Information processing method and system based on Internet of Things

Abstract

The invention discloses an information processing method and system based on the Internet of Things, and relates to the technical field of Internet of Things services, and the method comprises the steps: dynamically obtaining the computing power, energy consumption and network state of each node through an environment perception module, and constructing a node capability vector and a link performance vector; the task analysis module is used for carrying out structured modeling on tasks and identifying candidate segmentation points and unloading targets; the cost modeling module establishes a joint cost function of time delay and energy consumption; the strategy generation module generates an optimal unloading strategy based on multi-agent reinforcement learning; and the scheduling communication module executes task unloading and result integration according to the strategy, and feeds back execution data to update the system state. According to the method, intelligent, low-delay and energy-consumption optimization processing of task segmentation unloading is realized, and the method is suitable for a dynamic heterogeneous Internet of Things computing environment.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) service technology, and in particular to an information processing method and system based on IoT. Background Technology

[0002] With the rapid development of IoT technology, an increasing number of smart devices are being deployed in edge environments to perform various data acquisition, processing, and analysis tasks. However, IoT devices are often limited by computing power, battery life, and network resources, making it difficult for them to independently complete complex computing tasks. Especially under multi-tasking, high-concurrency, and dynamic network conditions, the system is prone to problems such as prolonged processing time, excessive energy consumption, and unbalanced resource allocation. Therefore, achieving efficient task scheduling under a collaborative architecture of edge, cloud, and endpoint has become an important research topic in the field of IoT computing architecture.

[0003] Existing offloading scheduling methods mostly employ static strategies or simple heuristic algorithms for task partitioning and target node selection, lacking in-depth modeling of device computing power, communication status, and task structure, making it difficult to adapt to dynamically changing physical environments and task loads. For example, some methods make decisions only on a task-by-task basis, ignoring the segmented execution characteristic within a task, resulting in coarse-grained offloading and poor flexibility. Other methods, while considering network and energy consumption factors, have significant shortcomings in multi-node collaboration, state updates, and long-term optimization, making it difficult to achieve efficient utilization of global resources and optimal task completion efficiency.

[0004] To address the aforementioned problems, this invention provides an information processing method and system based on the Internet of Things (IoT). By establishing a structured task model, environmental state modeling, and joint cost function, combined with a multi-agent reinforcement learning strategy, intelligent task segmentation and optimal offloading decisions are achieved. This method not only supports dynamic perception and feedback mechanisms but also boasts advantages such as strong scalability, high resource utilization efficiency, and strong adaptability to complex environments, significantly improving task processing performance and overall system efficiency in IoT environments. Summary of the Invention

[0005] To address the above problems, this invention provides an information processing method and system based on the Internet of Things, which solves the problems of coarse task unloading granularity and lack of global optimization and adaptive capabilities in execution decision-making in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An information processing method based on the Internet of Things includes the following steps: Step S1: The environment perception module acquires dynamic and static data from each computing node, performs static data modeling and dynamic data capability estimation respectively, obtains node capability vector and link performance vector, and encapsulates them into an environment state vector for real-time release. Step S2: The task parsing module performs structured modeling and load quantification assessment of the task to be processed, identifies suitable offloading boundaries, and matches feasible offloading destinations in combination with the current system environment status, and finally generates candidate segmentation point-target node mapping and its load parameter set. Step S3: Based on the environmental state vector obtained in S1 and the candidate segmentation point-target node mapping and its load parameter set formed in S2, the execution delay model and energy consumption model are constructed respectively through the task cost modeling module, and a unified delay-energy consumption joint cost function is constructed through normalization fusion. Step S4: Use reinforcement learning algorithm to generate optimal task segmentation and unloading decision strategy. The strategy generation module adopts multi-agent deep reinforcement learning method to model the unloading decision problem of multiple IoT devices as a Markov decision process. Step S5 involves unloading the scheduling execution and result integration feedback, which is uniformly controlled by the scheduling communication module. Based on the strategy decision output in step S4, the task segmentation and scheduling execution operations are completed, and the execution results are sent back to the environment state model.

