AIoT edge computing cooperation method, equipment and medium

Abstract

The invention discloses an AIoT edge computing cooperation method and device and a medium, and the method comprises the steps: collecting multi-modal data through a sensor disposed at an Internet of Things terminal device, generating a to-be-processed task for the multi-modal data, and transmitting the to-be-processed task to an edge computing node; dynamic resource allocation is carried out according to the corresponding resource state through the edge computing node, a target model matched with the resource state is selected from a preset model library, an execution strategy corresponding to the to-be-processed task is generated through the target model, and the execution strategy is issued to the executor to enable the executor to execute corresponding decision operation; and aggregating the model parameters corresponding to the target models of the edge computing nodes, generating global optimization models corresponding to the target models according to the aggregated model parameters, and issuing the global optimization models to the edge computing nodes to realize global optimization of the model library.

Description

Technical Field

[0001] The present application relates to the field of Internet of Things technology, and specifically to an AIoT edge computing collaboration method, device, and medium. Background Art

[0002] AIoT, a cutting-edge technology that integrates AI and IoT, is accelerating global digital transformation. However, traditional IoT systems rely on centralized cloud-based data processing, which presents drawbacks such as high latency, high bandwidth usage, and significant privacy risks. To address the shortcomings of cloud-based solutions, edge computing has been introduced to optimize real-time performance. However, existing edge computing nodes lack intelligence and dynamic optimization capabilities, making it impossible to dynamically adjust resource allocation and algorithm models based on environmental changes, resulting in low resource utilization. Furthermore, the inefficiency of fusion of heterogeneous data from multiple sources further hinders real-time decision-making. Summary of the Invention

[0003] To solve the above problems, this application proposes an AIoT edge computing collaboration method, including: Collect multimodal data through sensors deployed on IoT terminal devices, generate tasks to be processed based on the multimodal data, and send the tasks to be processed to edge computing nodes; Dynamically allocate resources according to the corresponding resource status through the edge computing node, and select a target model that matches the resource status from a preset model library to generate an execution strategy corresponding to the task to be processed through the target model, and send the execution strategy to the executor to enable the executor to perform the corresponding decision operation; The model parameters corresponding to the target model of each edge computing node are aggregated, and based on the aggregated model parameters, a global optimization model corresponding to each target model is generated, and the global optimization model is sent to the edge computing node to achieve global optimization of the model library.

[0004] In one implementation of the present application, dynamic resource allocation is performed according to the corresponding resource status, and a target model matching the resource status is selected from a preset model library, specifically including: Determine the task priority corresponding to the task to be processed according to the task type corresponding to the task to be processed; Evaluating the resource status to determine whether computing resources are sufficient; If yes, construct an action space, a state space, and a reward function according to the resource state and the preset resource indicators, and allocate corresponding computing resources according to the action space, the state space, and the reward function through a Q-Learning algorithm; According to the resource status, a target model matching the resource status is selected from a preset model library; For the tasks to be processed, corresponding execution strategies are generated in sequence through the target model according to their corresponding task priorities.

[0005] In one implementation of the present application, an action space, a state space, and a reward function are constructed based on the resource state and the preset resource indicators, specifically including: The action space, state space, and reward function are constructed based on the resource state and preset resource indicators using the following formula:

[0006]

[0007]

[0008] Among them, S represents the action space, A represents the state space, Indicates status, Indicates CPU utilization. Indicates the remaining power. Indicates the length of the task queue, Indicates action, is the number of CPU cores, Indicates whether GPU is enabled. is the model compression ratio, Delay for algorithm update, To consume power, and Represents the reward weight.

[0009] In one implementation of the present application, the Q-Learning algorithm is used to allocate corresponding computing resources according to the action space, the state space, and the reward function, specifically including: In an offline state, the Q value is iteratively trained based on the historical data set to generate an initial Q value; wherein the initial Q value represents the historically optimal resource allocation strategy; The initial Q value is fine-tuned according to the action space, the state space, and the reward function through a Q-Learning algorithm, so as to allocate corresponding computing resources according to the fine-tuned initial Q value.

