Heterogeneous computing power fusion method and device based on 5G CPE and WiFi

Abstract

The invention mainly relates to the technical field of wireless communication, and provides a heterogeneous computing power fusion method and device based on 5G CPE and WiFi in order to solve the problem that a 5G CPE and WiFi hybrid network lacks a reasonable computing power scheduling mechanism, and the core idea of the heterogeneous computing power fusion method and device based on 5G CPE and WiFi is that a plurality of node devices are arranged in the 5G CPE and WiFi hybrid network; respectively setting a large language model on the server and each node device to judge whether a to-be-processed task input into the node device needs scheduling computing power or not, respectively evaluating residual computing power of the server and the node device if the scheduling computing power is needed, and distributing task processing amount to the node device and the server according to the proportion of the residual computing power; the task transmission network is selected according to the signal states of the two networks to transmit the to-be-processed task to the server, the large language model on the server processes the received to-be-processed task and then returns the processing result to the node device for integration, and the task processing efficiency and response speed are improved.

Description

Technical Field

[0001] The present invention mainly relates to the field of wireless communication technology, and in particular to a method and device for integrating heterogeneous computing power based on 5G CPE and WiFi. Background Art

[0002] In modern home environments, with the popularity of smart devices and the increase in networking needs, the management of home networks and computing resources has become increasingly complex. 5G CPE and WiFi networks have become two common network connection methods in homes, each with different coverage characteristics and performance advantages. 5G CPE has higher bandwidth and lower latency, while WiFi performs better in terms of stability and coverage. As users move around the home, network latency and stable connection issues become important factors affecting user experience; in addition, the computing platforms in the home are heterogeneous, and the compatibility between devices is low, whether in chip architecture or processing unit. Edge devices usually have weak computing power and cannot independently and efficiently perform complex model reasoning. The lack of a unified resource scheduling and computing power integration mechanism means that edge devices cannot fully utilize the high computing power of fixed devices such as home servers when processing tasks with high computing requirements. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for integrating heterogeneous computing power based on 5G CPE and WiFi, so as to solve the problem that the 5G CPE and WiFi hybrid network lacks a reasonable computing power scheduling mechanism, resulting in low efficiency in processing tasks with high computing power requirements.

[0004] The technical solution adopted by the present invention to solve the above technical problems is: A heterogeneous computing power fusion method based on 5G CPE and WiFi is applied to a communication system of a 5G CPE and WiFi hybrid network including a server, and the method includes: Set up multiple node devices in the 5G CPE and WiFi hybrid network, and set up large language models on the server and each node device respectively; The large language model on the node device determines whether the pending tasks of the input node device need to schedule the computing power of the server. If necessary, the remaining computing power of the server and the node device is evaluated respectively, and the pending tasks are allocated to the node device and the server according to the proportion of the remaining computing power. The pending tasks assigned to the server are converted into a format that the server can recognize. The task transmission network is selected according to the signal status of the 5G CPE and the WiFi network. The pending tasks assigned to the server are transmitted to the server based on the selected task transmission network. The large language models on the server and the node devices process the assigned pending tasks respectively. The server returns the task processing results to the node device. The node device integrates the task processing results returned by the server with its own task processing results and transmits them to the user.

[0005] Furthermore, the specific method for the large language model on the node device to determine whether the pending task of the input node device requires the computing power of the scheduling server is: the large language model on the node device determines whether the computing power of the scheduling server is required based on the processing delay requirement of the pending task of the input node device. If the processing delay requirement of the pending task is greater than or equal to the set value, the large language model on the node device uses the local computing power of the node device to process the pending task, and does not schedule the computing power of the server; if the task processing delay requirement of the pending task is less than the set value, the remaining computing power of the server and the node device is evaluated respectively, and the pending task is allocated to the node device and the server according to the proportion of the remaining computing power.

[0006] Furthermore, the selecting of a task transmission network according to the signal status of the 5G CPE and the WiFi network specifically includes: calculating the root mean square and standard deviation of the 5G CPE and WiFi signal strengths in a set area around the node device, evaluating the average signal strengths of the 5G CPE and WiFi based on the root mean square, and evaluating the signal stability of the 5G CPE and WiFi based on the standard deviation; if the average signal strength and signal stability of the 5G CPE are better than those of WiFi, selecting the 5G CPE to transmit the pending tasks assigned to the server to the server; otherwise, selecting WiFi to transmit the pending tasks assigned to the server to the server.

