Method for predicting influence factors of network performance in urban scene

Abstract

The application discloses a method for predicting an unmanned driving network performance influence factor in a city scene, and is an unmanned driving network performance influence factor prediction model based on LSTM, which comprises the following steps: step 1, a feature fusion step, serving as an input of a neural network; step 2, an unmanned driving network performance influence factor establishment step, serving as an output of the neural network; step 3, a neural network structure building step; step 4, a loss function building step; step 5, a neural network training step; and step 6, a simulation experiment verification step.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, specifically to a method for predicting factors affecting the performance of autonomous driving networks in urban scenarios, and more specifically to a prediction model for factors affecting the performance of autonomous driving networks based on Long Short-Term Memory (LSTM) networks. Background Technology

[0002] Autonomous vehicles rely on onboard sensors to perceive their surroundings and control the vehicle based on these perceptions, enabling them to operate safely and automatically without human intervention. Initially, autonomous vehicles may be expensive and have limited performance; however, starting in the 2030s and 40s, their cost will gradually decrease, making them more accessible to the general public. It is projected that by the end of the 2020s, fully autonomous vehicles (Level 5) will be available and legally used in certain regions.

[0003] While autonomous vehicles can use onboard sensors to perceive their surroundings, blind spots and obstructed views are inherent problems during vehicle movement, especially in foggy or rainy / snowy weather. These conditions significantly reduce the sensing range of the onboard sensors, preventing autonomous vehicles from fully and accurately perceiving their environment and leading to potential safety hazards. Combining autonomous vehicles with wireless communication technology enables communication and information sharing between them, providing a new approach to addressing the blind spots and obstructed views issues that exist during autonomous vehicle movement.

[0004] While intelligent connected vehicle technology enables information sharing between vehicles, communication between autonomous vehicles is affected by external interference in urban scenarios, such as human-driven vehicles, roadside obstacles, traffic lights, and pedestrians. To accurately analyze the connectivity quality between autonomous vehicles and enable them to accurately select highly coupled and low-latency neighboring vehicles for perception result sharing, existing methods primarily predict and analyze the wireless communication quality between vehicle nodes in vehicle networks. However, applying these methods to autonomous driving urban scenarios still faces the following challenges:

[0005] 1) Failure to consider the impact of various external interference factors on vehicle communication: Existing wireless communication quality analysis and prediction methods for vehicle-to-everything (V2X) only consider the impact of neighboring vehicles on wireless communication between vehicle nodes. In urban scenarios, communication between autonomous vehicles may be affected by external interference such as driven vehicles, roadside obstacles, traffic lights, and pedestrians. Existing wireless communication quality prediction and analysis methods for V2X cannot analyze the impact of different external interferences on communication between autonomous vehicles in urban scenarios.

[0006] 2) Incomplete factors affecting wireless communication network performance: Existing wireless communication quality analysis and prediction methods for vehicle-to-everything (V2X) networks can only predict whether vehicles are connected and the duration of communication, lacking factors affecting wireless communication network performance such as delivery rate and latency.

[0007] To overcome this problem, this invention presents a predictive model for factors affecting the performance of autonomous driving networks. Summary of the Invention

[0008] Purpose of the invention:

[0009] To address the shortcomings of existing methods for analyzing and predicting the quality of vehicle-to-everything (V2X) wireless networks, such as failing to consider the impact of various external interference factors on vehicle communication and the incompleteness of factors affecting wireless communication network performance, this invention presents an LSTM-based prediction model for the performance influencing factors of autonomous driving networks. This model includes steps such as feature fusion, identification of factors affecting autonomous driving network performance, neural network structure construction, loss function construction, and neural network training. This allows for the filtering out of loosely coupled and high-latency neighbor nodes, ensuring stable information sharing among autonomous vehicles and achieving stable and orderly vehicle movement. This invention represents original basic research.