[0010] In one implementation of the present application, after determining whether computing resources are sufficient, the method further includes: Degrading resources of the pending task to obtain a resource status after degradation; A target model whose complexity matches the degraded resource state is screened out from the model library, and an execution strategy corresponding to the task to be processed is generated through the target model.

[0011] In one implementation of the present application, a global optimization model corresponding to each target model is generated based on the aggregated model parameters, specifically including: Adding differential privacy noise to the model gradient of the target model to generate a perturbed gradient; The perturbation gradients are aggregated, and the pre-trained large model is updated using the aggregated perturbation gradients to generate a global optimization model corresponding to each target model.

[0012] In one implementation of the present application, before selecting a target model that matches the resource state from a preset model library, the method further includes: The pre-built large model is migrated to the edge computing node, and the large model is compressed to generate a preset model library.

[0013] In one implementation of the present application, generating a task to be processed for the multimodal data specifically includes: Extracting feature vectors corresponding to the multimodal data through a preset neural network model; Assigning a corresponding attention weight to the feature vector according to the data source of the feature vector; According to the attention weights, feature fusion is performed on the feature vectors to generate a joint feature vector, and a task to be processed for processing the joint feature vector is generated.

[0014] The present application provides an AIoT edge computing collaborative device, comprising: at least one processor; and, a memory communicatively coupled to the at least one processor; In which, the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute an AIoT edge computing collaboration method as described in any one of the above items.

[0015] An embodiment of the present application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to: An AIoT edge computing collaboration method as described in any of the above items.

[0016] The AIoT edge computing collaboration method proposed in this application can bring the following beneficial effects: The dynamic resource allocation mechanism of edge computing nodes and the adaptive switching capabilities of the lightweight model library effectively compensate for the latency and privacy deficiencies of traditional cloud-based centralized processing. This also effectively improves the resource utilization of edge computing nodes and addresses the lack of dynamic optimization capabilities of existing edge computing nodes. The global model optimization mechanism enhances model generalization by integrating knowledge from multiple nodes while protecting edge data privacy. This improves the efficiency of multi-source heterogeneous data fusion, facilitates real-time decision-making, and enhances data processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic diagram of a process flow of an AIoT edge computing collaboration method provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of an AIoT edge computing collaborative device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0019] The following describes in detail the technical solutions provided by various embodiments of the present application in conjunction with the accompanying drawings.

[0020] like Figure 1 As shown, an AIoT edge computing collaboration method provided in an embodiment of the present application includes: S101: Collect multimodal data through sensors deployed on IoT terminal devices, generate tasks to be processed for the multimodal data, and send the tasks to be processed to edge computing nodes.

[0021] IoT devices are equipped with various types of sensors that collect multimodal data. This data can be integrated across heterogeneous devices through the adaptation of protocols such as MQTT, CoAP, and LoRaWAN. Multimodal data refers to data of varying types and sources. For example, in a smart home scenario, this data may include temperature data collected by a temperature sensor, image data captured by a camera, and sound data recorded by a microphone. Multimodal data encompasses a variety of data formats and can comprehensively reflect the various states and information of the environment in which the IoT device resides. Processing tasks are generated based on the collected multimodal data. These tasks are then sent to edge computing nodes. Based on the received tasks, the edge computing nodes invoke appropriate AI models to perform inference and generate execution strategies for the actuators to execute. Edge computing nodes are located close to the data source. Compared to transmitting all data to the cloud for processing, this approach can reduce data transmission latency, improve real-time data processing, reduce network bandwidth usage, and protect data privacy to a certain extent, as some data can be processed at the edge nodes without being uploaded to the cloud.

