A computing power intelligent allocation method and service system based on cloud-edge collaboration

Abstract

The present invention relates to a computing power intelligent allocation method and service system based on cloud-edge collaboration, which constructs a historical intelligent driving task data set, inputs the created intelligent driving task prediction model for training, and predicts the intelligent driving task data to be processed; divides the predicted intelligent driving task data into different delay gradients, scores the task priority of each task data in the first delay gradient, the second delay gradient, and the third delay gradient, and pre-allocates computing power; collects the intelligent driving task data to be processed in real time, and performs differential calculation on the intelligent driving task data to be processed collected in real time and the predicted intelligent driving task data to be processed, obtains the differential result, and re-sorts based on the differential result; performs node priority scoring on each edge node and cloud center node, and adjusts the computing power pre-allocation so that each node processes the received task data. The above process realizes balanced coordination of computing power allocation based on cloud-edge.

Description

Technical Field

[0001] The present invention belongs to the technical field of computing power allocation, and in particular, relates to a computing power intelligent allocation method and service system based on cloud-edge collaboration. Background Art

[0002] Intelligent driving is a technology that uses computer, communication, sensor and other technologies to achieve automatic driving of cars. In the process of intelligent driving, the surrounding environment information is collected through on-board sensors (such as radar, lidar, camera, etc.) to identify roads, vehicles, pedestrians, traffic signs and other information. The data from different sensors are integrated and processed to improve the accuracy and robustness of perception. The precise location of the car on the map is determined by the global positioning system or other navigation technology, and the driving route is planned according to the predetermined destination. According to the results of perception, data processing and navigation, the acceleration, deceleration, steering and other actions of the car are controlled to achieve automatic driving. Through interaction with the driver and passengers, real-time road condition information, navigation information, etc. are provided, so that the driver can better understand the car status and environmental information. Safety and emergency handling: In the event of an emergency, risk avoidance measures are automatically taken to ensure driving safety.

[0003] Therefore, in the process of intelligent driving, there are many tasks or data that need to be processed by intelligent driving vehicles. In the process of processing these tasks or data, it is necessary to consider how to allocate computing resources, that is, how to allocate tasks or data to the corresponding computing center for processing. The principle of computing power allocation in the prior art is to offload tasks or data with high latency sensitivity to edge computers for processing, and tasks or data with low latency sensitivity to central cloud servers for processing. However, with the rapid development of intelligent driving, the amount of tasks or data in the intelligent driving process has increased dramatically. If only latency sensitivity is used as a reference standard for computing power allocation, it is obviously unable to meet the existing intelligent driving needs.

[0004] Therefore, how to improve the existing computing power allocation method to achieve balanced and coordinated computing power allocation is a technical problem that needs to be solved urgently. Summary of the invention

[0005] The purpose of the present invention is to provide a computing power intelligent allocation method and service system based on cloud-edge collaboration, so as to improve the existing computing power allocation method and achieve balanced coordination of computing power allocation in multiple dimensions.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0007] In the first aspect, a computing power intelligent allocation method based on cloud-edge collaboration is provided, including the following specific processes:

[0008] S1: Acquire intelligent driving task data within a specified period, construct a historical intelligent driving task data set, and create an intelligent driving task prediction model, input the historical intelligent driving task data set into the intelligent driving task prediction model to train it, and predict the intelligent driving task data to be processed at time t through the intelligent driving task prediction model, and output the predicted intelligent driving task data;

[0009] S2: Divide the predicted intelligent driving task data into different delay gradients based on the minimum delay requirement, including the first delay gradient, the second delay gradient, and the third delay gradient; the task data in the first delay gradient is processed by the edge node, and the task data in the third delay gradient is processed by the cloud center node, and the processed data is returned to the intelligent driving terminal device via the edge node; the task data in the second delay gradient is processed by the edge node or the cloud center node;

[0010] S3: scoring the task priority of each task data in the first delay gradient, the second delay gradient and the third delay gradient respectively, sorting them according to the priority score value, and pre-allocating computing power according to the sorting result; wherein the indicators of the task priority score include delay, transmission time cost and transmission fee cost;

[0011] S4: collecting the intelligent driving task data to be processed at time t in real time, performing differential calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed at time t predicted to obtain a differential result, and reordering according to step S3 based on the differential result;

[0012] S5: Perform node priority scoring on each edge node and cloud center node. The node priority scoring indicators include computing power and remaining task data. The computing power pre-allocation is adjusted based on the node priority score and the task priority score to obtain a real-time computing power allocation result. Each node processes the received task data according to the real-time computing power allocation result.