[0007] An Internet of Things (IoT)-based information processing system, applied to the aforementioned IoT-based information processing method, includes: The environmental perception module is responsible for real-time monitoring and collection of key operating parameters in the IoT system, and for building and maintaining global environmental status information of the system. The task parsing module is responsible for the structural modeling and load assessment of the tasks to be processed. It constructs a structured representation of the task based on the task type, parses the computation and communication overhead of each subtask, selects reasonable candidate split points, and maps the feasible set of unloading target nodes in combination with the current environment state. The cost modeling module is responsible for constructing a joint cost model of delay and energy consumption for the task under different partitioning and offloading schemes, and outputting the comprehensive cost value corresponding to each scheme. The strategy generation module generates task unloading decision strategies based on a multi-agent reinforcement learning algorithm trained in a centralized manner. According to the current system state input and task characteristics, it calls the local agent model to output the optimal task split point and corresponding unloading target. The scheduling and communication module is responsible for scheduling and executing the offloading scheme output by the strategy generation module in the actual environment, while controlling the target server to receive and process backend tasks, and managing data interaction and resource scheduling between devices.

[0008] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves fine-grained analysis and unloadable boundary identification of computational tasks by constructing a structured task model and a segmentation analysis mechanism. The task analysis module can extract the computational load and data dependencies of each stage according to the task type, and combined with the communication load modeling results, select reasonable segmentation points and unloading destinations, which significantly improves the flexibility and adaptability of task scheduling, enabling tasks to achieve optimal collaborative execution between local and remote ends under the premise of meeting resource constraints.

[0009] This invention introduces a delay-energy consumption joint cost model, which unifies the measurement of task execution time overhead and energy consumption, forming a multi-index comprehensive evaluation function. The cost modeling module can accurately evaluate the performance of different offloading schemes based on device computing power, network bandwidth, propagation latency, and task structure parameters, providing a clear quantitative basis for strategy generation. This joint cost function is more in line with the actual needs of IoT scenarios, effectively balancing system performance and energy efficiency, and enhancing the economy and sustainability of system operation.

[0010] This invention employs a multi-agent reinforcement learning algorithm with centralized training and distributed execution to achieve intelligent optimization of offloading strategies in multi-device collaborative scenarios. Each IoT device agent can autonomously decide on the split point and offloading node based on its local state, while the strategy generation module learns the global environment rules through centralized training, effectively solving the problems of multi-task competition and resource conflicts.

[0011] This invention supports a complete offloading execution and feedback closed-loop mechanism. The scheduling and communication module not only realizes automatic task segmentation, local execution and remote transmission, but also dynamically monitors the execution results and provides feedback, realizing online correction of model parameters and adaptive updating of strategies. This ensures that the system still has stable optimization capabilities when facing uncertainties such as network fluctuations and resource changes, and significantly improves scheduling robustness and long-term system performance. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but is merely a selection of embodiments of the present invention.

[0015] Please refer to Figure 1 , Figure 1 A flowchart of an information processing method based on the Internet of Things provided in this embodiment of the invention includes the following steps: Step S1: The environment perception module acquires dynamic and static data from each computing node, performs dynamic data capability estimation and static data modeling respectively, generates node capability vectors and link performance vectors, and encapsulates them into environment state vectors for real-time release.

[0016] S1-1 monitors the status information of each node in the Internet of Things in real time through the environmental sensing module. The status information is divided into dynamic data and static data, among which: Dynamic data mainly reflects information such as the computing load, task queuing status and resource utilization of each node at the current moment, specifically including: terminal device status, edge node status, cloud server status and network communication status. Terminal device status is used to reflect the computing load and queuing status of IoT local devices, including current CPU utilization, remaining execution time of current tasks, and length of the queue of tasks to be processed. Edge node status is used to assess the resource carrying capacity and responsiveness of edge servers, including currently available computing resources and the cumulative computing load of currently queued tasks; Cloud server status is used to characterize the availability and queuing latency of remote cloud computing resources, including the current available computing power percentage and the current task queuing response time; Network communication status is used to estimate data transmission costs and response latency during task unloading, including available link bandwidth, inter-node link transmission latency, network signal strength, and packet loss rate.

[0017] Static data describes the hardware capabilities and basic configuration of each computing node. It is usually obtained through the registration process or system initialization phase when the device connects to the system. Specifically, it includes: terminal device, processor clock speed, floating-point operation capability, memory capacity, network interface type, transmission power, battery capacity, upper limit of edge server computing capability, upper limit of cloud server computing capability, and storage capacity.

[0018] S1-2, through the capability modeling unit in the environmental perception module, performs static data modeling and dynamic data capability estimation, and combines the results of the two to finally generate the capability vector of each computing node; Static data modeling involves modeling the hardware parameters in static data to construct the theoretical computing power and energy consumption limits of nodes. Specifically, this includes: estimating the theoretical computing power limit based on the device's processor clock speed, floating-point operation capability, and memory capacity; establishing an energy consumption limit model by recording the device's transmission power and battery capacity; and describing the theoretical computing resource limit based on the maximum supported computing power and storage capacity of edge servers and cloud servers.