[0022] In one embodiment, the collected multimodal data needs to be fused before the processing task can be generated. First, the feature vectors corresponding to the multimodal data are extracted through a preset neural network model. Neural network models include CNN models and LSTM models. For image data, a lightweight CNN model such as MobileNetV3 can be used to extract image feature vectors, while for sensor time series data, the corresponding time series feature vectors can be extracted through the LSTM model.

[0023] Different data types play different roles in data fusion. Therefore, it is necessary to assign corresponding attention weights to feature vectors based on their data sources. For image feature vectors, their corresponding attention weights can be calculated using the following formula:

[0024] in, is the feature vector of the i-th sensor or modality, W is a trainable weight matrix that maps the input features to the latent space. b is a bias vector used to enhance the model’s expressiveness. q is a trainable query vector used to evaluate feature importance, and tanh is a hyperbolic tangent activation function that compresses the linear transformation result to In the range, exp is an exponential function that converts the score into a positive number to facilitate normalization of the weight.

[0025] The eigenvectors of other modalities can also be calculated in this way. After determining the attention weights corresponding to each modal data, multiple eigenvectors need to be integrated into a joint eigenvector according to the attention weights. The fusion method uses a weighted summation method, that is, each eigenvector is multiplied by its corresponding attention weight and then added. Assuming that the multimodal data includes image data and sensor time series data, it can be calculated using the following formula:

[0026] in, represents the joint eigenvector, represents the image feature vector, represents the time series feature vector.

[0027] The feature vector incorporates key information from multimodal data, providing a more comprehensive picture of the environmental status of IoT devices and providing a richer information foundation for subsequent task processing. This fused joint feature vector needs to be sent to edge computing nodes for real-time analysis to help determine whether the IoT device is experiencing anomalies or security threats. Therefore, it is necessary to generate pending tasks to process the joint feature vector.

[0028] S102: Dynamically allocate resources according to the corresponding resource status through the edge computing node, and select the target model that matches the resource status from the preset model library to generate the execution strategy corresponding to the task to be processed through the target model, and send the execution strategy to the executor so that the executor performs the corresponding decision operation.

[0029] When receiving a pending task, the edge computing node will dynamically allocate resources based on the resource status, so that it can clearly identify the resources that can be used when performing data processing tasks. Resources include the number of CPU cores, whether the GPU is enabled, and the model compression rate. The edge computing node will execute the pending task based on these allocated computing resources. The processing and analysis of the pending task needs to rely on the AI ​​model in the preset model library. By selecting the target model that matches the resource status of the current edge computing node, it can perform data reasoning on the pending task, thereby generating the corresponding execution strategy for the executor to make decision-making operations. After receiving the instruction issued by the edge computing node, the executor will automatically parse the execution strategy corresponding to the instruction and automatically trigger the corresponding physical operation or system-level operation. The executor can include servo motors, smart water valves, traffic light controllers and other physical devices in various application scenarios.

[0030] In one embodiment, the edge computing node analyzes the type of task to be processed and then prioritizes it based on the task type, allowing for optimal resource allocation and processing order. Task types are divided into urgent tasks and routine tasks based on latency requirements. Urgent tasks, with a latency requirement of no more than 100ms, have a higher priority, while routine tasks, with a latency requirement of more than 100ms, have a lower priority.

[0031] Edge computing nodes evaluate their resource status in real time, including CPU utilization, remaining memory, and storage space, to determine whether current computing resources are sufficient to successfully process pending tasks. If resources are sufficient, resource allocation is performed using the Q-Learning algorithm. This process first constructs an action space, a state space, and a reward function based on resource status and pre-set resource metrics. The action space defines possible resource allocation actions, the state space describes the current resource status, and the reward function evaluates the effectiveness of resource allocation.

[0032] Specifically, the action space is represented as , the state space is , the reward function is . Among them, S represents the action space, A represents the state space, Indicates status, Indicates CPU utilization. Indicates the remaining power. Indicates the length of the task queue, Indicates action, is the number of CPU cores, Indicates whether GPU is enabled. is the model compression ratio, Delay for algorithm update, To consume power, and Represents the reward weight.