[0010] The technical solution provided by this invention specifically includes the following steps:

[0011] Step 1 Feature Fusion Step

[0012] Step 2: Identifying Factors Affecting Autonomous Driving Network Performance

[0013] Step 3: Neural Network Structure Construction Steps

[0014] Step 4: Loss Function Construction

[0015] Step 5: Neural Network Training Steps

[0016] Step 6: Simulation Experiment Verification

[0017] Beneficial effects

[0018] This invention is applied to the driving environment of autonomous vehicles in urban scenarios, providing a predictive model for factors affecting the performance of autonomous driving networks, ensuring the stability of information sharing between autonomous vehicles, and achieving the goal of stable and orderly movement behavior of autonomous vehicles. Attached Figure Description

[0019] Figure 1 SUMO simulation software and manually designed simulation scene diagrams

[0020] Figure 2 SUMO simulation software and simulation scenario diagram of actual road network

[0021] Figure 3 Framework diagram of prediction model for factors affecting network performance of autonomous driving

[0022] Figure 4 Mean absolute error at maximum speeds of different vehicles in a manually designed scenario

[0023] Figure 5 Mean square error of different vehicle maximum speeds in a manually designed scenario

[0024] Figure 6 Coefficient of determination for different maximum vehicle speeds in artificially designed scenarios

[0025] Figure 7 Mean absolute error under different numbers of vehicles in real road network scenarios

[0026] Figure 8 Mean square error under different numbers of vehicles in real road network scenarios

[0027] Figure 9 Determination coefficients under different numbers of vehicles in real-world road network scenarios

[0028] Figure 10 Average absolute error at different maximum vehicle speeds in real-world road network scenarios

[0029] Figure 11 Mean square error of different maximum vehicle speeds in real-world road network scenarios

[0030] Figure 12 Determination coefficients for different maximum vehicle speeds in real-world road network scenarios

[0031] Figure 13 This is a flowchart of the method of the present invention.

[0032] Figure 14 Table 1. Simulation Experiment Configuration in Artificially Designed Scenarios

[0033] Figure 15 Table 2 Simulation Experiment Configuration in Actual Road Network Scenarios Detailed Implementation

[0034] The specific implementation process of this invention includes the following six aspects:

[0035] Step 1 Feature Fusion

[0036] Step 2: Identifying Factors Affecting Autonomous Driving Network Performance

[0037] Step 3: Building the Neural Network Structure

[0038] Step 4: Construction of the loss function

[0039] Step 5: Neural Network Training

[0040] Step 6: Simulation Experiment Verification

[0041] Details are as follows:

[0042] In this article, "network performance" is short for "autonomous driving network performance".

[0043] In the text, "NPIFP" is short for "Network Performance Influencing Factors Prediction".

[0044] Step 1 Feature Fusion Step

[0045] This step involves fusing features from the autonomous vehicle and external disturbances to generate the input features for the LSTM. The relevant definitions of autonomous vehicles and external disturbances in urban scenarios are as follows:

[0046] Definition 1: In urban scenarios, autonomous vehicles and external interference are...

[0047]

[0048] in, It represents a collection of driverless vehicles. This refers to an unmanned vehicle. Represents the set of positive integers. This indicates a group of manned vehicles. This indicates a manned vehicle. This represents a collection of roadside obstacles. This represents a roadside obstacle. Indicates the collection of traffic lights. This represents a traffic light. Indicates a group of pedestrians. It represents a pedestrian.

[0049] Definition 2: Driverless vehicles exist The characteristics of time are

[0050]

[0051] in, yes exist The speed of time, yes exist acceleration at any moment and They represent exist The horizontal and vertical coordinates of the current position at any given time. yes exist The direction of travel at any given moment.

[0052] Definition 3: Manned vehicle exist The characteristics of time are

[0053]

[0054] in, yes exist The speed of time, yes exist acceleration at any moment and They represent exist The x and y coordinates of the current position at any given time. yes exist The direction of travel at any given moment.

[0055] Definition 4: Roadside obstacles exist The characteristics of time are

[0056]

[0057] in, and They represent The x and y coordinates of the location, and They represent Length and width.

[0058] Definition 5: Traffic lights exist The characteristics of time are

[0059]

[0060] in, and They represent The x and y coordinates of the location, yes exist The state at any given moment, among which, 1 and 2 represent red light, green light and yellow light respectively.