[0019] Dynamic data capability estimation, which combines dynamic data to estimate the actual available computing power and resource consumption of the current device, specifically involves: estimating the task execution capacity and queue waiting impact of a node based on real-time task queue length and remaining execution time, forming a real-time timeline of its task processing efficiency; estimating actual power consumption based on historical power data and current load, and generating a computing power consumption curve; estimating the currently allocable computing power by combining remaining computing resources and task queue load; estimating task execution latency based on current response time and queue length; and generating a corresponding computing energy consumption curve based on historical energy consumption data.

[0020] Based on the results of static data modeling and dynamic data capability estimation, a standardized node capability vector is generated, as shown in the following formula:

[0021] in, The capability vector of the i-th computation node Based on current available computing power, Calculate energy consumption per unit. This is a status indicator for the task queue. This represents the average task response time. This represents the effective bandwidth of the i-th link.

[0022] S1-3, based on the collected dynamic data of network communication status and static location information of nodes, the network modeling unit of the environment perception module generates a link performance vector according to the following process: First, the bandwidth measurement results are reduced by the packet loss rate to obtain the effective bandwidth. Then, the propagation delay and queuing-processing delay are calculated using the distance between nodes and the real-time queuing length, and corrected in combination with the signal-to-noise ratio. Next, the terminal transmission power is multiplied by the data transmission time corresponding to the effective bandwidth to calculate the unit transmission energy consumption. Finally, the "effective bandwidth, comprehensive delay, and unit transmission energy consumption" are integrated into a link performance vector and written into the state database.

[0023] The dynamic data of network communication status includes: available link bandwidth, link transmission delay, network signal strength and packet loss rate; the static location information of nodes refers to the physical location information uploaded by each node during the system registration phase.

[0024] S1-4, through the state publishing unit of the environment perception module, merges the node capability vector generated in S1-2 and the link performance vector formed in S1-3 by timestamp, supplements the calculation of the remaining execution time and queuing delay of each node based on the current queue and available computing power, and finally, encapsulates all the information into an environment state vector, and publishes it to the global state library at a fixed period or event triggering mode, to ensure that the task parsing module, cost modeling module and policy generation module always obtain the latest system state, and make segmentation, unloading and scheduling decisions accordingly.

[0025] Step S2 involves using the task parsing module to perform structured modeling and load quantification assessment of the task to be processed, identifying suitable offloading boundaries, and matching feasible offloading destinations based on the current system environment status. Finally, candidate split point-target node mappings and their load parameters are generated, providing basic input for cost modeling and strategy generation.

[0026] S2-1, the task parsing module first extracts the execution form of the current task to be processed, and constructs a structured task model according to the task type: for conventional data processing tasks, a data flow graph (DAG) is generated, with nodes representing computation stages and edges representing data dependencies; for deep neural network inference tasks, a network layer sequence structure arranged in sequence is constructed.

[0027] Subsequently, the computational load and the amount of data transmitted to downstream units for each computing unit are extracted. The above data comes from the combined extraction of static analysis and runtime performance analysis tools for the task code.

[0028] S2-2, based on the obtained structured task model, quantifies the computational load and communication cost of each computational stage under different offloading schemes. Let the computational load of the i-th stage be... This refers to the number of computational instructions required for this stage; if the task is executed on a local device, an edge node, and a cloud server, their available computing power is respectively... , and The theoretical computation delay at different computing nodes during this stage is:

[0029] in, , , These represent the execution time required for stage i on the local, edge, and cloud servers, respectively. Let be the computational cost for the i-th stage. For the available computing power of local devices, The available computing power for edge nodes. The available computing power of the cloud server; Let the amount of intermediate result data in the i-th stage of the data transmission path from the local device to the target node be... The effective bandwidth of the i-th link is The propagation delay is The data transmission delay is:

[0030] in, The total communication latency for data transmission from the local machine to the server in stage i; Based on the unit computational energy consumption parameter in the node capability vector generated in S1-2, and the unit transmission energy consumption parameter in the link performance vector in S1-3, the execution energy consumption and transmission energy consumption under different offloading schemes are estimated respectively. The estimation method is as follows:

[0031] in, Let i be the energy consumption of the i-th stage executed locally. Calculate energy consumption for local equipment units. Let be the computational cost for the i-th stage. The energy consumption for sending the output data of stage i to the target node. Energy consumption per unit of transmission This represents the amount of output data for the i-th stage. By analyzing the combination of local execution, edge execution, cloud execution, and transmission path at each stage, a task execution cost table that can be used for offloading decisions is obtained.