[0007] Furthermore, the method for evaluating the remaining computing power of the server specifically includes: assigning set weights to the server's floating-point operations per second, memory bandwidth, throughput and CPU / GPU frequency, performing min-max normalization on the server's floating-point operations per second, memory bandwidth, throughput and CPU / GPU frequency, and performing weighted calculation on the standardized values ​​according to the assigned weights, and using the result of the weighted calculation as the server's remaining computing power.

[0008] Furthermore, the remaining computing power evaluation method of the node device specifically includes: assigning set weights to the power consumption, delay, number of floating-point operations per second and bandwidth of the node device, performing min-max normalization on the power consumption, delay, number of floating-point operations per second and bandwidth of the node device, and performing weighted calculation on the standardized values ​​according to the assigned weights, and using the result of the weighted calculation as the remaining computing power of the node device.

[0009] Furthermore, before the large language model on the node device determines whether the pending task of the input node device needs to schedule computing power, the large language model on the node device preprocesses the pending task, and the preprocessing specifically includes: Standardize the original pixel values ​​of the image data, and convert the standardized image data into vector data through image encoding; After the voice data is sampled and quantized, normalization and feature extraction are performed to convert the voice data into vector data; The text data is segmented and stop words are removed, and the text data is converted into vector data through word embedding.

[0010] Furthermore, the large language model includes multiple layers of a feedforward neural network consisting of a multi-head attention module and a multi-layer perceptron module.

[0011] Furthermore, before the large language model is respectively set on the node device and the server, the method further includes: training the large language model, performing distillation and pruning processing on the trained large language model, and then setting it on the node device and the server.

[0012] Furthermore, when the large language model processes the task to be processed, if the privacy requirement of the task to be processed is higher than the set standard, federated learning is selected to process the task to be processed; otherwise, distributed learning is selected to process the task to be processed.

[0013] On the other hand, the present invention also provides a heterogeneous computing power fusion device based on 5G CPE and WiFi, the device comprising 5G CPE, a WiFi component, a plurality of node devices, a data conversion module and a server; The 5G CPE and WiFi components are used to establish a 5G CPE and WiFi hybrid network. A large language model is set on the node device and the server. Multiple node devices are respectively connected to the 5G CPE and WiFi hybrid network. The large language model on the node device determines whether the pending task of the input node device needs to schedule the server computing power. If necessary, the remaining computing power of the server and the node device is evaluated respectively, and the amount of pending tasks is allocated to the node device and the server according to the proportion of the remaining computing power. The data conversion module is used to convert the pending tasks assigned to the server into a format that can be stored and parsed by the server, select a task transmission network according to the signal status of the 5G CPE and WiFi, and transmit the pending tasks assigned to the server to the server based on the selected task transmission network; The large language models on the server and node devices process the assigned tasks to be processed respectively. The server returns the task processing results to the node device. The large language model on the node device integrates the task processing results returned by the server with its own task processing results and transmits them to the user.

[0014] The beneficial effects of the present invention are: (1) The present invention arranges multiple node devices in a 5G CPE and WiFi hybrid network, and sets a large language model on each node device and server respectively. The large language model on the node device determines whether the pending task input to the node device needs to schedule the computing power of the node device and the server. If necessary, the remaining computing power of the server and the node device is evaluated respectively, and the pending task is allocated to the node device and the server in proportion. A task transmission signal network is selected according to the signal status of the 5G CPE and WiFi network. Based on the selected task transmission signal network, the pending task allocated to the server is transmitted to the server. The large language models on the server and the node device process the allocated pending task respectively. The server returns the task processing result to the node device. The node device integrates the task processing result returned by the server with its own task processing result and transmits it to the user. By efficiently integrating the heterogeneous computing power of the node device and the server, the node device can dynamically schedule the computing power assistance of the home server according to the timeliness requirements of the reasoning task, thereby improving the efficiency and response speed of task processing.