[0013] S6: Incorporate the real-time collected intelligent driving task data to be processed at time t into the historical intelligent driving task data set in step S1, and iteratively execute steps S1 to S5.

[0014] Preferably, the intelligent driving task prediction model in step S1 is a long short-term memory network model, and an output layer, multiple LSTM layers and an output layer are set, and the number of neurons and hyperparameters of each layer are defined; the process of training the intelligent driving task prediction model is as follows:

[0015] Inputting the historical intelligent driving task data set into the intelligent driving task prediction model for forward propagation, updating the cell state and hidden state of the LSTM layer and calculating the output of the intelligent driving task prediction model;

[0016] Comparing the output of the intelligent driving task prediction model with the true label, and calculating the loss function value to obtain the difference between the output of the intelligent driving task prediction model and the true value;

[0017] Back propagation is performed to calculate the gradient according to the loss function value, and the weights and biases of the intelligent driving task prediction model are updated based on the conjugate gradient method to fit the training data;

[0018] Iterate the above three steps until the preset number of iterations is reached or the loss function value is less than the set threshold;

[0019] The intelligent driving task data of another specified period is obtained and input into the intelligent driving task prediction model, the model is verified, and it is determined whether the verification result meets the preset requirements. If so, the next step is executed; if not, retraining is performed.

[0020] Preferably, the time extension length of the first delay gradient in step S2 is the shortest, the time extension length of the third delay gradient is the longest, and the time extension length of the second delay gradient is longer than the time extension length of the first delay gradient and less than the time extension length of the third delay gradient.

[0021] Preferably, the specific process of step S3 is as follows:

[0022] S31: setting corresponding weights for the delay, transmission time cost, and transmission fee cost in the indicators of the task priority scoring, and obtaining a delay weight, a transmission time cost weight, and a transmission fee cost weight respectively;

[0023] S32: Based on the task priority scoring indicator, each task data is scored for task priority, and the scoring formula is as follows:

[0024] Priority score value = delay weight * α + transmission time cost weight * β, transmission fee cost weight * γ.

[0025] Among them, α is the delay weight parameter, the value range is [0-1], which indicates the influence of delay on priority; β is the transmission time cost weight parameter, the value range is [0-1], which indicates the influence of transmission time cost on priority; γ is the transmission cost weight parameter, the value range is [0-1], which indicates the influence of transmission cost on priority;

[0026] S33: pre-allocating task data in the first delay gradient to edge nodes according to priority scores, and pre-allocating task data with higher priority scores to edge nodes with stronger computing capabilities;

[0027] Pre-allocate task data with a priority score greater than a specified threshold in the second delay gradient to edge nodes, and pre-allocate task data with a higher priority score to edge nodes with stronger computing power;

[0028] Pre-allocate task data with a priority score value less than or equal to a specified threshold in the second delay gradient to cloud center nodes, and pre-allocate task data with a higher priority score value to cloud center nodes with stronger computing power;

[0029] The task data in the third delay gradient are pre-allocated to the cloud center nodes according to the priority score values; and the task data with higher priority score values ​​are pre-allocated to the cloud center nodes with stronger computing power.

[0030] Preferably, the specific process of performing differential calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed at the predicted time t in step S4 is as follows:

[0031] S41: Construct a differential data set, and define an original data set M and a differential data set N, where N[ i ] is the number of data elements from the first to the i The difference between the data elements, that is, N[ i ]= M[ i ]- M [ i -1];

[0032] S42: By passing N[ l ]Add data c ,N[ r +1] Subtract data c , to achieve the specified interval [ l , r ]To add a data c modification of

[0033] S43: Restore the original data group M by calculating the prefix of the differential data group N, where the prefix is ​​the sum of all items starting from the first item of the differential array N and accumulated to the current item, and this sum is equal to the data element at the corresponding position in the original data group M.

[0034] Preferably, the specific process of performing node priority scoring on each edge node and cloud center node in S5 is as follows:

[0035] S51: assigning corresponding weights to the node priority scoring indicators to obtain computing power weights and remaining task data weights;

[0036] S52: Calculate the node priority score based on the computing power, computing power weight, remaining task data, and remaining task data weight. The calculation formula is as follows:

[0037] Node priority score = computing power * computing power weight - remaining task data * remaining task data weight;

[0038] The computing power = the maximum number of clock cycles executed by the CPU per second * the number of floating-point operations performed per clock cycle * the number of CPUs.

[0039] Preferably, the adjustment process of pre-allocating computing power based on the node priority score and the task priority score in step S5 is: allocating task data with higher task priority score values ​​to nodes with higher node priority score values.