[0032] S2-3, the boundaries of each stage in the task structure model are initially regarded as potential separable points, forming a set of candidate boundaries. Subsequently, based on the analysis results of S2-2, the segmentation points in B were evaluated one by one: If the data output at a certain boundary Much greater than the total computational load of its front end This results in transmission time or energy consumption far exceeding the computational cost, i.e.:

[0033] If a significant communication bottleneck is identified at a split point, it will be removed. If there are strong dependencies between certain stages or if the entire process needs to be executed, the corresponding boundaries will also be removed from B. This indicates the effective bandwidth of the link from the local device to the edge server; The set of filtered results is denoted as the set of valid candidate split points. And record each Corresponding local computational load and the amount of data transmitted .

[0034] S2-4, for each candidate split point In the current environment state vector published in S1-4, identify the set of target nodes that meet the performance conditions. It satisfies: Available computing power

[0035] Effective network bandwidth Enough to complete Data transmission, and satisfying:

[0036] in, The maximum tolerable delay allowed for the current task scheduling.

[0037] Ultimately, a set of candidate unloading solutions is formed:

[0038] The computational load and transmission cost corresponding to each combination will be used in the next step for cost modeling and optimal offloading strategy generation. The task parsing module will finally output the load parameter set corresponding to the candidate split point-target node mapping. .

[0039] Step S3: Based on the environmental state vector obtained in S1 and the candidate segmentation point-target node mapping formed in S2, as well as the calculation of communication parameters, this step constructs the execution delay model and energy consumption model respectively through the task cost modeling module, and constructs a unified delay-energy consumption joint cost function through normalization fusion, so as to realize the quantitative evaluation of the performance cost of each offloading scheme under the current system state.

[0040] S3-1, for each candidate unloading scheme, i.e., by the split point With unload node The binary pair formed Construct a model to estimate the total execution time of the application in its current state; The latency consists of four parts: the execution time of the first half of the task locally, the unloading and data transfer time, the computation time of the second half of the task on the remote node, and the queuing time on the current target node; the total execution time of the scheme is obtained by adding these four components together, as shown in the formula:

[0041] in, Total execution time This refers to the execution time of the first half of the task locally. To unload data transmission time, This refers to the computation time of the latter half of the task on the remote node. This represents the queuing time at the current target node.

[0042] If you choose to execute all locally, then only If the target is a cloud server, then and It will increase significantly, and factors such as link latency and task backlog need to be fully considered.

[0043] S3-2 considers the energy consumption of a task as the sum of its local execution and data transmission portions, constructing the following two energy consumption models: one is the local computation energy consumption, defined as... Secondly, there is the power consumption of wireless transmission, denoted as... In summary:

[0044] in, This represents the energy consumed by an IoT device in the initial stage of performing tasks locally. The power consumption per unit of IoT devices, This indicates the execution time of the task's initial phase on the local device; This represents the energy consumption during the process of uploading intermediate results of a task to the target node. Indicates the wireless transmission power of IoT devices. Indicates the upload and transfer time; Indicates the task is in the At each split point, the subsequent part is unloaded to the node. The total energy consumption generated.

[0045] If the entire task is completed locally, then If offloading is attempted, the upload power consumption in the edge or cloud direction must be comprehensively considered, which is particularly critical for battery-powered devices and serves as one of the rigid constraints on the feasibility of offloading.

[0046] S3-3, In order to unify delay and energy consumption into a single optimization index, a weighted comprehensive cost function is proposed to construct a unified joint delay-energy consumption cost function, as shown in the following equation:

[0047] in, At the task split point Unload the target node The comprehensive value of the next generation and These are the weighting coefficients. These are dimensionless indices after normalizing the total latency and total energy consumption, respectively. The total delay after normalization. This represents the normalized total energy consumption.

[0048] Normalization methods can be linearly scaled based on their respective historical maximum values: ,

[0049] in, Total task execution delay Total energy consumption for the task; , These are the maximum time delay and maximum energy consumption observed in history, respectively, and are used as the normalization denominator.