[0015] (2) The present invention selects a task transmission signal network according to the signal status of the 5G CPE and the WiFi network, and transmits the pending tasks assigned to the server to the server based on the selected task transmission signal network. With the help of the 5G CPE and WiFi hybrid network, the optimal network is intelligently selected according to the network signal status of the user's location, which can effectively ensure the timeliness of the server and node devices in performing the pending tasks and the data transmission success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the heterogeneous computing power fusion device of the 5G CPE and WiFi hybrid network described in the present invention. DETAILED DESCRIPTION

[0017] The heterogeneous computing power fusion method of the 5G CPE and WiFi hybrid network described in the present invention is based on the 5G CPE and WiFi hybrid network, and schedules the external computing power such as home servers and the local computing power of node devices to improve the processing efficiency of tasks. The heterogeneous computing power fusion method of the 5G CPE and WiFi hybrid network described in the present invention includes the following steps: Step 1: Deploy the large language model on node devices and servers.

[0018] First, multiple node devices are deployed in the 5G CPE and WiFi hybrid network to build a communication network architecture with wide coverage and stable connection to ensure smooth data communication.

[0019] A large language model is set up in the server and each node device. The internal architecture of the large language model consists of multiple layers, each of which includes a multi-head attention (MHA) module and a feedforward neural network composed of a multi-layer perceptron (MLP). In order to reduce the storage requirements of the large language model, the established large language model is trained before being deployed to the node device and server. The trained large language model is first lightweighted by pruning and distillation. Pruning technology compresses the large language model by reducing the redundant parameters of the large language model, making the large language model suitable for deployment on node devices with limited data storage. After the large language model is deployed on the node device, in order to restore the performance of the large language model, the large language model before pruning is used as the teacher network, and the pruned large language model is used as the student network for distillation. The distillation process aims to transfer knowledge through the teacher network to restore the performance of the pruned student network, ensuring that only minimal performance loss is introduced while reducing the amount of data in the large language model, so that the final lightweight large language model can run effectively on the node device and server.

[0020] Step 2: The large language model on the node device preprocesses the task to be processed input by the user.

[0021] Before the large language model makes computing power scheduling decisions for different types of tasks, in order to ensure that the input data meets the computing requirements of the large language model, it is necessary to pre-process the tasks to be processed input by the user and convert them into a unified vector form, thereby ensuring that the large language model can effectively process different types of data and perform task reasoning.

[0022] In this embodiment, the preprocessing of text data by the large language model includes word segmentation, removal of stop words and other text normalization operations, and then the text is converted into a numerical vector representation through word embedding technology. The preprocessing of image data includes resizing and normalization operations, and after the original pixel values ​​are normalized, the image is converted into a vector representation through image coding technology. The preprocessing process of speech data includes signal sampling and quantization, and then the speech signal is converted into a numerical vector representation through normalization operations and feature extraction.

[0023] Step 3: The large language model on the node device determines whether the pre-processed task needs to schedule server computing power.

[0024] The large language model on the node device determines whether computing power scheduling is required based on the processing delay requirements of the pending tasks input to the node device. If the processing delay requirements of the pending tasks are greater than or equal to the set value, the large language model on the node device uses the local computing power of the node device to process the pending tasks, and does not schedule the computing power of the server. For example, if the types of tasks such as offline data analysis and batch image processing have low latency requirements, the large language model will give priority to using the local computing power of the node device. If the task processing delay requirements of the pending tasks are less than the set value, the remaining computing power of the server and node device will be evaluated respectively, and the pending tasks will be allocated to the node device and server according to the proportion of the remaining computing power, and the server computing power and the local computing power of the node device will be scheduled. For example, if tasks such as video stream processing and speech recognition have high latency requirements, the large language model will apply for scheduling server computing power to ensure the real-time performance of the tasks.

[0025] Step 4: For tasks that require scheduling server computing power, convert the pending tasks into a format that the server can recognize, select the task transmission network based on the signal status of the 5G CPE and WiFi network, and transmit the pending tasks assigned to the server to the server based on the selected task transmission network. The large language models on the server and node devices process the assigned pending tasks respectively.

[0026] The present invention evaluates the remaining computing power of the server based on the number of floating-point operations per second, memory bandwidth, throughput and CPU / GPU frequency of the server, assigns device weights to the number of floating-point operations per second, memory bandwidth, throughput and CPU / GPU frequency of the server respectively, performs min-max normalization on the number of floating-point operations per second, memory bandwidth, throughput and CPU / GPU frequency values, performs weighted calculation on the standardized values ​​according to the assigned weights, and uses the weighted calculation result as the remaining computing power of the server. The remaining computing power of the node device is evaluated based on the power consumption, delay, number of floating-point operations per second and bandwidth of the node device, assigns set weights to the power consumption, delay, number of floating-point operations per second and bandwidth of the node device respectively, performs min-max normalization on the power consumption, delay, number of floating-point operations per second and bandwidth, and performs weighted calculation on the standardized values ​​according to the assigned weights, and uses the weighted calculation result as the remaining computing power of the node device.

