Vehicle-mounted edge computing task scheduling model and method based on CNN automatic feature generation
By adopting CNN-based automatic feature generation model and multi-head VEC task offloading decision model in on-board edge computing, the problem of inefficient scheduling in a highly dynamic environment is solved, and more efficient, accurate and adaptable task offloading decisions are achieved. .
Patent Information
- Application Number
- CN202411511574.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-28
- Publication Date
- 2025-02-18
AI Technical Summary
Traditional on-board edge computing task scheduling strategies cannot effectively deal with highly dynamic and heterogeneous environments, resulting in increased task scheduling delays and even task failures.
Using CNN-based automatic feature generation model and multi-head VEC task offload decision model (DMNet), the task scheduling features are automatically extracted and reorganized through convolutional layer, pooling layer and recombination layer, and the multi-head model is used to perform independent calculations based on different access technologies. and decision making, and ultimately comprehensive outputs the optimal task offload decision.
It improves computational unloading efficiency, reduces decision complexity, enhances adaptability, improves feature extraction accuracy and model generalization capabilities, and ensures effective unloading decisions in a highly dynamic environment.
Abstract
Description
Technical Field
[0001] The present invention relates to the field of vehicle-mounted edge computing technology, and more specifically to a vehicle-mounted edge computing task scheduling model and method based on CNN automatic feature generation. Background Art
[0002] In the vehicle edge computing (VEC) system, task scheduling is a key link to ensure that vehicle users can effectively offload computing tasks to edge servers. Traditional task scheduling strategies usually rely on static rules or simple heuristic algorithms, which often fail to make optimal decisions when faced with complex network environments and dynamically changing resource states. For example, some early strategies may only consider the computing requirements and network bandwidth of the task, while ignoring the real-time nature of the task and the mobility of the vehicle, resulting in low scheduling efficiency in highly dynamic environments.
[0003] In addition, although some scheduling strategies based on machine learning can optimize decisions by learning from historical data, they usually require a lot of manual feature engineering to extract key information about tasks and networks. These manually designed features are not only time-consuming and labor-intensive, but may not cover all factors that affect scheduling decisions, limiting the generalization ability and adaptability of the model.
[0004] In the VEC system, the decision of computing offloading needs to consider not only the computing requirements of the task and the network conditions, but also the real-time prediction of the mobility of the vehicle and the resource status of the edge server. Traditional scheduling strategies are often unable to effectively handle such highly dynamic and heterogeneous environments, resulting in increased delays in task scheduling and even task failures.
[0005] To solve these problems, an automatic feature generation model based on convolutional neural network (CNN) is proposed, which can automatically learn the features that have an impact on task scheduling from raw data without manual intervention. This model automatically extracts features through convolutional layers and eliminates the influence of feature input order by designing reorganization layers, thereby improving the robustness and adaptability of the model. Compared with traditional manual feature engineering, the CNN-based automatic feature generation model can more comprehensively capture the key information of tasks and networks, providing a more efficient and accurate solution for VEC task scheduling. Summary of the invention
[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a vehicle-mounted edge computing task scheduling model and method based on CNN automatic feature generation.
[0007] The technical solution of the present invention is as follows:
[0008] A vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation includes: a CNN-based automatic feature generation model, a multi-head VEC task offloading decision model (DMNet), and a historical task scheduling data processing module, wherein:
[0009] 1) CNN-based automatic feature generation model: including convolution layer, pooling layer, and reorganization layer;
[0010] Convolution layer: uses 3×3 convolution kernel, step size of 1, and number of 64. The features of input data are extracted through convolution operation;
[0011] Pooling layer: uses 2×2 maximum pooling with a step size of 2 to reduce feature dimensions and reduce the amount of calculation;
[0012] Reorganization layer: By randomly shuffling and recombining features, the influence of feature input order on model performance is avoided;
[0013] Working principle: The input raw task scheduling data first passes through the convolution layer for feature extraction to obtain a preliminary feature map. Then it passes through the pooling layer for dimensionality reduction to reduce the number of features. Finally, it passes through the reorganization layer for feature reorganization to eliminate the influence of the input order.
[0014] 2) Multi-head VEC task offloading decision model (DMNet): It includes multiple heads, each head corresponds to an access technology (WLAN, WAN, CN).
[0015] Each head contains a fully connected layer and an output layer.