[0050] Robustness can also be improved by standard deviation normalization. In actual deployment, all candidate solutions can be normalized according to standard deviation normalization. The algorithm sorts the data and selects the minimum cost solution as the current optimal unloading path. It can also be extended to a multi-task global optimization objective function.

[0051] Step S4: Use reinforcement learning algorithms to generate optimal task segmentation and offloading decision strategies. The strategy generation module adopts the multi-agent deep reinforcement learning (MARL) method to model the offloading decision problem of multiple IoT devices as a Markov decision process (MDP). By training the agents of each device in a centralized manner, the overall system cost is optimized collaboratively.

[0052] S4-1, MDP Modeling and State Construction, models the multi-terminal task offloading process and uses Markov Decision Process (MDP) to describe the system dynamics. Each IoT device is an agent, and its state information consists of its remaining task load, task queue, available resources of candidate servers, and network link quality. Actions are defined as where the current task is split and to which target node the subsequent part is offloaded. The feedback after each action is executed is determined by the system operating cost, including execution latency and energy consumption, which is used as reward input to the learning system, ultimately obtaining a standard state-action space structure that can be input into the reinforcement learning policy network.

[0053] S4-2, Reward Function Design and Optimization Objective Definition: A joint evaluation function is constructed for the system's operating cost. The total task execution time and local device energy consumption are weighted and combined according to set weights to form a comprehensive cost value. The negative value of this value is used as a reward signal to feed back to the learning process, thereby achieving the optimization objective of "the lower the cost, the higher the reward". This mechanism encourages each agent to make strategy choices that minimize costs in a dynamic environment, forming a stable unloading behavior pattern.

[0054] S4-3, a multi-agent reinforcement learning algorithm design, adopts a centralized training and distributed execution architecture to design a multi-agent deep reinforcement learning algorithm. During the training phase, state-action-reward data of all devices are collected centrally, and a unified global value network is used to evaluate the current policy. Subsequently, the local policy networks of each device are updated synchronously. The policy network takes its local state as input and outputs the corresponding split point and offloading target node. In the training iteration, the policy optimization algorithm PPO is used to maintain the stability and convergence speed of policy updates, and finally obtains an approximately optimal offloading policy.

[0055] S4-4, Policy Output and Online Adaptive Update: Deploy the trained policy network and connect it to each IoT device as its local offloading decision-maker; whenever a task arrives at a device, the agent outputs the segmentation position and target node in the current state through the policy network, forms an offloading plan, and hands it over to the scheduling and communication module for execution. The strategy generation module supports a dynamic feedback mechanism, which transmits the latency and energy consumption results during the actual operation of the system back for subsequent strategy fine-tuning and retraining, thereby ensuring that the strategy can always adapt to environmental changes and maintain high efficiency during long-term operation.

[0056] Step S5 involves unloading the scheduling execution and integrating the feedback of the results. This is uniformly controlled by the scheduling communication module. Based on the strategy decision output in step S4, the task splitting and scheduling execution operations are completed, and the execution results are sent back to the environment state model to realize the task unloading control closed loop.

[0057] S5-1: Upon receiving the unloading policy, the scheduling communication module first divides the task into a local execution segment and a remote execution segment according to the specified splitting position. If the policy is to execute entirely locally, the local processor is started directly to complete the entire task; if there is an unloading requirement, the first subtask is executed locally first, and after reaching the splitting point, the execution is paused and intermediate output data is extracted as input for subsequent remote tasks. Task splitting can be implemented through a static splitting framework or a runtime breakpoint management mechanism to ensure that the intermediate state of the computation process can be captured and is complete.

[0058] S5-2, intermediate data packaging and remote transmission: After obtaining the segmentation point output, the local device encapsulates the data and sends it to the designated server node; during transmission, the scheduling communication module selects the optimal transmission protocol and path according to the network model, and dynamically adjusts the data sending rate and transmission mode according to the current network status; to ensure transmission quality, the system supports real-time monitoring of link status and anomaly response mechanism, such as enabling retransmission or path switching in a timely manner according to link congestion, signal strength or packet loss.

[0059] S5-3, Remote Task Scheduling and Execution Processing: After receiving intermediate data, the target server puts the task fragment into its internal computing scheduling process and arranges the execution order according to its current resource usage and queue status. The server can allocate lightweight isolation containers for unloaded tasks to reduce deployment costs, and during task execution, it feeds back progress information and resource usage status to the local scheduling and communication module in real time. If the server is currently idle, the task can start immediately; if there are already tasks to be processed in the queue, they are queued for execution and the estimated waiting time and total task time are recorded.