[0033] Through training and learning, the Q-Learning algorithm can find the optimal resource allocation strategy and realize the allocation of computing resources.

[0034] Specifically, resource allocation for edge computing nodes is divided into two phases, offline and online, depending on their online status. The offline phase leverages historical data to quickly construct initial strategies, reducing online trial-and-error costs. The online phase periodically fine-tunes Q values ​​and updates strategies to accommodate real-time environmental changes.

[0035] First, in an offline state, the Q value is iteratively trained based on the historical data set to generate the initial Q value. The initial Q value represents the historically optimal resource allocation strategy, and its update rule is expressed as:

[0036] in, is the learning rate, which is 0.1 and is used to control the update step size. represents the discount factor, 、 、 and Represents state, action, immediate reward, and next state respectively.

[0037] Secondly, in the online state, the initial Q value needs to be fine-tuned according to the action space, state space and reward function through the Q-Learning algorithm. By fine-tuning the initial Q value, it is more consistent with the current actual situation and optimization goals, and then the corresponding computing resources are reasonably allocated according to the fine-tuned Q value to achieve efficient and reasonable use of computing resources and meet the needs of various tasks. In other words, real-time data is collected at fixed intervals. , update the Q value using the following formula:

[0038] Among them, the online learning rate It is usually smaller than the offline phase to prevent real-time data noise interference.

[0039] The online stage requires passing - Greedy search algorithm for adaptive adjustment, It can be set to 0.1. By setting a certain probability to try new actions, it can also select the currently known best action. This not only allows the strategy to have a certain degree of flexibility to adapt to dynamic changes in the environment, but also ensures that the existing better strategy can be used for resource allocation most of the time, thereby achieving effective optimization of dynamic resource allocation.

[0040] If resources are insufficient, it means that the remaining computing resources of the current edge computing node cannot support the execution of the pending task. In this case, the pending task needs to be downgraded. After resource downgrade, the edge computing node will receive an updated resource status. Based on the downgraded resource status, a target model with a complexity that matches it is selected from the preset model library. The target model can operate normally under the current downgraded resource conditions, ensuring the execution of the task. In the case of insufficient resources, downgrading the resource status and then accurately matching the model with the current actual resources can effectively utilize existing resources, prevent resource waste, and improve the resource utilization of the edge computing node, enabling it to efficiently and reasonably complete task processing even in resource-limited scenarios, thus ensuring the stability and availability of the edge computing system.

[0041] In addition, the embodiment of the present application also introduces a dynamic model switching mechanism, which can select a suitable target model for reasoning based on the current resource status, enhance the adaptability and flexibility of the edge computing node, and enable it to cope with various task requirements. Specifically, the edge computing node will select a matching target model from the preset model library based on the current resource status. For example, if the current resources are sufficient, a more complex model will be selected to obtain more accurate results; if resources are limited, a lightweight model will be selected to save resources.

[0042] After allocating available computing resources and selecting the appropriate target model, we need to generate execution policies for the tasks to be processed, based on the task priorities and the selected target model. The execution policy dictates how tasks should be processed and what actions should be taken. Tasks are processed in order of priority, ensuring that high-priority tasks are processed and executed first.

[0043] It should be noted that the model library is built based on the pre-trained large model sent by the cloud server. The cloud will initially train the large model based on the historical data set and Migrate to edge computing nodes to form a lightweight model , the edge side needs to compress the large model to make the model meet , The model size calculation function is used to form the AI ​​model library. The model library supports dynamic loading and can load target models of corresponding complexity in real time based on resource status.

[0044] S103: Aggregate the model parameters corresponding to the target model of each edge computing node, and generate a global optimization model corresponding to each target model based on the aggregated model parameters, and send the global optimization model to the edge computing node to achieve global optimization of the model library.