[0061] Definition 6: Pedestrian exist The characteristics of time are

[0062]

[0063] in, yes exist The speed of time and They represent exist The x and y coordinates of the current position at any given time. express exist The direction of travel at any given moment.

[0064] Definition 7: Driverless vehicles With self-driving vehicles or With manned vehicles The relative characteristics between them are

[0065]

[0066] in, , , and They represent and or and Between The relative velocity, acceleration, distance, and direction of motion at any given moment.

[0067] This invention presents an LSTM-based prediction model for factors influencing the performance of autonomous driving networks. The method comprises two parts: feature fusion and prediction of these factors. Feature fusion combines features from the autonomous vehicle and external interference to generate the input features for the LSTM. In predicting the performance of autonomous vehicles... and When considering factors affecting network performance, the input characteristics of LSTM are:

[0068]

[0069] in, express and external interference Characteristics of time, for time A group of driverless vehicles gathered around. , , and They represent time There are drivers, roadside obstacles, traffic lights, and pedestrians gathering nearby. Similarly, express and external interference Characteristics of a moment.

[0070] Step 2: Identifying Factors Affecting Autonomous Driving Network Performance

[0071] This step analyzes the metrics affecting connectivity quality between autonomous vehicles and determines the LSTM output. The main factors influencing autonomous vehicle network performance include the probability of communication interruption between autonomous vehicles, data correct delivery rate, link capacity, connectivity duration, data transmission latency, and round-trip time, with the following specific meanings:

[0072] 1) Probability of communication interruption : Indicates driverless vehicles and The probability that the communication link between them may be interrupted at some point in the future.

[0073] 2) Data correct delivery rate : Indicates driverless vehicles and The probability of correct data delivery between them.

[0074] 3) Link capacity : Indicates driverless vehicles and The amount of data successfully transmitted per unit time in the communication link between them.

[0075] 4) Connection duration : Indicates an autonomous vehicle that is currently communicating. and The longest time that the communication link between them can remain connected.

[0076] 5) Data transmission delay : Towards Sending data indicates The time required to send from the first bit to the last bit.

[0077] 6) Communication round-trip time : Towards Sending data indicates from Data sent received The confirmation time.

[0078] Step 3: Neural Network Structure Construction

[0079] This step involves constructing an LSTM-based neural network structure. Because the motion of autonomous vehicles is time-dependent—meaning the current communication state between autonomous vehicles is closely related to the previous state—and LSTM is suitable for processing and predicting time-series data, this invention inputs the fused features into the LSTM neural network, processing them through the LSTM input layer, hidden layer, and output layer. Then, the features from the LSTM output layer are input into a fully connected layer for linear transformation. Finally, the fully connected layer outputs the predicted values ​​of the factors influencing the performance of the autonomous driving network.

[0080] Step 4: Loss Function Construction Steps

[0081] As shown in step 2, predicting the factors influencing the performance of autonomous driving networks is a regression problem. Therefore, we choose the mean squared error as the loss function, and the objective function for predicting the factors influencing the performance of autonomous driving networks based on LSTM is:

[0082]

[0083] in, Represents the weights in a neural network. Indicates the number of training samples. Indicates the first Input features of each sample Indicates the first In the nth sample The true values ​​of factors affecting the performance of autonomous driving networks Indicates the first In the nth sample Predicted values ​​of factors affecting the performance of autonomous driving networks.

[0084] Step 5: Training the Neural Network

[0085] The specific steps of the neural network training method for predicting the factors affecting the performance of autonomous driving networks based on LSTM are as follows:

[0086] (1) Initialize the weight parameters of the LSTM neural network and divide the training set into several sub-training sets of the same size.

[0087] (2) For all input samples in a subset, firstly, the predicted values ​​of the factors affecting the performance of the autonomous driving network are calculated by forward propagation through the LSTM neural network, then the loss is calculated, and finally the weights of the LSTM neural network are updated by backpropagation. .

[0088] (3) Repeat step (2) for all sub-training sets to update the weight parameters of the LSTM neural network.

[0089] (4) After all sub-training sets have been iterated, save the structure of the LSTM neural network and the weights between nodes. The training of the LSTM neural network is now complete.