[0040] In a second aspect, a computing power intelligent allocation service system based on cloud-edge collaboration is used to implement any one of the computing power intelligent allocation methods based on cloud-edge collaboration, including a data acquisition module, a model creation module, an intelligent driving task prediction model, a priority calculation module, a data acquisition module, a differential calculation module, a computing power allocation module and a main control module, wherein the data acquisition module, the model creation module, the intelligent driving task prediction model, the priority calculation module, the data acquisition module, the differential calculation module, and the computing power allocation module are all connected to the main control module;

[0041] The data acquisition module is used to acquire intelligent driving task data within a specified period and construct a historical intelligent driving task data set;

[0042] The model creation module is used to create an intelligent driving task prediction model, and input the historical intelligent driving task data set into the intelligent driving task prediction model for training;

[0043] The intelligent driving task prediction model is used to predict the intelligent driving task data to be processed at time t and output the predicted intelligent driving task data;

[0044] The priority calculation module is used to calculate the task priority score value for each task data in the first delay gradient, the second delay gradient and the third delay gradient, and to calculate the node priority score value for each edge node and the cloud center node based on the computing power and the remaining task data;

[0045] The data acquisition module is used to collect the intelligent driving task data to be processed at time t in real time;

[0046] The difference calculation module is used to perform a difference calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed at the predicted time t to obtain a difference result;

[0047] The computing power allocation module is used to pre-allocate the task data in the first delay gradient to the edge node, pre-allocate the task data in the third delay gradient to the cloud center node, pre-allocate the task data in the second delay gradient to the edge node or the cloud center node, and adjust the pre-allocation based on the node priority score value;

[0048] The main control module is used to control the data acquisition module, model creation module, intelligent driving task prediction model, priority calculation module, data acquisition module, differential calculation module, and computing power allocation module.

[0049] The beneficial effects of the present invention include:

[0050] The cloud-edge collaboration-based intelligent computing power allocation method and service system provided by the present invention construct a historical intelligent driving task data set, and input the created intelligent driving task prediction model for training, and predict the intelligent driving task data to be processed; divide the predicted intelligent driving task data into different delay gradients, and perform task priority scoring and pre-allocate computing power for each task data in the first delay gradient, the second delay gradient and the third delay gradient; collect the intelligent driving task data to be processed in real time, and perform differential calculation on the intelligent driving task data to be processed collected in real time and the predicted intelligent driving task data to be processed to obtain a differential result, and re-sort based on the differential result; perform node priority scoring on each edge node and cloud center node, and adjust the computing power pre-allocation so that each node processes the received task data.

[0051] First, by building an intelligent driving task prediction model and training it based on historical intelligent driving task data sets, we can accurately predict the intelligent driving task data to be processed, and pre-allocate computing power based on the prediction results of the intelligent driving task data to be processed. Through the pre-allocation of computing power, task data can be directly allocated after being collected by the terminal device, effectively reducing data allocation time and data processing latency.

[0052] Second, by dividing the predicted intelligent driving task data into different delay gradients according to its delay requirements, namely the first delay gradient, the second delay gradient and the third delay gradient, these three delay gradients represent the different delay requirements of the intelligent driving task data to be processed. For example, some data has high delay sensitivity and needs to be processed in a very short time, so it is divided into the first delay gradient. Some data has low delay sensitivity and can be completed in a longer time, so it is divided into the third delay gradient. Some data is between the delay requirements of the first delay gradient and the third delay gradient, so it is divided into the second delay gradient. The tasks in the second delay gradient can be processed flexibly, and real-time cloud-edge computing power can be allocated according to the specific needs of the task and the current status of the edge-cloud node, thereby achieving balanced coordination of cloud-edge computing power.

[0053] Third, by scoring the task priority of each task data in the first delay gradient, the second delay gradient, and the third delay gradient, the task priority is scored again in each gradient. That is, the tasks are allocated and processed in order based on the task priority in each gradient, and then combined with the priority scoring of each edge node and cloud center node, the tasks of each gradient can achieve more refined computing power allocation and processing.

[0054] Fourth, differential calculation is performed on the real-time collected intelligent driving task data to be processed and the predicted intelligent driving task data to be processed to find the difference between the predicted task to be processed and the real-time data to be processed, and pre-allocation adjustment is performed based on the difference results, which greatly saves the computing power allocation time for real-time data to be processed, reduces the task data processing delay, and effectively improves the task data processing efficiency.