[0045] The cloud server aggregates the model parameters corresponding to the target models of each edge computing node and, based on these aggregated model parameters, generates a global optimization model for each target model. This global model is then distributed to each edge node, updating its local model library. This distributed collaborative training approach enables iterative model optimization without sharing original data, improving model performance while ensuring privacy and security.

[0046] In one embodiment, the cloud can achieve global optimization of the model through federated learning. First, differential privacy noise is added to the model gradient of the target model to generate a perturbed gradient. The noise masks the true value of the gradient, making it impossible for attackers to infer the original data through the gradient, thus ensuring data security. The perturbed gradient can be generated using the following formula: ,in, is from a value with a mean of 0 and a variance of The random noise sampled from the Gaussian distribution has a larger variance, which leads to stronger privacy protection. However, the accuracy of the model may also decrease. Therefore, the value of the variance needs to be balanced through experiments.

[0047] After generating the perturbation gradients, they are aggregated to aggregate global information. Furthermore, the pre-trained large model is updated using the aggregated perturbation gradients to generate a global optimization model for each target model.

[0048] The update formula of the global model is: ,in, is the learning rate, which controls the update step size of the global model, It is to aggregate global information by summing the perturbation gradients of N edge computing nodes.

[0049] While ensuring the data privacy of edge nodes, the global model integrates multi-node knowledge to enhance the generalization capability of the model, enabling it to adapt to diverse tasks in heterogeneous environments. It also reduces the resource overhead of edge deployment through lightweight adaptation, ultimately achieving safe, efficient, and accurate edge intelligent collaborative optimization.

[0050] In addition, an embodiment of the present application also provides an AIoT edge computing collaboration system, which includes a multimodal perception module, an edge computing node, a cloud collaboration module, and a cross-protocol communication interface.

[0051] The multimodal perception module is deployed on the IoT terminal device to collect heterogeneous data streams containing at least image data, sensor time series data and environmental sound data.

[0052] The edge computing node includes a dynamic resource allocation unit, a lightweight AI model library and a local decision engine. The dynamic resource allocation unit dynamically allocates computing resources and selects an adapted AI model for data processing tasks based on real-time task load and energy consumption constraints.

[0053] The cloud collaboration module aggregates the model parameters of multiple edge computing nodes through the federated learning algorithm, generates a global optimization model and sends it to the edge.

[0054] The cross-protocol communication interface supports data transmission of MQTT, CoAP and LoRaWAN protocols, enabling data interoperability between heterogeneous devices.

[0055] The system provided in this application solves real-time performance issues in resource-constrained scenarios through a dynamic, adaptive framework for "edge-cloud" collaboration. By integrating multimodal data with reinforcement learning, it achieves a balance between high-precision decision-making and low energy consumption. A compatible communication interface is designed to support seamless integration of devices from multiple vendors, making it suitable for scenarios such as industrial automation, smart cities, and smart homes.

[0056] The above are embodiments of the method proposed in this application. Based on the same idea, some embodiments of this application also provide devices and non-volatile computer storage media corresponding to the above methods.

[0057] Figure 2 This is a schematic diagram of the structure of an AIoT edge computing collaborative device provided in an embodiment of the present application. Figure 2 As shown, including: at least one processor; and, at least one processor communicatively connected to a memory; wherein, The memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to implement an AIoT edge computing collaborative method as described in any of the above items.

[0058] An embodiment of the present application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: An AIoT edge computing collaboration method as described in any of the above items.

[0059] The various embodiments in this application are described in a progressive manner. Similar portions between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device and medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For relevant portions, refer to the descriptions of the method embodiments.