[0060] S5-4, Result Integration and Return Delivery: After the server completes the remote task processing, it transmits the result data back to the local device or the designated result recipient. In scenarios requiring integration, the scheduling communication module combines the local and remote results to construct the complete task output. For cases where integration is not required, such as neural network inference results, the system directly delivers them to the upper-layer application. The return data process supports integrity verification mechanisms, such as result hash comparison or digest confirmation, to improve system reliability and error recovery capabilities.

[0061] S5-5 executes feedback acquisition and system state updates. After the task is completed, the scheduling and communication module sends back key indicators such as the actual task running time, data transmission time, and local energy consumption to the system state model to correct previous performance prediction results and improve the accuracy of future unloading decisions. At the same time, the system updates the resource status on the local and server sides (such as releasing computing resources and updating queuing status), and the environment awareness module re-aggregates and publishes the current system status for use in subsequent new task decisions. If the online reinforcement learning mechanism is enabled, the system can also add the current state-action-feedback sample to the experience pool for periodic policy iteration training to achieve continuous adaptation to the long-term dynamic environment.

[0062] Please refer to Figure 2 , Figure 2 An information processing system architecture diagram based on the Internet of Things provided in this embodiment of the invention includes the following modules: The environmental perception module is responsible for real-time monitoring and collection of key operating parameters in the IoT system, including node computing power, energy consumption characteristics, task queuing status, network bandwidth and transmission latency, in order to construct and maintain global environmental status information of the system. The continuously updated environmental status of this module will serve as input for the task parsing module, cost modeling module and policy generation module to use.

[0063] The task parsing module is responsible for structural modeling and load assessment of the tasks to be processed. This module constructs a structured representation of the task based on its type, analyzes the computational and communication overhead of each subtask, selects reasonable candidate split points, and maps a set of feasible offloading target nodes based on the current environment state. The parsing results are then passed as input to the cost modeling module and the policy generation module.

[0064] The cost modeling module is responsible for constructing a joint cost model of latency and energy consumption for tasks under different partitioning and offloading schemes. This module receives output data from the environment perception module and the task parsing module, calculates the time overhead of local and remote execution, wireless transmission time and energy consumption, and then outputs the comprehensive cost value corresponding to each scheme through multi-indicator weighted fusion.

[0065] The strategy generation module generates task unloading decision strategies based on a centrally trained multi-agent reinforcement learning algorithm. This module calls the local agent model to output the optimal task split point and corresponding unloading target based on the current system state input and task characteristics, and optimizes the unloading path to minimize the overall cost. The strategy generation module supports online adaptive updates and can continuously optimize the strategy model based on execution feedback to cope with environmental changes.

[0066] The scheduling and communication module is responsible for scheduling and executing the offloading scheme output by the policy generation module in the actual environment. This module coordinates the splitting of local tasks, front-end execution, and intermediate result transmission, while controlling the target server to receive and process back-end tasks and managing data interaction and resource scheduling between devices. After the task is completed, this module will also collect actual execution indicators and feed the results back to the environment perception and policy generation module for model correction and policy update, thereby closing the system scheduling optimization loop.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations will be apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An information processing method based on Internet of Things, characterized by, Comprise the following steps: Step S1, the environment perception module acquires the dynamic data and static data of each computing node, respectively executes static data modeling and dynamic data capability estimation, obtains node capability vector and link performance vector, and encapsulates as environment state vector real-time release; Step S2, through the task analysis module, the structured modeling and load quantitative evaluation of the to-be-processed task are carried out, the segmentation boundary suitable for unloading is identified, and the feasible unloading direction is matched in combination with the current system environment state, and finally the candidate segmentation point-target node mapping and its load parameter set are generated; Step S3, based on the environment state vector obtained in S1 and the candidate segmentation point-target node mapping and its load parameter set formed in S2, the execution delay model and the energy consumption model are respectively constructed through the task cost modeling module, and a unified delay-energy consumption joint cost function is constructed through normalized fusion; Step S4, the optimal task segmentation and unloading decision strategy is generated by using the reinforcement learning algorithm, and the strategy generation module adopts the multi-agent deep reinforcement learning method to model the unloading decision problem of multiple Internet of Things devices as a Markov decision process; Step S5, the unloading scheduling execution and result integration feedback are controlled by the scheduling communication module, the task segmentation and scheduling execution operation are completed according to the strategy decision output in step S4, and the execution result is returned to the environment state model.