[0055] Fifth, the real-time collected intelligent driving task data to be processed is incorporated into the historical intelligent driving task data set, and the entire process is iteratively executed, so that the prediction of the task data to be processed at a specified time is made by the intelligent driving task prediction model based on the intelligent driving task data added at any time, ensuring the accuracy of each task data prediction, and then ensuring the accuracy of subsequent computing power allocation, and realizing balanced and coordinated computing power allocation based on cloud edge. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the method for intelligent allocation of computing power based on cloud-edge collaboration of the present invention.

[0057] Figure 2 This is a schematic diagram of the cloud-edge architecture of the present invention.

[0058] Figure 3 This is a schematic diagram of the architecture of the computing power intelligent allocation service system based on cloud-edge collaboration of the present invention. DETAILED DESCRIPTION

[0059] The following is combined with Figure 1~Figure 3 The present invention is further described in detail:

[0060] Example 1

[0061] See attached Figure 1 As shown in the figure, a computing power intelligent allocation method based on cloud-edge collaboration includes the following specific processes:

[0062] S1: Obtain the intelligent driving task data within the specified period, construct a historical intelligent driving task data set, and create an intelligent driving task prediction model. Input the historical intelligent driving task data set into the intelligent driving task prediction model to train it, and use the intelligent driving task prediction model to predict the intelligent driving task data to be processed at time t, and output the predicted intelligent driving task data. By constructing an intelligent driving task prediction model and training it based on the historical intelligent driving task data set, accurate prediction of the intelligent driving task data to be processed can be achieved, and pre-allocation of computing power can be achieved based on the prediction results of the intelligent driving task data to be processed. Through the pre-allocation of computing power, it can be directly allocated after the terminal device collects the task data, effectively reducing the data allocation time and reducing the data processing delay.

[0063] S2: Divide the predicted intelligent driving task data into different delay gradients based on the minimum delay requirement, including the first delay gradient, the second delay gradient, and the third delay gradient; the task data in the first delay gradient is processed by the edge node, and the task data in the third delay gradient is processed by the cloud center node, and the processed data is returned to the intelligent driving terminal device via the edge node; the task data in the second delay gradient is processed by the edge node or the cloud center node.

[0064] The above process divides the predicted intelligent driving task data into different delay gradients according to its delay requirements, namely the first delay gradient, the second delay gradient and the third delay gradient. These three delay gradients represent the different delay requirements of the intelligent driving task data to be processed. For example, some data has high delay sensitivity and needs to be processed in a very short time, so it is divided into the first delay gradient. Some data has low delay sensitivity and can be completed in a longer time, so it is divided into the third delay gradient. Some data is between the delay requirements of the first delay gradient and the third delay gradient, so it is divided into the second delay gradient. The tasks in the second delay gradient can be processed flexibly, and real-time cloud-edge computing power can be allocated according to the specific needs of the task and the current status of the edge-cloud node, thereby achieving balanced coordination of cloud-edge computing power.

[0065] S3: The task priority scores are respectively performed for each task data in the first delay gradient, the second delay gradient, and the third delay gradient, and the task data are sorted according to the priority score value, and the computing power is pre-allocated according to the sorting result; the computing power is pre-allocated based on the prediction results of the intelligent driving task data to be processed. Through the pre-allocation of computing power, the task data can be directly allocated after being collected by the terminal device, which effectively reduces the data allocation time and reduces the data processing delay. The indicators of task priority scoring include delay, transmission time cost, and transmission fee cost.

[0066] S4: Collect the intelligent driving task data to be processed at time t in real time, and perform differential calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed predicted at time t to obtain the differential result, and reorder according to step S3 based on the differential result. Perform differential calculation on the intelligent driving task data to be processed collected in real time and the predicted intelligent driving task data to be processed to find the difference between the predicted task to be processed and the real-time data to be processed, and perform pre-allocation adjustment based on the difference result, which greatly saves the computing power allocation time for the real-time data to be processed, reduces the task data processing delay, and effectively improves the task data processing efficiency.

[0067] S5: Node priority scoring is performed on each edge node and cloud center node. The node priority scoring indicators include computing power and remaining task data. The pre-allocation of computing power is adjusted based on the node priority score and the task priority score to obtain the real-time computing power allocation result. Each node processes the received task data according to the real-time computing power allocation result. By scoring the task priority of each task data in the first delay gradient, the second delay gradient, and the third delay gradient respectively, and scoring the task priority again in each gradient, that is, the task priority in each gradient is allocated and processed in order, and then combined with the priority score of each edge node and cloud center node, the tasks of each gradient can achieve more refined computing power allocation and processing.