[0016] The number of neurons in the fully connected layer is 256, and the activation function is ReLU.
[0017] The number of neurons in the output layer is determined according to the specific decision category, and the Softmax activation function is used for classification.
[0018] Working principle: The features output by the CNN automatic feature generation model are input into each head of DMNet. Each head performs independent calculations and decisions based on the characteristics and requirements of the corresponding access technology, and finally the output results of multiple heads are combined to determine the optimal task offloading decision.
[0019] 3) Historical task scheduling data processing module: including data collection, cleaning and preprocessing.
[0020] Data collection: Task scheduling related data is collected from the server of the vehicle-mounted edge computing system every 5 minutes, including task type, computing requirements, network bandwidth, vehicle location, server resource status, etc.
[0021] Data cleaning: remove duplicate data, outliers and missing values. The criterion for outliers is data exceeding the mean ± 3 times the standard deviation.
[0022] Data preprocessing: Standardize the data to make its mean 0 and variance 1.
[0023] Working principle: The collected historical task scheduling data is cleaned and preprocessed, and then used to train the CNN-based automatic feature generation model and DMNet model to improve the prediction accuracy and generalization ability of the model.
[0024] A method for scheduling vehicle-mounted edge computing tasks based on CNN automatic feature generation includes the following steps:
[0025] Step 1: Real-time monitoring of task generation and resource status changes in the vehicle-mounted edge computing system. The monitoring frequency is once per second to obtain information such as the task's computing volume, deadline, priority, and resource status information such as the vehicle's location, speed, network connection status, and edge server computing power and bandwidth.
[0026] Step 2: Input the monitored real-time data into the CNN-based automatic feature generation model to generate task features.
[0027] Step 3: Input the generated task features into the multi-head VEC task offloading decision model (DMNet).
[0028] Step 4: The DMNet model calculates the task offloading decision results under each access technology based on the characteristics of different access technologies and the current task characteristics, including whether to offload or not, the selection of offloading servers, etc.
[0029] Step 5: Combine the decision results of multiple heads and select the best decision for task offloading. For example, if the decision result of the WLAN head is to offload to a nearby WLAN access point, the decision result of the WAN head is to offload to a remote WAN server, and the decision result of the CN head is to offload to a cellular network base station, consider network latency, bandwidth, server load and other factors to select the best decision.
[0030] Step 6: Execute the task offloading operation according to the optimal decision, and monitor the task execution and resource usage in real time. If an abnormality occurs during the task execution, such as network interruption, server failure, etc., adjust the task scheduling strategy in time.
[0031] Technical effects and advantages of the present invention:
[0032] 1. Improve computing offloading efficiency: By introducing a deep learning-based on-board edge computing task scheduling strategy, it is possible to more accurately predict resource status and scheduling results, thereby improving the efficiency of V2I computing offloading.
[0033] 2. Reduce decision complexity: The DMNet model reduces the workload of manually designing combined features through automatic feature generation, and simplifies the process of the system making unloading decisions.
[0034] 3. Enhanced adaptability: The DMNet model can adapt to highly dynamic computing and network resource environments and make effective offloading decisions even when resource status changes rapidly.
[0035] 4. Improve the accuracy of feature extraction: Utilize the convolutional layer and recombination layer of CNN to automatically extract and reorganize features, avoid the influence of feature input order on model performance, and improve the accuracy of feature extraction and the generalization ability of the model. DETAILED DESCRIPTION
[0036] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0037] Embodiment 1
[0038] A vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation includes: a CNN-based automatic feature generation model, a multi-head VEC task offloading decision model (DMNet), and a historical task scheduling data processing module, wherein:
[0039] 1) CNN-based automatic feature generation model: including convolution layer, pooling layer, and reorganization layer;
[0040] Convolution layer: uses 3×3 convolution kernel, step size of 1, and number of 64. The features of input data are extracted through convolution operation;
[0041] Pooling layer: uses 2×2 maximum pooling with a step size of 2 to reduce feature dimensions and reduce the amount of calculation;
[0042] Reorganization layer: By randomly shuffling and recombining features, the influence of feature input order on model performance is avoided;
[0043] Working principle: The input raw task scheduling data first passes through the convolution layer for feature extraction to obtain a preliminary feature map. Then it passes through the pooling layer for dimensionality reduction to reduce the number of features. Finally, it passes through the reorganization layer for feature reorganization to eliminate the influence of the input order.