[0060] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects to their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0061] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0063] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0065] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0066] Memory may include non-permanent storage in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0067] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can be implemented using any method or technology to store information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0068] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0069] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. An AIoT edge computing collaboration method, characterized in that: The method comprises: Collect multimodal data through sensors deployed on IoT terminal devices, generate tasks to be processed based on the multimodal data, and send the tasks to be processed to edge computing nodes; Dynamically allocate resources according to the corresponding resource status through the edge computing node, and select a target model that matches the resource status from a preset model library to generate an execution strategy corresponding to the task to be processed through the target model, and send the execution strategy to the executor to enable the executor to perform the corresponding decision operation; The model parameters corresponding to the target model of each edge computing node are aggregated, and based on the aggregated model parameters, a global optimization model corresponding to each target model is generated, and the global optimization model is sent to the edge computing node to achieve global optimization of the model library.

2. The AIoT edge computing collaborative method according to claim 1, characterized in that: Dynamic resource allocation is performed based on the corresponding resource status, and a target model that matches the resource status is selected from a preset model library, specifically including: Determine the task priority corresponding to the task to be processed according to the task type corresponding to the task to be processed; Evaluating the resource status to determine whether computing resources are sufficient; If yes, construct an action space, a state space, and a reward function according to the resource state and the preset resource indicators, and allocate corresponding computing resources according to the action space, the state space, and the reward function through a Q-Learning algorithm; According to the resource status, a target model matching the resource status is selected from a preset model library; For the tasks to be processed, corresponding execution strategies are generated in sequence through the target model according to their corresponding task priorities.

3. The AIoT edge computing collaborative method according to claim 2, characterized in that: According to the resource status and preset resource indicators, the action space, state space and reward function are constructed, specifically including: The action space, state space, and reward function are constructed based on the resource state and preset resource indicators using the following formula: Among them, S represents the action space, A represents the state space, Indicates status, Indicates CPU utilization. Indicates the remaining power. Indicates the length of the task queue, Indicates action, is the number of CPU cores, Indicates whether GPU is enabled. is the model compression ratio, Delay for algorithm update, To consume power, and Represents the reward weight.

4. The AIoT edge computing collaboration method according to claim 2, characterized in that: The root allocates corresponding computing resources according to the action space, the state space and the reward function through the Q-Learning algorithm, specifically including: In an offline state, the Q value is iteratively trained based on the historical data set to generate an initial Q value; wherein the initial Q value represents the historically optimal resource allocation strategy; The initial Q value is fine-tuned according to the action space, the state space, and the reward function through a Q-Learning algorithm, so as to allocate corresponding computing resources according to the fine-tuned initial Q value.

5. The AIoT edge computing collaboration method according to claim 2, characterized in that: After determining whether the computing resources are sufficient, the method further includes: Degrading resources of the pending task to obtain a resource status after degradation; A target model whose complexity matches the degraded resource state is screened out from the model library, and an execution strategy corresponding to the task to be processed is generated through the target model.

6. The AIoT edge computing collaboration method according to claim 1, characterized in that: Based on the aggregated model parameters, a global optimization model corresponding to each target model is generated, specifically including: Adding differential privacy noise to the model gradient of the target model to generate a perturbed gradient; The perturbation gradients are aggregated, and the pre-trained large model is updated using the aggregated perturbation gradients to generate a global optimization model corresponding to each target model.

7. The AIoT edge computing collaboration method according to claim 6, characterized in that: Before selecting a target model that matches the resource status from a preset model library, the method further includes: The pre-built large model is migrated to the edge computing node, and the large model is compressed to generate a preset model library.

8. The AIoT edge computing collaboration method according to claim 1, characterized in that: Generating tasks to be processed for the multimodal data, specifically including: Extracting feature vectors corresponding to the multimodal data through a preset neural network model; Assigning a corresponding attention weight to the feature vector according to the data source of the feature vector; According to the attention weights, feature fusion is performed on the feature vectors to generate a joint feature vector, and a task to be processed for processing the joint feature vector is generated.

9. An AIoT edge computing collaborative device, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; In which, the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute an AIoT edge computing collaboration method as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured to: An AIoT edge computing collaboration method as described in any one of claims 1 to 8.

Images