[0068] S6: Incorporate the real-time collected intelligent driving task data to be processed at time t into the historical intelligent driving task data set in step S1, and iteratively execute steps S1 to S5. Incorporate the real-time collected intelligent driving task data to be processed into the historical intelligent driving task data set, and iteratively execute the entire process, so that the prediction of the task data to be processed at a specified time is performed by the intelligent driving task prediction model based on the newly added intelligent driving task data at any time, ensuring the accuracy of each task data prediction, and then ensuring the accuracy of subsequent computing power allocation, and realizing balanced coordination of computing power allocation based on cloud edge.

[0069] Example 2

[0070] On the basis of Example 1, the intelligent driving task prediction model in step S1 is a long short-term memory network model, and an output layer, multiple LSTM layers and an output layer are set, and the number of neurons and hyperparameters of each layer are defined. Multiple LSTM layers include hidden layers, forget gates, and input gates. Each time step in the hidden layer receives an input and the cell state of the previous time step. Then, through a series of gating mechanisms and cell state updates, the cell state and hidden state of the current time step are calculated and output. The gating mechanism controls the flow of information, determines which information is forgotten, which new information is added, and which information is output.

[0071] The process of training the intelligent driving task prediction model is to input the historical intelligent driving task data set into the intelligent driving task prediction model for forward propagation, update the cell state and hidden state of the LSTM layer, and calculate the output of the intelligent driving task prediction model. The output of the intelligent driving task prediction model is compared with the true label, and the loss function value is calculated to obtain the difference between the output of the intelligent driving task prediction model and the true value. Back propagation is performed according to the loss function value to calculate the gradient, and the weights and biases of the intelligent driving task prediction model are updated based on the conjugate gradient method to fit the training data. The above three steps are iterated until the preset number of iterations is reached or the loss function value is less than the set threshold. The intelligent driving task data of another specified period is obtained and input into the intelligent driving task prediction model, the model is verified, and it is determined whether the verification result meets the preset requirements. If so, the next step is executed, and if not, retraining is performed.

[0072] By introducing gating mechanisms and cell states, the intelligent driving task prediction model can maintain information over a longer time span, extract features of long-term dependencies between data and time, and selectively forget specified information in the cell state, enabling the model to adapt well to the temporal dynamics of intelligent driving task data.

[0073] The time extension length of the first delay gradient in step S2 is the shortest, the time extension length of the third delay gradient is the longest, and the time extension length of the second delay gradient is longer than the time extension length of the first delay gradient and shorter than the time extension length of the third delay gradient.

[0074] Example 3

[0075] Based on Example 1 or Example 2, the specific process of step S3 is as follows:

[0076] S31: setting corresponding weights for the delay, transmission time cost and transmission fee cost in the indicators of task priority scoring, and obtaining the delay weight, transmission time cost weight and transmission fee cost weight respectively. Since the delay, transmission time cost and transmission fee cost have a certain influence on the priority of task processing, but the degree of their influence is different, different weights are set for them.

[0077] S32: Based on the task priority scoring indicators, each task data is scored for task priority. The scoring formula is as follows:

[0078] Priority score value = delay weight * α + transmission time cost weight * β, transmission cost weight * γ;

[0079] Among them, α is the delay weight parameter, the value range is [0-1], which indicates the influence of delay on priority; β is the transmission time cost weight parameter, the value range is [0-1], which indicates the influence of transmission time cost on priority; γ is the transmission cost weight parameter, the value range is [0-1], which indicates the influence of transmission cost on priority.

[0080] S33: Pre-allocate the task data in the first delay gradient to the edge nodes according to the priority score values, and pre-allocate the task data with higher priority scores to the edge nodes with stronger computing capabilities. Pre-allocate the task data in the second delay gradient whose priority score values ​​are greater than the specified threshold to the edge nodes, and pre-allocate the task data with higher priority scores to the edge nodes with stronger computing capabilities. Pre-allocate the task data in the second delay gradient whose priority score values ​​are less than or equal to the specified threshold to the cloud center nodes, and pre-allocate the task data with higher priority scores to the cloud center nodes with stronger computing capabilities. Pre-allocate the task data in the third delay gradient to the cloud center nodes according to the priority score values; and pre-allocate the task data with higher priority scores to the cloud center nodes with stronger computing capabilities.

[0081] By scoring the task priority of each task data in the first delay gradient, the second delay gradient, and the third delay gradient respectively, and scoring the task priority again in each gradient, that is, sorting the tasks based on the task priority in each gradient, and then combining the subsequent priority scoring of each edge node and cloud center node, the tasks of each gradient can achieve more refined computing power allocation and processing.

[0082] Example 4

[0083] On the basis of Embodiment 1 or Embodiment 2 or Embodiment 3, the specific process of performing differential calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed at the predicted time t in step S4 is as follows:

[0084] S41: Construct a differential data set, and define an original data set M and a differential data set N, where N[ i ] is the number of data elements from the first to the i The difference between the data elements, that is, N[ i ]= M[ i ]- M [ i -1].

