;

[0137] Optimize the objective function through quantum bit mapping, and then solve the optimized objective function, for example, convert the sensor weight distribution problem into an Ising model for solution;

[0138] Edge nodes integrate photonic matrix computing units (such as Lightmatter chips) to achieve real-time generation of PPO strategy reasoning and NSGA-III solutions to reduce latency.

[0139] Through the above-mentioned quantum-photonic hybrid computing architecture, quantum computing is responsible for global model optimization, and photonic computing handles local real-time decision-making, which can form a "cloud-edge-end" collaborative computing network, solve the real-time bottleneck of complex algorithms, and improve the response rate.

[0140] Example 3

[0141] See also Figure 2 As shown, this embodiment provides a tunnel traffic flow prediction and control system, including:

[0142] Network construction module: Deploys a sensor network in the tunnel, presetting the initial sensor weights and redundant activation rules for extreme scenarios;

[0143] Data acquisition module: The sensor network collects and transmits sensor data and traffic data to edge nodes, analyzes and obtains traffic vectors; uses sensor data and initial weights as input to the fusion model to obtain sensor vectors;

[0144] Traffic prediction module: takes the initial weight, sensor vector and traffic vector as inputs of the traffic prediction model to obtain the traffic prediction value of the future time window;

[0145] Compensation Analysis Module: This module combines sensor vectors, traffic vectors, and traffic flow predictions, uses the sensor output degradation mechanism triggered by edge nodes to simulate extreme scenarios, dynamically compensates sensor outputs, and uses the PPO reinforcement learning algorithm to adjust the compensation results to obtain real-time weights.

[0146] Intelligent control module: Based on real-time weights, it provides feedback to the fusion model and traffic prediction model, updates the traffic prediction value, builds a three-dimensional tunnel model, and then generates the optimal control strategy for the guidance system, tunnel control system, and upstream and downstream traffic lights based on the NSGA-III algorithm in combination with the real-time weights.

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person 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 based on the scope of protection of the claims.

[0148] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A tunnel traffic flow prediction and control method, characterized in that: The steps include: Deploy a sensor network in the tunnel, presetting the initial sensor weights and redundant activation rules for extreme scenarios; The sensor network collects and transmits sensor data and traffic data to edge nodes, analyzes and obtains traffic vectors; uses sensor data and initial weights as inputs to the fusion model to obtain sensor vectors; The initial weights, sensor vectors, and traffic vectors are used as inputs to the traffic prediction model to obtain traffic prediction values ​​for the future time window. Combining sensor vectors, traffic vectors, and traffic prediction values, the sensor output degradation mechanism triggered by edge nodes is used to simulate extreme scenarios, dynamically compensate the sensor output, and adjust the real-time weight based on the compensation results using the PPO reinforcement learning algorithm. Based on the real-time weights, feedback is given to the fusion model and traffic prediction model to update the traffic prediction value and build a three-dimensional tunnel model. Then, combined with the real-time weights, the optimal control strategy for controlling the guidance system, tunnel control system, and upstream and downstream traffic lights is generated based on the NSGA-III algorithm.

2. The tunnel traffic flow prediction and control method according to claim 1, characterized in that: The method of simulating extreme scenarios by utilizing the sensor output degradation mechanism triggered by edge nodes includes: When the sensor network detects that the sensor data or traffic data exceeds the preset threshold, the sensor output degradation mechanism is triggered to simulate the scenario and modify the parameters of the sensor output.

3. The tunnel traffic flow prediction and control method according to claim 2, characterized in that: The method for the output degradation mechanism to perform scenario simulation and parameter correction on the sensor output includes: For camera sensors, a light intensity degradation model is built based on the attenuation coefficient, detection distance, and noise to correct the actual light intensity. For radar sensors: For laser sensors, the effective detection distance is calculated based on the scattering coefficient, absorption coefficient, transmission power, receiving sensitivity, and target reflectivity; for millimeter wave sensors, the corrected signal-to-noise ratio is calculated based on the original signal-to-noise ratio, combined with rainfall and detection distance. For radar sensors, the signal power loss is also calculated based on the reflection coefficient, dynamic delay, and multipath effect. For traffic detection sensors, a corrected signal voltage is calculated based on the actual signal voltage, external magnetic field strength, interference frequency, and a preset calibration coefficient, and an updated signal-to-noise ratio is calculated based on the corrected signal voltage. According to the above output degradation mechanism, the scenario simulation and parameter correction of sensor output are performed on the data obtained to perform supervised learning training for various sensors.