[0027] In this embodiment, the weights assigned to the server's floating-point operations per second, memory bandwidth, and throughput are all 0.3, and the weights assigned to the CPU / GPU frequencies are 0.1 respectively; the weights assigned to the node device's power consumption, latency, and floating-point operations per second are all 0.3, and the weight assigned to the bandwidth is 0.1.

[0028] The root mean square and standard deviation of the 5G CPE and WiFi signal strength within the set range of the area where the node device is located are calculated. The average signal strength of the 5G CPE and WiFi signals is evaluated based on the root mean square of the signal strength within the set range. The stability of the 5G CPE and WiFi signals is evaluated based on the standard deviation of the signal strength within the set range. When the average signal strength and stability indicators of the 5G CPE are better than those of WiFi, the 5G CPE is selected to transmit the pending tasks to the server; otherwise, WiFi is selected.

[0029] Step 5: After the large language model on the server and node devices processes the task to be processed, the server returns the task processing result to the node device for integration, generates a complete task processing result on the node device, and returns it to the task processor.

[0030] When the large language model processes tasks, the large language models on the server and node devices respectively process the assigned tasks to be processed and select distributed learning and federated learning to process the tasks according to the privacy requirements of the tasks to be processed. If the privacy requirements of the tasks to be processed are higher than the set standard, the federated learning method is selected to process the tasks to be processed; otherwise, the distributed learning method is selected to process the tasks to be processed.

[0031] When the server returns the task processing results to the node device, in order to ensure that the node device can smoothly integrate all the task processing results, it is also necessary to convert the task processing results into a set format and return them to the node device.

[0032] The heterogeneous computing power fusion device based on 5G CPE and WiFi described in the present invention is as follows Figure 1 As shown, it includes 5G CPE, WiFi components, node equipment, data conversion modules and servers; The node device is provided with a large language model, and the node device is respectively connected to the 5G CPE and the WiFi network. The large language model on the node device determines whether the pending task input to the node device requires the computing power of the scheduling server. If necessary, the remaining computing power of the server and the node device is evaluated respectively, and the amount of pending tasks is allocated to the node device and the server in proportion. The data conversion module is used to convert the pending tasks assigned to the server into a mode that can be stored and parsed by the server, select a task transmission network according to the signal status of the 5G CPE and WiFi, and transmit the pending tasks assigned to the server to the server based on the selected task transmission network; The large language models on the server and node devices process the assigned tasks to be processed respectively. The server returns the task processing results to the node device. The node device integrates the task processing results returned by the server with its own task processing results and transmits them to the user.

Claims

1. A heterogeneous computing power fusion method based on 5G CPE and WiFi is applied to a communication system of a 5G CPE and WiFi hybrid network including a server, characterized in that: The method comprises: Set up multiple node devices in the 5G CPE and WiFi hybrid network, and set up large language models on the server and each node device respectively; The large language model on the node device determines whether the pending tasks of the input node device need to schedule the computing power of the server. If necessary, the remaining computing power of the server and the node device is evaluated respectively, and the pending tasks are allocated to the node device and the server according to the proportion of the remaining computing power. The pending tasks assigned to the server are converted into a format that the server can recognize. The task transmission network is selected according to the signal status of the 5G CPE and the WiFi network. The pending tasks assigned to the server are transmitted to the server based on the selected task transmission network. The large language models on the server and the node devices process the assigned pending tasks respectively. The server returns the task processing results to the node device. The node device integrates the task processing results returned by the server with its own task processing results and transmits them to the user.

2. The heterogeneous computing power fusion method based on 5G CPE and WiFi according to claim 1, characterized in that: The specific method for the large language model on the node device to judge whether the pending task of the input node device needs to schedule the computing power of the server is as follows: the large language model on the node device judges whether the computing power of the server needs to be scheduled according to the processing delay requirement of the pending task of the input node device; if the processing delay requirement of the pending task is greater than or equal to the set value, the large language model on the node device uses the local computing power of the node device to process the pending task, and does not schedule the computing power of the server; if the task processing delay requirement of the pending task is less than the set value, the remaining computing power of the server and the node device is evaluated respectively, and the pending task is allocated to the node device and the server according to the proportion of the remaining computing power.