[0085] S42: By passing N[ l ]Add data c ,N[ r +1] Subtract data c , to achieve the specified interval [ l , r ]To add a data c Since modifying the data elements in the differential data group N will affect the corresponding position and all subsequent elements in the metadata group M, by changing N[ l ]Add data c ,N[ r +1] Subtract data c , to achieve the specified interval [ l , r ]To add a data c Modifications efficiently update data values ​​in a specific range of metadata groups.

[0086] S43: The original data group M is restored by calculating the prefix of the differential data group N. The prefix is ​​the sum of all items starting from the first item of the differential array N and accumulated to the current item. This sum is equal to the data element at the corresponding position in the original data group M.

[0087] The core of the above differential calculation process is to use the differential data group N to perform efficient prefix sum calculation and interval modification process, which realizes a large amount of data processing, allowing us to complete the differential calculation in a shorter time complexity.

[0088] The specific process of performing node priority scoring on each edge node and cloud center node in step S5 is as follows:

[0089] S51: assign corresponding weights to the node priority scoring indicators to obtain computing power weights and remaining task data weights;

[0090] S52: Calculate the node priority score based on the computing power, computing power weight, remaining task data, and remaining task data weight. The calculation formula is as follows:

[0091] Node priority score = computing power * computing power weight - remaining task data * remaining task data weight;

[0092] The computing power = the maximum number of clock cycles executed by the CPU per second * the number of floating-point operations performed per clock cycle * the number of CPUs.

[0093] The adjustment process of computing power pre-allocation based on node priority score and task priority score is as follows: task data with higher task priority score values ​​are allocated to nodes with higher node priority score values.

[0094] See attached Figure 2 and attached Figure 3 As shown, a computing power intelligent allocation service system based on cloud-edge collaboration is used to implement any one of the computing power intelligent allocation methods based on cloud-edge collaboration, including a data acquisition module, a model creation module, an intelligent driving task prediction model, a priority calculation module, a data acquisition module, a differential calculation module, a computing power allocation module and a main control module. The data acquisition module, the model creation module, the intelligent driving task prediction model, the priority calculation module, the data acquisition module, the differential calculation module, and the computing power allocation module are all connected to the main control module.

[0095] The data acquisition module is used to acquire the intelligent driving task data within the specified period and construct a historical intelligent driving task data set. The model creation module is used to create an intelligent driving task prediction model and input the historical intelligent driving task data set into the intelligent driving task prediction model for training. The intelligent driving task prediction model is used to predict the intelligent driving task data to be processed at time t and output the predicted intelligent driving task data. The priority calculation module is used to calculate the task priority score value for each task data in the first delay gradient, the second delay gradient and the third delay gradient, and to calculate the node priority score value for each edge node and the cloud center node based on the computing power and the remaining task data. The data collection module is used to collect the intelligent driving task data to be processed at time t in real time. The differential calculation module is used to perform differential calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed predicted at time t to obtain the differential result. The computing power allocation module is used to pre-allocate the task data in the first delay gradient to the edge node, pre-allocate the task data in the third delay gradient to the cloud center node, pre-allocate the task data in the second delay gradient to the edge node or the cloud center node, and adjust the pre-allocation based on the node priority score. The main control module is used to control the data acquisition module, the model creation module, the intelligent driving task prediction model, the priority calculation module, the data acquisition module, the differential calculation module, and the computing power allocation module.

[0096] In summary, the intelligent allocation method and service system of computing power based on cloud-edge collaboration provided by the present invention realizes accurate prediction of the intelligent driving task data to be processed by constructing an intelligent driving task prediction model and training based on the historical intelligent driving task data set, and realizes pre-allocation of computing power based on the prediction results of the intelligent driving task data to be processed. Through the pre-allocation of computing power, it can be directly allocated after the task data is collected by the terminal device, effectively reducing the data allocation time and reducing the data processing delay. By dividing the predicted intelligent driving task data into different delay gradients according to its delay requirements, namely the first delay gradient, the second delay gradient and the third delay gradient, these three delay gradients represent the different delay requirements of the intelligent driving task data to be processed. For example, some data has a high delay sensitivity and needs to be processed in a very short time, so it is divided into the first delay gradient. Some data has a low delay sensitivity and can be completed in a longer time, so it is divided into the third delay gradient. Some data is between the delay requirements of the first delay gradient and the third delay gradient, and is divided into the second delay gradient. The tasks in the second delay gradient can be processed flexibly, and real-time cloud-edge computing power can be allocated according to the specific requirements of the task and the current status of the edge-cloud nodes, thereby achieving balanced coordination of cloud-edge computing power.