[0044] 2) Multi-head VEC task offloading decision model (DMNet): It includes multiple heads, each head corresponds to an access technology (WLAN, WAN, CN).
[0045] Each head contains a fully connected layer and an output layer.
[0046] The number of neurons in the fully connected layer is 256, and the activation function is ReLU.
[0047] The number of neurons in the output layer is determined according to the specific decision category, and the Softmax activation function is used for classification.
[0048] Working principle: The features output by the CNN automatic feature generation model are input into each head of DMNet. Each head performs independent calculations and decisions based on the characteristics and requirements of the corresponding access technology, and finally the output results of multiple heads are combined to determine the optimal task offloading decision.
[0049] 3) Historical task scheduling data processing module: including data collection, cleaning and preprocessing.
[0050] Data collection: Task scheduling related data is collected from the server of the vehicle-mounted edge computing system every 5 minutes, including task type, computing requirements, network bandwidth, vehicle location, server resource status, etc.
[0051] Data cleaning: remove duplicate data, outliers and missing values. The criterion for outliers is data exceeding the mean ± 3 times the standard deviation.
[0052] Data preprocessing: Standardize the data to make its mean 0 and variance 1.
[0053] Working principle: The collected historical task scheduling data is cleaned and preprocessed, and then used to train the CNN-based automatic feature generation model and DMNet model to improve the prediction accuracy and generalization ability of the model.
[0054] Embodiment 2
[0055] The specific implementation steps of a vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation are as follows:
[0056] Step 1: In the vehicle-mounted edge computing system, set up a data collection module to collect task scheduling related data from the server every 5 minutes, including task type (such as multimedia processing, data analysis, real-time control, etc.), computing requirements (expressed in the number of CPU cycles, such as 10000-50000 cycles), network bandwidth (in Mbps, such as 10-100Mbps), vehicle location (latitude and longitude coordinates), server resource status (CPU utilization, memory utilization, etc., expressed in percentage, such as 0-100%), etc.
[0057] Step 2: Clean the collected data and remove duplicate data and outliers. For example, data exceeding the mean ±3 times the standard deviation of the calculation requirement is considered as an outlier and removed, and missing values are filled with the mean.
[0058] Step 3: Preprocess the cleaned data to standardize it with a mean of 0 and a variance of 1.
[0059] Step 4: Construct an automatic feature generation model based on CNN. The convolution layer uses a 3×3 convolution kernel with a step size of 1 and 64 kernels. The pooling layer uses a 2×2 maximum pooling with a step size of 2. The reorganization layer performs random shuffling and recombination.
[0060] Step 5: Construct a multi-head VEC task offloading decision model (DMNet). Each head contains 256 neurons in the fully connected layer, the activation function is ReLU, the number of neurons in the output layer is determined according to the specific decision category, and the Softmax activation function is used for classification.
[0061] Step 6: Use the preprocessed historical data to train the model with 500 training rounds, a learning rate of 0.001, and a batch size of 128.
[0062] Step 7: In real-time task scheduling, task generation and resource status changes are monitored every second, and the computational effort of the task is obtained as 20,000 cycles, the deadline is 30 seconds, the priority is high, the vehicle location is (latitude: 30.5°, longitude: 120.5°), the network connection status is WLAN, the bandwidth is 50Mbps, and the computing power of the edge server is 80% CPU utilization and 20% memory utilization.
[0063] Step 8: Input the real-time data into the CNN-based automatic feature generation model to generate task features.
[0064] Step 9: Input the generated task features into the multi-head VEC task offloading decision model (DMNet).
[0065] Step 10: The DMNet model calculates that the decision result under WLAN access technology is to offload to the nearby WLAN access point A, the decision result under WAN access technology is not to offload, and the decision result under CN access technology is not to offload.
[0066] Step 11: Based on the decision results, the optimal decision is selected to offload to WLAN access point A, and the task offloading operation is performed.
[0067] Step 12: During the task execution, monitor the task execution status and resource usage every 5 seconds. If the network bandwidth drops to 20Mbps and the server CPU utilization rises to 95%, readjust the task scheduling strategy, such as migrating the task to other available access points or servers.