[0090] After the neural network is trained, a predictive model of factors affecting the performance of autonomous driving networks can be obtained, applicable to urban scenarios. This model helps autonomous vehicles filter out neighboring autonomous vehicles with low coupling and high latency, enabling them to establish data-sharing relationships with neighboring autonomous vehicles that have better connectivity (lower communication interruption probability, lower data transmission latency and round-trip time, higher data correct delivery rate, link capacity, and longer connection duration). This allows autonomous vehicles to share perception results and decision-making information with neighboring autonomous vehicles in real time and accurately during operation, improving their perception of the driving environment and driving safety, ultimately achieving stable and orderly autonomous vehicle behavior.

[0091] Step 6: Simulation Experiment Verification

[0092] The present invention applies the model trained in step 5 to two urban scenarios to verify the effectiveness of the model in the two urban scenarios.

[0093] (1) Simulation Experiment Data and Methods

[0094] NS3 is an open-source discrete event network simulator. SUMO is a tool capable of simulating microscopic and continuous traffic flow. This invention uses SUMO and NS3 respectively to simulate the movement and communication behavior of autonomous vehicles. This invention uses two scenarios—manually designed and real-world road networks—to verify the effectiveness of the predictive model for factors influencing the performance of autonomous driving networks.

[0095] Artificially designed scenarios, such as Figure 1 As shown, the road network structure is a crossroads with a length of 10,200 meters horizontally and vertically. The simulation includes 356 driverless vehicles, 44 manned vehicles, 41 traffic lights, 40 roadside obstacles, and 400 pedestrians. A 200-second segment of the simulation process is captured. To cover various driving scenarios from low to high speeds, the maximum speed of the driverless vehicles is set to 5, 10, 15, 20, 25, and 30 m / s, respectively. Other parameters are shown in Table 1.

[0096] Real-world road network scenarios, such as Figure 2As shown, the road network structure is a 3km by 2km section of the Shanghai urban area captured using OpenStreetMap. This road network contains 3469 road segments and 94 traffic lights, simulating 60 roadside obstacles and 600 pedestrians. A 200-second segment of the entire simulation process was captured. To cover various driving scenarios from low to high density, the number of autonomous vehicles was set to 600, 700, 800, 900, and 1000 respectively. To cover various driving scenarios from low to high speed, the maximum speed of the autonomous vehicles was set to 30, 40, 50, 60, and 70 km / h respectively. Other parameters are shown in Table 2.

[0097] The framework of the prediction model for factors affecting the performance of autonomous driving networks is as follows: Figure 3 As shown, the system comprises two modules: feature fusion and prediction of factors influencing the performance of the autonomous driving network. Feature fusion combines and concatenates features from the autonomous vehicle and external interference to generate the input features for the LSTM. The lengths of traffic lights, pedestrians, roadside obstacles, and relative features are 12, 40, 16, 50, and 170, respectively. The LSTM consists of an input layer, hidden layers, and an output layer, with 288, 512, and 256 nodes in each layer, respectively. The fully connected layer has a 5-layer structure, with 256, 128, 64, 32, and 6 nodes in each layer, respectively. The output of the fully connected layer is the predicted value of the factors influencing the performance of the autonomous driving network.

[0098] To verify the effectiveness of the Network Performance Influencing Factors Prediction (NPIFP) model for autonomous driving networks in this invention, the simulation experiment is designed as follows:

[0099] 1) Import the manually designed and actual road network scenarios into SUMO and configure the relevant parameters, then export vehicle movement models under different speeds and densities in the two scenarios.

[0100] 2) Use NS3 to realize communication between autonomous vehicles, and collect statistics on the interference information of autonomous vehicles and their surroundings, as well as the network performance impact factors of other autonomous vehicles per second.

[0101] 3) Use the collected data to predict the factors affecting the performance of autonomous driving networks, and calculate the absolute error, mean square error, and coefficient of determination.

[0102] 4) Combine NPIFP with Chen and Guestrin [1] The proposed XGBoost (a scalable end-to-end tree boosting system) algorithm and Gao et al. [2]The proposed MLP (Multi-Layer Perceptron) algorithm is compared with NPIFP, and the advantages and disadvantages of NPIFP are analyzed.