4. The tunnel traffic flow prediction and control method according to claim 1, characterized in that: Methods for obtaining traffic vectors include: The traffic volume is obtained by counting the number of vehicle pulses passing through the geomagnetic coil; Traffic detection sensors are used to collect the time difference between vehicles passing through the tunnel to calculate the average speed of vehicles in the corresponding lanes. Detect and obtain the vehicle speed distribution of each lane, divide the speed intervals according to the preset speed intervals, and count the proportion of vehicles in each lane in each interval to obtain the vehicle speed distribution; Calculate the average vehicle speed and calculate the traffic density based on the traffic flow; The actual average time it takes for vehicles to pass through the tunnel is calculated, and the ratio of the actual average time to the free flow time is calculated to obtain the travel time ratio. The traffic density, vehicle speed distribution and travel time ratio are spliced ​​together to obtain the traffic vector.

5. The tunnel traffic flow prediction and control method according to claim 4, characterized in that: If the average vehicle speed in a lane falls below R% of the preset free-flow speed, it is marked as potentially congested. The NSGA-III algorithm is directly triggered to generate an optimal control strategy based on the historical sensor weights, and the weights and optimal control strategy are updated in real time based on updated sensor data. R is a preset threshold.

6. The tunnel traffic flow prediction and control method according to claim 1, characterized in that: Methods for obtaining sensor vectors include: Obtain sampling data from each sensor. For camera sensors, obtain the mean square error (MSE) between the sampling data and the processed data. Calculate the peak signal-to-noise ratio (PSNR) based on the MSE and the maximum possible value of the pixel value. The sampling data of the camera sensor is used as the input of the camera detection model to obtain the detection confidence of the camera sensor; For radar sensors, the number of point clouds per unit area is calculated to obtain the point cloud density. The point cloud density is compared with a preset density threshold. Areas below the preset density threshold are marked as low quality, and the marking results of all areas are counted. Through Kalman filtering, a unique ID is assigned to each physical entity (i.e., target) detected by the radar sensor. The occurrence and loss of each target ID in the continuous frame data of the radar sensor is recorded, and the corresponding loss rate is recorded. For environmental sensors, calculate the mean of historical data and the deviation rate between real-time data and the mean of historical data; The peak signal-to-noise ratio, detection confidence, labeling results, loss rate and deviation rate, as well as the corresponding initial weights are used as inputs of the fusion model to obtain the sensor vector.

7. The tunnel traffic flow prediction and control method according to claim 1, characterized in that: Methods for adjusting and obtaining real-time weights based on compensation results include: Define the state space: The state space consists of a state vector, which includes environmental parameters, sensor vectors, traffic vectors, traffic flow prediction values, and the deviation between the traffic flow prediction value and the actual traffic flow; Define the action space: The action space consists of action vectors, which include weight adjustments of sensor nodes. The weight adjustment of each sensor satisfies the corresponding preset weight adjustment range, and the sum of the weights of all sensors is 1. Design a reward function: This function is calculated by combining the deviation between the traffic prediction value and the actual traffic flow, the prediction stability within the preset time period, the sensor energy consumption, and the volatility of the sensor fusion weight. Simulate extreme tunnel scenarios and compensate for sensor output changes based on sensor output degradation mechanisms; Sensor data, environmental data, and historical data are used as inputs to the Actor network, and a multi-layer Transformer+CNN is used as the middle layer to obtain the dynamic weight vector of each sensor. The sensor data, environmental data, historical data, and the error between the fused traffic prediction value and the true value are used as the input of the Critic network, and the long-term benefits of the current weight distribution strategy are output through the state value function. Select the weight allocation strategy with the highest long-term return as the final adjustment strategy to obtain the corresponding dynamic weight.