[0068] Embodiment 3
[0069] Due to the innovativeness of this technical solution, it can be widely used in application fields such as vehicle edge computing (VEC), intelligent transportation system, and Internet of Vehicles (IoV). With the development of intelligent connected vehicle technology, the demand for vehicle computing resources is increasing. This technical solution can effectively improve the decision-making efficiency of vehicle computing offloading and meet the needs of intelligent connected vehicles for efficient computing resources. At the same time, with the continuous advancement of Internet of Vehicles technology, communication and data exchange between vehicles have become more frequent. The application of this technical solution can optimize the computing offloading choices of vehicle users in different network environments and improve the response speed and quality of Internet of Vehicles services. In addition, as an important part of urban traffic management, the intelligent transportation system also has high requirements for the decision-making efficiency and accuracy of vehicle computing offloading. The application of this technical solution can improve the operating efficiency of the intelligent transportation system and promote the development of the intelligent transportation system. Therefore, this technical solution has broad market demand and good application prospects.
[0070] Embodiment 4
[0071] A reinforcement learning-based vehicle edge computing task scheduling strategy, which builds a multi-agent reinforcement learning environment and allows each vehicle user to act as an agent to learn the optimal unloading decision through interaction with the environment. The solution includes the following parts: state representation, action space definition, reward function design, and learning algorithm implementation.
[0072] 1) Status representation includes the vehicle’s computing power, network status, mission characteristics, etc.
[0073] 2) The action space is defined as the set of unloading decisions that vehicle users can choose.
[0074] 3) The reward function is designed to be defined based on factors such as the completion time, energy consumption, and network load of the offloading task.
[0075] 4) The learning algorithm is implemented using a multi-agent reinforcement learning algorithm, such as Multi-Agent Deep Deterministic Policy Gradient (MADDPG).
[0076] Embodiment 5
[0077] A vehicle-mounted edge computing task scheduling strategy based on federated learning. This strategy establishes a federated learning network between vehicle users, allowing vehicle users to share learning models and jointly optimize task scheduling strategies while protecting privacy. The solution includes the following parts: data preprocessing, model training, model aggregation, and model update.
[0078] 1) Data preprocessing includes cleaning and standardizing the historical task scheduling data of vehicle users.
[0079] 2) Model training is performed locally on each vehicle user, using local data to train a personalized scheduling model.
[0080] 3) Model aggregation: Through federated learning algorithms, such as the Federated Averaging (FedAvg) algorithm, the model parameters of each vehicle user are aggregated to form a global model.
[0081] 4) Model update: The aggregated global model parameters are distributed to each vehicle user for model update.
[0082] Implementation List VI
[0083] A genetic algorithm-based on-vehicle edge computing task scheduling strategy optimizes the task scheduling strategy by simulating natural selection and genetic mechanisms. The scheme includes the following parts: encoding, selection, crossover, and mutation.
[0084] 1) Encoding represents the task scheduling strategy as a chromosome.
[0085] 2) Select excellent chromosomes for reproduction based on the performance of task scheduling.
[0086] 3) Crossover: New chromosomes are generated through the crossover operation of chromosomes.
[0087] 4) Mutation introduces new genetic variation through chromosome mutation operations.
[0088] 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 principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation, characterized in that: include: CNN-based automatic feature generation model, multi-head VEC task offloading decision model (DMNet), historical task scheduling data processing module, including: 1) CNN-based automatic feature generation model: including convolution layer, pooling layer, and reorganization layer; Convolution layer: uses a 3×3 convolution kernel with a step size of 1 and 64 kernels; extracts the features of the input data through convolution operations; Pooling layer: uses 2×2 maximum pooling with a step size of 2 to reduce feature dimensions and reduce the amount of calculation; Reorganization layer: By randomly shuffling and recombining features, the influence of feature input order on model performance is avoided; 2) Multi-head VEC task offloading decision model: It includes multiple heads, each head corresponds to an access technology, and the access technologies include WLAN, WAN, and CN; Each head contains a fully connected layer and an output layer; The number of neurons in the fully connected layer is 256, and the activation function is ReLU; The number of neurons in the output layer is determined according to the specific decision category, and the Softmax activation function is used for classification; 3) Historical task scheduling data processing module: including data collection, cleaning, and preprocessing; Data collection: Task scheduling related data is collected from the server of the vehicle-mounted edge computing system every 5 minutes, including task type, computing requirements, network bandwidth, vehicle location, server resource status, etc. Data cleaning: remove duplicate data, outliers and missing values; the criterion for judging outliers is data exceeding the mean ± 3 times the standard deviation; Data preprocessing: Standardize the data to make its mean 0 and variance 1.