[0097] By scoring the task priority of each task data in the first delay gradient, the second delay gradient and the third delay gradient respectively, and scoring the task priority again in each gradient, that is, the task priority in each gradient is used to allocate and process the tasks in order, and then combined with the priority scoring of each edge node and cloud center node, the tasks of each gradient can achieve more refined computing power allocation and processing. The real-time collected intelligent driving task data to be processed and the predicted intelligent driving task data to be processed are differentially calculated to find the difference between the predicted tasks to be processed and the real-time data to be processed, and the pre-allocation adjustment is made according to the difference results, which greatly saves the computing power allocation time for the real-time data to be processed, reduces the task data processing delay, and effectively improves the task data processing efficiency. The real-time collected intelligent driving task data to be processed is incorporated into the historical intelligent driving task data set, and the whole process is iteratively executed, so that the prediction of the task data to be processed at a specified time is performed by the intelligent driving task prediction model based on the newly added intelligent driving task data at any time, ensuring the accuracy of each task data prediction, and then ensuring the accuracy of the subsequent computing power allocation, and realizing the balanced coordination of computing power allocation based on the cloud edge.

Claims

1. A method for intelligent allocation of computing power based on cloud-edge collaboration, characterized in that: The specific process includes the following: S1: Acquire intelligent driving task data within a specified period, construct a historical intelligent driving task data set, and create an intelligent driving task prediction model, input the historical intelligent driving task data set into the intelligent driving task prediction model to train it, and predict the intelligent driving task data to be processed at time t through the intelligent driving task prediction model, and output the predicted intelligent driving task data; S2: Divide the predicted intelligent driving task data into different delay gradients based on the minimum delay requirement, including the first delay gradient, the second delay gradient, and the third delay gradient; the task data in the first delay gradient is processed by the edge node, and the task data in the third delay gradient is processed by the cloud center node, and the processed data is returned to the intelligent driving terminal device via the edge node; the task data in the second delay gradient is processed by the edge node or the cloud center node; S3: scoring the task priority of each task data in the first delay gradient, the second delay gradient and the third delay gradient respectively, sorting them according to the priority score value, and pre-allocating computing power according to the sorting result; wherein the indicators of the task priority score include delay, transmission time cost and transmission fee cost; S4: collecting the intelligent driving task data to be processed at time t in real time, performing differential calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed at time t predicted to obtain a differential result, and reordering according to step S3 based on the differential result; S5: Perform node priority scoring on each edge node and cloud center node. The node priority scoring indicators include computing power and remaining task data. The computing power pre-allocation is adjusted based on the node priority score and the task priority score to obtain a real-time computing power allocation result. Each node processes the received task data according to the real-time computing power allocation result. S6: Incorporate the real-time collected intelligent driving task data to be processed at time t into the historical intelligent driving task data set in step S1, and iteratively execute steps S1 to S5; The specific process of step S3 is as follows: S31: setting corresponding weights for the delay, transmission time cost, and transmission fee cost in the indicators of the task priority scoring, and obtaining a delay weight, a transmission time cost weight, and a transmission fee cost weight respectively; S32: Based on the task priority scoring indicator, each task data is scored for task priority, and the scoring formula is as follows: Priority score value = delay weight * α + transmission time cost weight * β, transmission cost weight * γ; Among them, α is the delay weight parameter, the value range is [0-1], which indicates the influence of delay on priority; β is the transmission time cost weight parameter, the value range is [0-1], which indicates the influence of transmission time cost on priority; γ is the transmission cost weight parameter, the value range is [0-1], which indicates the influence of transmission cost on priority; S33: pre-allocating task data in the first delay gradient to edge nodes according to priority scores, and pre-allocating task data with higher priority scores to edge nodes with stronger computing capabilities; Pre-allocate task data with a priority score greater than a specified threshold in the second delay gradient to edge nodes, and pre-allocate task data with a higher priority score to edge nodes with stronger computing power; Pre-allocate task data with a priority score value less than or equal to a specified threshold in the second delay gradient to cloud center nodes, and pre-allocate task data with a higher priority score value to cloud center nodes with stronger computing power; Pre-allocate the task data in the third delay gradient to the cloud center nodes according to the priority score values; and pre-allocate higher priority score values ​​to cloud center nodes with stronger computing power; The specific process of node priority scoring for each edge node and cloud center node in S5 is as follows: S51: assigning corresponding weights to the node priority scoring indicators to obtain computing power weights and remaining task data weights; S52: Calculate the node priority score based on the computing power, computing power weight, remaining task data, and remaining task data weight. The calculation formula is as follows: Node priority score = computing power * computing power weight - remaining task data * remaining task data weight; The computing power = the maximum number of CPU clock cycles per second * the number of floating point operations per clock cycle * the number of CPUs; The adjustment process of pre-allocating computing power based on the node priority score and the task priority score in step S5 is: allocating task data with higher task priority score values ​​to nodes with higher node priority score values.