[0103] (2) Analysis of simulation experiment results

[0104] 1) Prediction results in manually designed scenarios

[0105] Mean absolute error (MAE) measures the absolute error between the actual and predicted values ​​of factors affecting the performance of autonomous driving networks. A smaller MAE indicates a more accurate assessment of these factors. In a manually designed scenario, the maximum speed of the autonomous vehicle was set to 5, 10, 15, 20, 25, and 30 m / s, and the MAE of the three methods at different maximum speeds was statistically analyzed. The results are as follows: Figure 4 As shown, the mean absolute error (MAO) of MLP is greater than that of NPIFP and XGBoost. When predicting data transmission delay, the MAOs of NPIFP and XGBoost are the same. However, when predicting communication outage probability, data correct delivery rate, link capacity, connection duration, and round-trip time, the MAO of NPIFP is lower than that of XGBoost. Therefore, NPIFP has a smaller MAO and higher prediction accuracy.

[0106] The mean squared error (MSE) is the sum of squares of the errors between the predicted and actual values ​​of factors affecting the performance of autonomous driving networks. A smaller MSE indicates more accurate predictions of these factors. In a manually designed scenario, the maximum speed of the autonomous vehicle was set to 5, 10, 15, 20, 25, and 30 m / s, and the MSE of the three methods at different maximum speeds was statistically analyzed. The results are as follows: Figure 5 As shown, NPIFP and XGBoost have the same mean squared error when predicting communication outage probability, connection duration, and data transmission delay. However, NPIFP has a smaller mean squared error than XGBoost when predicting data correct delivery rate, link capacity, and round-trip time. Therefore, NPIFP has a smaller mean squared error and higher prediction accuracy.

[0107] The coefficient of determination reflects the goodness of fit of the autonomous driving network performance prediction model. A coefficient of determination closer to 1 indicates a better fit. In a manually designed scenario, the maximum speed of the autonomous vehicle was set to 5, 10, 15, 20, 25, and 30 m / s, and the coefficients of determination for the three methods at different maximum speeds were calculated. The results are as follows: Figure 6 As shown, NPIFP's coefficient of determination is closer to 1 and higher than XGBoost and MLP. NPIFP has the lowest coefficient of determination when predicting data transmission delay because data transmission delay is more variable than other network performance factors.

[0108] 2) Prediction results in actual road network scenarios

[0109] In real-world road network scenarios, the number of autonomous vehicles was set to 600, 700, 800, 900, and 1000, respectively. The mean absolute error of the three methods under different vehicle densities was statistically analyzed, and the results are as follows: Figure 7 As shown, the mean absolute error (MAO) of MLP is greater than that of NPIFP and XGBoost. When predicting data delivery accuracy, link capacity, and communication round-trip time, the MAO of NPIFP is lower than that of XGBoost. Compared to manually designed scenarios, the MAO of NPIFP is higher because real-world road network scenarios are more complex, making communication relationships between vehicles more volatile. Based on the above results, NPIFP has a smaller MAO and higher prediction accuracy.

[0110] In real-world road network scenarios, the mean square errors of the three methods under different vehicle densities are as follows: Figure 8 As shown, the mean squared error of MLP is greater than that of NPIFP and XGBoost. Similar to the mean absolute error, NPIFP has a smaller mean squared error than XGBoost when predicting data delivery accuracy, link capacity, and round-trip time. Therefore, NPIFP has higher prediction accuracy.

[0111] In real-world road network scenarios, the determination coefficients of the three methods under different vehicle densities are as follows: Figure 9 As shown, the coefficient of determination for NPIFP is higher than that for MLP. When predicting link capacity, the coefficients of determination for NPIFP and XGBoost are essentially the same. However, when predicting other factors affecting network performance, the coefficient of determination for NPIFP is higher than that for XGBoost.

[0112] In a real-world road network scenario, the maximum speeds of autonomous vehicles were set to 30, 40, 50, 60, and 70 km / h, respectively. The average absolute error of the three methods at different maximum speeds was statistically analyzed, and the results are as follows: Figure 10 As shown, the mean absolute error of MLP is greater than that of NPIFP and XGBoost. When predicting communication outage probability, data correct delivery rate, link capacity, and communication round-trip time, the mean absolute error of NPIFP is lower than that of XGBoost. Therefore, NPIFP has higher prediction accuracy.