2. The information processing method based on Internet of Things according to claim 1, characterized in that: In step S1, further comprising the following sub-steps: S1-1, the state information of each node in the Internet of Things is monitored in real time by using the environment perception module, and the state information is divided into dynamic data and static data; S1-2, the static data modeling and dynamic data capability estimation are carried out by the capability modeling unit in the environment perception module, and the capability vector of each computing node is finally generated in combination with the results of the two; Static data modeling, namely modeling the hardware parameters in the static data, constructing the theoretical computing capacity and energy consumption upper limit of the node; Dynamic data capability estimation, namely estimating the actual available computing capacity and resource consumption of the current device in combination with the dynamic data; according to the results of static data modeling and dynamic data capability estimation, a standardized node capability vector is generated; S1-3, based on the collected network communication state dynamic data and node static position information, the network modeling unit of the environment perception module reduces the bandwidth measurement results according to the packet loss rate to obtain the effective bandwidth, then calculates the propagation delay and queuing-processing delay according to the distance between nodes and the real-time queue length and corrects them according to the signal-to-noise ratio, then multiplies the terminal transmission power and the data transmission time corresponding to the effective bandwidth to calculate the unit transmission energy consumption, and finally assembles the "effective bandwidth, comprehensive delay, unit transmission energy consumption" into a link performance vector and writes it into the state library; S1-4, through the state publishing unit of the environment perception module, the node capability vector and the link performance vector are combined according to the time stamp, the remaining execution time and the queuing delay of each node based on the current queue and available computing power are supplemented, and all information is uniformly encapsulated as an environment state vector and released to the global state library in a fixed period or event triggered manner.

3. The Internet of Things-based information processing method according to claim 1, characterized in that: In step S2, the following sub-steps are further included: S2-1, the task analysis module extracts the execution form of the current to-be-processed task, and constructs a structured task model according to the task type: for a regular data processing task, a data flow graph is generated; for a deep neural network inference task, a network layer sequence structure arranged in sequence is constructed; then the calculation amount of each calculation unit and the data amount transmitted to the downstream unit are extracted; S2-2, based on the obtained structured task model, the calculation load and communication cost of each calculation stage under different offloading schemes are quantified, and the time required for execution of the i-th stage on the local, edge, and cloud server is obtained, as well as the total communication delay of the data transmitted from the local to the server in the i-th stage; Based on the unit calculation energy consumption parameter in the node capability vector generated in S1-2 and the unit transmission energy consumption parameter in the link performance vector in S1-3, the execution energy consumption and transmission energy consumption under different offloading schemes are estimated respectively, and through the combination analysis of the local execution, edge execution, cloud execution, and transmission path of each stage, a task execution cost table that can be used for offloading decision is obtained; S2-3, the boundaries of each stage in the structured task model are initially regarded as potential divisible points to form a candidate boundary set, and the dividing points in the candidate boundary set are evaluated and screened one by one in combination with the analysis results of S2-2, the screened set is recorded as the effective candidate dividing point set, and the local part calculation amount and transmission data amount corresponding to each are recorded; S2-4, for each candidate split point In the current environment state vector issued at S1-4, identify a set of target nodes that satisfy the performance condition which satisfies: Available computing power ; Network effective bandwidth sufficient to complete data transmission, and meet: Wherein, The maximum tolerable delay allowed by the current task scheduling; finally form a candidate unloading scheme set The task analysis module finally outputs the load parameter set corresponding to the candidate split point-target node mapping pair .

4. The Internet of Things-based information processing method according to claim 1, characterized in that: In step S3, the following sub-steps are further included: S3-1, for each candidate offloading scheme, i.e. a binary tuple composed of a dividing point and an offloading node, a total execution time estimation model in the current state is constructed; the delay is composed of four parts, and the total execution time of the scheme is obtained by accumulation; S3-2, considering the energy consumption of the task as the sum of the local execution part and the data transmission part, two energy consumption models are constructed as follows: one is the local computing energy consumption; the other is the wireless transmission energy consumption, and the following is obtained: wherein, represents the energy consumed by the IoT device before executing the local part of the task, represents the energy consumption in the process of uploading the intermediate result of the task to the target node, represents the total energy consumption generated by the task at the split point, and the subsequent part is unloaded to the node . S3-3, construct a unified delay-energy joint cost function, as shown in the formula: wherein, At the task split point , the comprehensive cost value under the offload target node , and is a weight coefficient, respectively, are dimensionless indexes after normalization processing of the total time delay and the total energy consumption, is the normalized total time delay, is the normalized total energy consumption.