2. According to claim 1, a vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation is characterized in that: The operating principle of the CNN-based automatic feature generation model is as follows: the input original task scheduling data first undergoes feature extraction through the convolution layer to obtain a preliminary feature map; then undergoes dimensionality reduction processing through the pooling layer to reduce the number of features; and finally undergoes feature reorganization through the reorganization layer to eliminate the influence of the input order.
3. According to claim 1, a vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation is characterized in that: The operating principle of the multi-head VEC task offloading decision model is as follows: the features output by the CNN automatic feature generation model are input into each head of DMNet respectively; each head performs independent calculations and decisions based on the characteristics and requirements of the corresponding access technology, and finally the output results of multiple heads are combined to determine the optimal task offloading decision.
4. According to claim 1, a vehicle-mounted edge computing task scheduling model based on CNN automatic feature generation is characterized in that: Historical task scheduling data processing module: Operating principle: The collected historical task scheduling data is cleaned and preprocessed, and then used to train the CNN-based automatic feature generation model and DMNet model to improve the prediction accuracy and generalization ability of the model.
5. A method for scheduling vehicle-mounted edge computing tasks based on CNN automatic feature generation, characterized in that: The following steps are involved: Step 1: Real-time monitoring of task generation and resource status changes in the vehicle-mounted edge computing system; Step 2: Input the monitored real-time data into the CNN-based automatic feature generation model to generate task features; Step 3: Input the generated task features into the multi-head VEC task offloading decision model; Step 4: The DMNet model calculates the task offloading decision results under each access technology based on the characteristics of different access technologies and the current task characteristics, including whether to offload or not, the server selection for offloading, etc. Step 5: Integrate the decision results of multiple heads and select the optimal task offloading decision; Step 6: Execute the task offloading operation according to the optimal decision, and monitor the task execution status and resource usage in real time; if an abnormality occurs during the task execution, adjust the task scheduling strategy in time.
6. The method for scheduling vehicle-mounted edge computing tasks based on CNN automatic feature generation according to claim 5 is characterized in that: In step 1: the monitoring frequency is once per second, and information such as the computing amount, deadline, priority, etc. of the task, as well as resource status information such as the vehicle's location, speed, network connection status, computing power and bandwidth of the edge server are obtained.
7. The method for scheduling vehicle-mounted edge computing tasks based on CNN automatic feature generation according to claim 5 is characterized in that: In step 5: if the decision result of the WLAN head is to offload to a nearby WLAN access point, the decision result of the WAN head is to offload to a remote WAN server, and the decision result of the CN head is to offload to a cellular network base station, comprehensively consider factors such as network delay, bandwidth, server load, etc. and select the optimal decision.
8. The method for scheduling vehicle-mounted edge computing tasks based on CNN automatic feature generation according to claim 5 is characterized in that: In step 6: abnormalities occurring during task execution include: network interruption and server failure.
9. The method for scheduling vehicle-mounted edge computing tasks based on CNN automatic feature generation according to claim 5 is characterized in that: Before step 1, the following preparation steps are also included: Step 01: In the vehicle-mounted edge computing system, set up a data collection module to collect task scheduling related data from the server every 5 minutes, including task type, computing requirements, network bandwidth, vehicle location, and server resource status; Step 02: Clean the collected data to remove duplicate data and outliers; for example, data exceeding the mean ±3 times the standard deviation required for calculation is considered as outliers and removed, and missing values are filled with the mean; Step 03: Preprocess the cleaned data to standardize it to a mean of 0 and a variance of 1; Step 04: Build an automatic feature generation model based on CNN. The convolution layer uses a 3×3 convolution kernel with a step size of 1 and 64 kernels. The pooling layer uses a 2×2 maximum pooling with a step size of 2. The reorganization layer performs random shuffling and recombining. Step 05: Build a multi-head VEC task offloading decision model. Each head contains 256 neurons in the fully connected layer, the activation function is ReLU, the number of neurons in the output layer is determined according to the specific decision category, and the Softmax activation function is used for classification; Step 06: Use the preprocessed historical data to train the model with 500 training rounds, a learning rate of 0.001, and a batch size of 128.