8. The tunnel traffic flow prediction and control method according to claim 1, characterized in that: The methods for generating the optimal control strategy include: Preset optimization objectives and constraints; Preset chromosome encoding scheme: set the gene structure of the chromosome including weight genes and strategy genes; Randomly generate real-time weights and control strategy parameters for N groups of sensors. The control strategy parameters include control instructions for the induction system, tunnel control system, and upstream and downstream signal light control instructions. The real-time weights and control strategy parameters of the N groups of sensors are encoded into the genetic structure of the N1 group of chromosomes. Individuals that violate the constraints are eliminated to ensure that the initial population contains all feasible solutions. The population is divided into M frontier layers according to the Pareto dominance relationship and arranged in the order of increasing dominance number. The edge node obtains the first L frontier layers and uploads them; Based on the target space formed by the first L frontier layers, a uniformly distributed set of reference points is generated in each dimension of the preset optimization target, where the number of reference point sets is E times the number of chromosomes in the first L frontier layers; the position of the reference point set is periodically updated according to the change of the objective function range; The parent chromosome parameters are crossed to generate offspring. For weight genes, arithmetic crossover is used to ensure that the sum of weights after crossover is 1; for strategy genes, discrete crossover is used to randomly exchange parent parameter values. Random perturbations are applied to offspring through mutation. For weight genes, random perturbations are applied to each weight and then renormalized. For strategy genes, Gaussian noise is added within the parameter range. Calculate the Euclidean distance between each offspring individual and the reference point, associate the individual to the reference point with the smallest Euclidean distance, retain the first Q individuals associated with the reference point to enter the offspring, and directly retain the first P individuals on the optimal frontier of the parent generation to enter the offspring; obtain the offspring population; Repeat the iteration of the offspring individuals until the preset termination condition is reached.

9. The tunnel traffic flow prediction and control method according to claim 8, characterized in that: Methods for obtaining the first L frontier layers include: Step A1, Domination Rule: For two individuals A and B, if A's target values ​​on all preset optimization targets are not lower than B's, and it is better than B on at least one target, then A is said to dominate B, and B's dominated number is increased by 1. Count the dominated numbers of all individuals according to the above domination rule; Step A2: extract all individuals in the population that are not dominated by any other individuals, that is, individuals with a domination count of 0, to form the first frontier layer; Step A3: Repeat step A1 to extract the dominated numbers of all individuals in the remaining individuals after the first frontier layer, and extract individuals with a dominated number of 0 to form a second frontier layer; Step A4: Repeat step A3 to extract the currently undominated individuals one by one to form the next frontier layer until all individuals are assigned to the corresponding frontier layer or the preset maximum number of layers L is reached.

10. A tunnel traffic flow prediction and control system, implementing the tunnel traffic flow prediction and control method according to any one of claims 1 to 9, characterized in that: include: Network construction module: Deploys a sensor network in the tunnel, presetting the initial sensor weights and redundant activation rules for extreme scenarios; Data acquisition module: The sensor network collects and transmits sensor data and traffic data to edge nodes, analyzes and obtains traffic vectors; uses sensor data and initial weights as input to the fusion model to obtain sensor vectors; Traffic prediction module: takes the initial weight, sensor vector and traffic vector as inputs of the traffic prediction model to obtain the traffic prediction value of the future time window; Compensation Analysis Module: This module combines sensor vectors, traffic vectors, and traffic flow predictions, uses the sensor output degradation mechanism triggered by edge nodes to simulate extreme scenarios, dynamically compensates sensor outputs, and uses the PPO reinforcement learning algorithm to adjust the compensation results to obtain real-time weights. Intelligent control module: Based on real-time weights, it provides feedback to the fusion model and traffic prediction model, updates the traffic prediction value, builds a three-dimensional tunnel model, and then generates the optimal control strategy for the guidance system, tunnel control system, and upstream and downstream traffic lights based on the NSGA-III algorithm in combination with the real-time weights.

Images