[0113] In real-world road network scenarios, the mean square errors of the three methods at different maximum speeds are as follows: Figure 11As shown, the mean absolute error of MLP is greater than that of NPIFP and XGBoost. However, NPIFP has a lower mean absolute error than XGBoost in terms of data correct delivery rate, link capacity, and round-trip time. Therefore, NPIFP has a lower mean squared error and higher prediction accuracy.

[0114] In real-world road network scenarios, the determination coefficients of the three methods at different maximum speeds are as follows: Figure 12 As shown, the coefficient of determination for NPIFP is closer to 1 and higher than that for XGBoost and MLP. Therefore, NPIFP has a higher prediction accuracy.

[0115] In summary, NPIFP's prediction results achieve low absolute error and mean square error, as well as a high coefficient of determination, under different speeds and densities in both artificially designed and real-world road network scenarios. Therefore, NPIFP's prediction accuracy is relatively high.

[0116] Innovation and application value

[0117] This invention addresses the shortcomings of existing vehicle-to-everything (V2X) wireless network quality analysis and prediction methods, which fail to consider the impact of various external interference factors on vehicle communication and lack a comprehensive understanding of factors affecting wireless communication network performance. For the first time, this invention creatively discloses a predictive model for the influencing factors of autonomous driving network performance in urban scenarios. In urban applications, each autonomous vehicle can use this model to predict and select neighboring autonomous vehicles with superior communication quality within its current network and establish data sharing relationships with them. This allows each autonomous vehicle to connect to its vehicle group during movement through neighboring vehicles with superior communication quality, thereby sharing perception results with the group in real time and obtaining decision-making information from the group. This ensures the autonomous vehicle's perception of the driving environment and driving safety, achieving stable and orderly autonomous vehicle behavior.

[0118] As can be seen from the design principle of the prediction model in this invention, the prediction model evaluates the connectivity quality indicators between autonomous vehicles and neighboring vehicles, including:

[0119] 1. The lower the three indicators—probability of communication interruption, data transmission delay, and round-trip time—the better the connectivity quality.

[0120] 2. The higher the values ​​of data correct delivery rate, link capacity, and connection duration, the better the connection quality.

[0121] Better neighboring autonomous vehicles can be identified and data-sharing relationships can be established with them.

[0122] [1] Chen T, Guestrin C. Xgboost: A scalable tree boosting system[C] / / Proceedings of the 22nd Acm Sigkdd International Conference on KnowledgeDiscovery and Data Mining. 2016: 785-794.

[0123] [2] Gao S, Zhou M, Wang Y, et al. Dendritic neuron model witheffective learning algorithms for classification, approximation, andprediction[J]. IEEE Transactions on Neural Networks and Learning Systems,2018, 30(2): 601-614.