5. The Internet of Things-based information processing method according to claim 1, characterized in that: In step S4, the following sub-steps are further included: S4-1, a Markov decision process is used to describe the system dynamics, wherein each IoT device is regarded as an intelligent agent, and the state information is composed of the residual task load of the device itself, the task queue, the available resources of the candidate server, and the network link quality; the action is defined as the division of the current task from which positions and the offloading of the subsequent part to which target node; The feedback after each action execution is determined by the system running cost, and finally a standard state-action space structure that can be input to the reinforcement learning strategy network is obtained; S4-2, the total execution time of the task and the local device energy consumption are combined into a comprehensive value according to the set weight, and the negative value thereof is used as a reward signal feedback to the learning process; S4-3, a multi-agent deep reinforcement learning algorithm is designed using a centralized training and distributed execution architecture. In the training phase, the state-action-reward data of all devices is collected centrally, and a unified global value network is used to evaluate the current policy. Then, the local policy network of each device is updated synchronously. The policy network takes the local state as input and outputs the corresponding segmentation point and offloading target node; In the training iteration, the policy optimization algorithm PPO is used; S4-4, the trained policy network is deployed and accessed by each IoT device as its local offloading decision maker. Whenever a task arrives at the device, the agent outputs the segmentation position and target node under the current state through the policy network, forms an offloading scheme, and hands it over to the scheduling communication module for execution. The policy generation module supports a dynamic feedback mechanism to return the delay and energy consumption results during the actual operation of the system.

6. The Internet of Things-based information processing method of claim 1, further comprising the following sub-steps in step S5: S5-1, the scheduling communication module divides the task into local execution segments and remote execution segments according to the specified segmentation position. If the policy is complete local execution, the local processor directly completes the entire task. If there is a need for offloading, the local device executes the sub-tasks before the segmentation point, pauses after executing to the segmentation point, and extracts the intermediate output data as the input for the subsequent remote task. Task segmentation can be achieved through a static segmentation framework or a runtime breakpoint management mechanism; S5-2, after obtaining the segmentation point output, the local device encapsulates and sends the data to the specified server node. During transmission, the scheduling communication module selects the most suitable transmission protocol and path according to the network model, and dynamically adjusts the data transmission rate and transmission mode according to the current network state; S5-3, after receiving the intermediate data, the target server puts the task segment into the internal computing scheduling process, arranges the execution order according to its current resource occupation and queue state, and feeds back the progress information and resource occupation state to the local scheduling communication module in real time during task execution; S5-4, when the server completes the remote task processing, the result data is transmitted back to the local device or the specified result receiver. In the case of integration, the scheduling communication module combines the local and remote results to construct the complete task output; For cases that do not require integration, the task is directly handed over to the upper layer application for use; S5-5, after the task execution is completed, the scheduling communication module returns the key indicators to the system state model for correcting the previous performance prediction results. At the same time, the system updates the local and server-side resource states, and the environment perception module re-summarizes and publishes the current system state; If the online reinforcement learning mechanism is enabled, the system adds the state-action-feedback sample to the experience pool for periodic policy iteration training. It includes:

7. An information processing system based on Internet of Things, applied to the information processing method based on Internet of Things in any of claims 1-6, characterized in that, An environment perception module that is responsible for real-time monitoring and collection of key operating parameters in the Internet of Things system, and for building and maintaining system global environment state information; ​ The task analysis module is responsible for the structural modeling and load evaluation of the to-be-processed task, constructs a structured representation according to the task type, analyzes the calculation and communication overhead of each subtask, screens reasonable candidate segmentation points, and maps the feasible unloading target node set in combination with the current environment state; The cost modeling module is responsible for constructing a delay and energy consumption combined cost model of the task under different segmentation and unloading schemes, and outputting the comprehensive cost value corresponding to each scheme; The strategy generation module generates a task unloading decision strategy based on a centralized trained multi-agent reinforcement learning algorithm, inputs the current system state and task characteristics, and calls a local agent model to output the optimal task segmentation point and the corresponding unloading target; The scheduling communication module is responsible for scheduling and executing the unloading scheme output by the strategy generation module in the actual environment, controlling the target server to receive and process the subsequent tasks, and managing the data interaction and resource scheduling between devices.

Images