2. According to claim 1, a computing power intelligent allocation method based on cloud-edge collaboration is characterized in that: The intelligent driving task prediction model in step S1 is a long short-term memory network model, and sets an output layer, multiple LSTM layers and an output layer, and defines the number of neurons and hyperparameters of each layer; the process of training the intelligent driving task prediction model is as follows: Inputting the historical intelligent driving task data set into the intelligent driving task prediction model for forward propagation, updating the cell state and hidden state of the LSTM layer and calculating the output of the intelligent driving task prediction model; Comparing the output of the intelligent driving task prediction model with the true label, and calculating the loss function value to obtain the difference between the output of the intelligent driving task prediction model and the true value; Back propagation is performed to calculate the gradient according to the loss function value, and the weights and biases of the intelligent driving task prediction model are updated based on the conjugate gradient method to fit the training data; Iterate the above three steps until the preset number of iterations is reached or the loss function value is less than the set threshold; The intelligent driving task data of another specified period is obtained and input into the intelligent driving task prediction model, the model is verified, and it is determined whether the verification result meets the preset requirements. If so, the next step is executed; if not, retraining is performed.

3. According to the method of intelligent computing power allocation based on cloud-edge collaboration in claim 1, it is characterized in that: The time extension length of the first delay gradient in step S2 is the shortest, the time extension length of the third delay gradient is the longest, and the time extension length of the second delay gradient is longer than the time extension length of the first delay gradient and shorter than the time extension length of the third delay gradient.

4. According to claim 1, a computing power intelligent allocation method based on cloud-edge collaboration is characterized in that: The specific process of performing differential calculation on the real-time collected intelligent driving task data to be processed and the predicted intelligent driving task data to be processed at time t in step S4 is as follows: S41: Construct a differential data set, and define an original data set M and a differential data set N, where N[ i ] is the number of data elements from the first to the i The difference between the data elements, that is, N[ i ]= M[ i ]- M[ i -1]; S42: By passing N[ l ]Add data c ,N[ r +1] Subtract data c , to achieve the specified interval [ l , r ]To add a data c modification of S43: Restore the original data group M by calculating the prefix of the differential data group N, where the prefix is ​​the sum of all items starting from the first item of the differential array N and accumulated to the current item, and this sum is equal to the data element at the corresponding position in the original data group M.

5. A computing power intelligent allocation service system based on cloud-edge collaboration, used to implement a computing power intelligent allocation method based on cloud-edge collaboration as described in any one of claims 1-4, characterized in that: It includes a data acquisition module, a model creation module, an intelligent driving task prediction model, a priority calculation module, a data acquisition module, a differential calculation module, a computing power allocation module and a main control module, wherein the data acquisition module, the model creation module, the intelligent driving task prediction model, the priority calculation module, the data acquisition module, the differential calculation module and the computing power allocation module are all connected to the main control module; The data acquisition module is used to acquire intelligent driving task data within a specified period and construct a historical intelligent driving task data set; The model creation module is used to create an intelligent driving task prediction model, and input the historical intelligent driving task data set into the intelligent driving task prediction model for training; The intelligent driving task prediction model is used to predict the intelligent driving task data to be processed at time t and output the predicted intelligent driving task data; The priority calculation module is used to calculate the task priority score value for each task data in the first delay gradient, the second delay gradient and the third delay gradient, and to calculate the node priority score value for each edge node and the cloud center node based on the computing power and the remaining task data; The data acquisition module is used to collect the intelligent driving task data to be processed at time t in real time; The difference calculation module is used to perform a difference calculation on the intelligent driving task data to be processed collected in real time and the intelligent driving task data to be processed at the predicted time t to obtain a difference result; The computing power allocation module is used to pre-allocate the task data in the first delay gradient to the edge node, pre-allocate the task data in the third delay gradient to the cloud center node, pre-allocate the task data in the second delay gradient to the edge node or the cloud center node, and adjust the pre-allocation based on the node priority score value; The main control module is used to control the data acquisition module, model creation module, intelligent driving task prediction model, priority calculation module, data acquisition module, differential calculation module, and computing power allocation module.

Images