Claims

1. A method for predicting factors affecting the performance of autonomous driving networks in an urban scenario, characterized by: This paper presents a prediction model for factors affecting the performance of autonomous driving networks based on LSTM, including feature fusion, identification of factors affecting the performance of autonomous driving networks, construction of neural network structure, construction of loss function, and neural network training steps. Step 1: Feature fusion step, used as input to the neural network; Step 2: Identifying the factors influencing the performance of autonomous driving networks, which will be used as the output of the neural network; Step 3: Neural Network Structure Construction Steps Step 4: Loss Function Construction Step 5: Neural Network Training Steps Step 1. Feature Fusion Step The relevant definitions of autonomous vehicles and external interference in urban scenarios are as follows: Definition 1: In urban scenarios, autonomous vehicles and external interference are... in, It represents a collection of driverless vehicles. This refers to an unmanned vehicle. Represents the set of positive integers. This indicates a group of manned vehicles. This indicates a manned vehicle. This represents a collection of roadside obstacles. This represents a roadside obstacle. Indicates the collection of traffic lights. This represents a traffic light. Indicates a group of pedestrians. Indicates a pedestrian; Definition 2: Driverless vehicles exist The characteristics of time are in, yes exist The speed of time, yes exist acceleration at any moment and They represent exist The x and y coordinates of the current position at any given time. yes exist The direction of travel at any given time; Definition 3: Manned vehicle exist The characteristics of time are in, yes exist The speed of time, yes exist acceleration at any moment and They represent exist The x and y coordinates of the current position at any given time. yes exist The direction of travel at any given time; Definition 4: Roadside obstacles exist The characteristics of time are in, and They represent The x and y coordinates of the location, and They represent Length and width; Definition 5: Traffic lights exist The characteristics of time are in, and They represent The x and y coordinates of the location, yes exist The state at any given moment, among which, 1 and 2 represent red light, green light and yellow light respectively; Definition 6: Pedestrian exist The characteristics of time are in, yes exist The speed of time and They represent exist The x and y coordinates of the current position at any given time. express exist The direction of travel at any given moment; Definition 7: Driverless vehicles With self-driving vehicles or With manned vehicles The relative characteristics between them are in, , , and They represent and or and Between The relative velocity, acceleration, distance, and direction of motion at any given moment; Predicting self-driving vehicles and When considering factors affecting network performance, the input characteristics of LSTM are: in, express and external interference Characteristics of time, for time A group of driverless vehicles gathered around. , , and They represent time There are drivers, roadside obstacles, traffic lights, and pedestrians gathered nearby; similarly, express and external interference Characteristics of a moment.

2. The prediction method as described in claim 1, characterized in that, Step 2: Identifying Factors Affecting Autonomous Driving Network Performance Analyze the indicators that affect the connectivity quality between autonomous vehicles and determine the output of the LSTM; This includes the probability of communication interruption between autonomous vehicles, data correct delivery rate, link capacity, connection duration, data transmission delay, and communication round-trip time, with the following specific meanings: 1) Probability of communication interruption : Indicates driverless vehicles and The probability that the communication link between them may be interrupted at some point in the future; 2) Data correct delivery rate : Indicates driverless vehicles and The probability of correct data delivery between them; 3) Link capacity : Indicates driverless vehicles and The amount of data successfully transmitted per unit time in the communication link between them; 4) Connection duration : Indicates an autonomous vehicle that is currently communicating. and The longest time that the communication link between them can remain connected; 5) Data transmission delay : Towards Sending data indicates The time required to send from the first bit to the last bit; 6) Communication round-trip time : Towards Sending data indicates from Data sent received The confirmation time.

3. The prediction method as described in claim 1, characterized in that, Step 3: Neural Network Structure Construction This step involves constructing an LSTM-based neural network structure; The fused features are input into an LSTM neural network and processed through the LSTM input layer, hidden layer, and output layer. Then, the features from the LSTM output layer are input into a fully connected layer for linear transformation. Finally, the fully connected layer outputs the predicted values ​​of the factors affecting the performance of the autonomous driving network.

4. The prediction method as described in claim 1, characterized in that, Step 4. Loss Function Construction Steps The objective function for predicting the factors affecting the performance of autonomous driving networks based on LSTM is: in, Represents the weights in a neural network. Indicates the number of training samples. Indicates the first Input features of each sample Indicates the first In the nth sample The true values ​​of factors affecting the performance of autonomous driving networks Indicates the first In the nth sample Predicted values ​​of factors affecting the performance of autonomous driving networks.

5. The prediction method as described in claim 1, characterized in that, Step 5: Neural Network Training Steps The specific steps are as follows: (1) Initialize the weight parameters of the LSTM neural network and divide the training set into several sub-training sets of the same size; (2) For all input samples in a subset, firstly, the predicted values ​​of the factors affecting the performance of the autonomous driving network are calculated by forward propagation through the LSTM neural network, then the loss is calculated, and finally the weights of the LSTM neural network are updated by backpropagation. ; (3) Repeat step (2) for all sub-training sets to update the weight parameters of the LSTM neural network; (4) After all sub-training sets have been iterated, save the structure of the LSTM neural network and the weights between nodes. The training of the LSTM neural network is then complete. Once the neural network is trained, a predictive model for factors affecting the performance of autonomous driving networks can be obtained and applied to urban scenarios.

Images