3. The heterogeneous computing power fusion method based on 5G CPE and WiFi according to claim 1, characterized in that: The method of selecting a task transmission network according to the signal status of the 5G CPE and the WiFi network specifically includes: calculating the root mean square and standard deviation of the 5G CPE and WiFi signal strengths in a set area around the node device, evaluating the average signal strengths of the 5G CPE and WiFi based on the root mean square, and evaluating the signal stability of the 5G CPE and WiFi based on the standard deviation; if the average signal strength and signal stability of the 5G CPE are better than those of WiFi, then selecting the 5G CPE to transmit the pending tasks assigned to the server to the server; otherwise, selecting WiFi to transmit the pending tasks assigned to the server to the server.

4. The heterogeneous computing power fusion method based on 5G CPE and WiFi according to claim 1, characterized in that: The method for evaluating the remaining computing power of a server specifically includes: assigning set weights to the server's floating-point operations per second, memory bandwidth, throughput, and CPU / GPU frequency, performing min-max normalization on the server's floating-point operations per second, memory bandwidth, throughput, and CPU / GPU frequency, and performing weighted calculation on the standardized values ​​according to the assigned weights, and using the result of the weighted calculation as the remaining computing power of the server.

5. The heterogeneous computing power fusion method based on 5G CPE and WiFi according to claim 1, characterized in that: The remaining computing power evaluation method of the node device specifically includes: assigning set weights to the power consumption, delay, number of floating-point operations per second and bandwidth of the node device, performing min-max normalization on the power consumption, delay, number of floating-point operations per second and bandwidth of the node device, and performing weighted calculation on the standardized values ​​according to the assigned weights, and using the result of the weighted calculation as the remaining computing power of the node device.

6. The heterogeneous computing power integration method based on 5G CPE and WiFi according to any one of claims 1 to 5, characterized in that: Before the large language model on the node device determines whether the pending task of the input node device needs to schedule computing power, the large language model on the node device preprocesses the pending task, and the preprocessing specifically includes: Standardize the original pixel values ​​of the image data, and convert the standardized image data into vector data through image encoding; After the voice data is sampled and quantized, normalization and feature extraction are performed to convert the voice data into vector data; The text data is segmented and stop words are removed, and the text data is converted into vector data through word embedding.

7. The heterogeneous computing power integration method based on 5G CPE and WiFi according to any one of claims 1 to 5, characterized in that: The large language model includes multiple layers of a feedforward neural network consisting of a multi-head attention module and a multi-layer perceptron module.

8. The heterogeneous computing power integration method based on 5G CPE and WiFi according to any one of claims 1 to 5, characterized in that: Before respectively setting the large language model on the node device and the server, the method also includes: training the large language model, performing distillation and pruning processing on the trained large language model, and then setting it on the node device and the server.

9. The method for integrating heterogeneous computing power based on 5G CPE and WiFi according to any one of claims 1 to 5, characterized in that: When a large language model processes a task to be processed, if the privacy requirement of the task to be processed is higher than the set standard, federated learning is selected to process the task to be processed; otherwise, distributed learning is selected to process the task to be processed.

10. A heterogeneous computing power fusion device based on 5G CPE and WiFi, used to implement the heterogeneous computing power fusion method based on 5G CPE and WiFi as described in any one of claims 1 to 9, characterized in that: The device includes a 5G CPE, a WiFi component, a plurality of node devices, a data conversion module and a server; The 5G CPE and WiFi components are used to establish a 5G CPE and WiFi hybrid network. A large language model is set on the node device and the server. Multiple node devices are respectively connected to the 5G CPE and WiFi hybrid network. The large language model on the node device determines whether the pending task of the input node device needs to schedule the server computing power. If necessary, the remaining computing power of the server and the node device is evaluated respectively, and the amount of pending tasks is allocated to the node device and the server according to the proportion of the remaining computing power. The data conversion module is used to convert the pending tasks assigned to the server into a format that can be stored and parsed by the server, select a task transmission network according to the signal status of the 5G CPE and WiFi, and transmit the pending tasks assigned to the server to the server based on the selected task transmission network; The large language models on the server and node devices process the assigned tasks to be processed respectively. The server returns the task processing results to the node device. The large language model on the node device integrates the task processing results returned by the server with its own task processing results and transmits them to the user.

Images