Internet of Things edge computing task unloading method based on integrated gateway

Abstract

The invention discloses an Internet of Things edge computing task unloading method based on a comprehensive gateway, which belongs to the technical field of computer data processing, and comprises the following steps: acquiring tasks, equipment and network data of Internet of Things equipment, and fusing to generate a unified information view; determining task priorities based on the view by applying a dynamic priority evaluation mechanism, and generating a dynamic priority list; applying the time sequence analysis model to generate short-term prediction data; in combination with the dynamic priority list and the short-term prediction data, multi-objective optimization is performed through a collaborative decision model, and a task unloading strategy is generated; generating feedback data according to the task unloading strategy; and generating an adjustment signal by using the feedback data, and updating the dynamic priority evaluation mechanism and the collaborative decision model. According to the method, the unified information view is generated by fusing the multi-source data, collaborative decision making is carried out in combination with dynamic priority evaluation and short-term prediction data, and the model is continuously optimized by utilizing feedback data, so that self-adaptive, forward-looking and efficient task unloading can be realized.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and in particular to a method for offloading IoT edge computing tasks based on an integrated gateway. Background Technology

[0002] IoT edge computing is an emerging computing paradigm that moves computing power and data storage from the centralized cloud to the network edge, closer to the IoT devices at the data source. By executing compute-intensive or latency-sensitive tasks on integrated gateways or edge servers, network bandwidth pressure and response latency can be effectively reduced. Task offloading, as one of the core technologies of edge computing, lies in deciding whether a computing task should be executed locally on resource-constrained IoT devices or offloaded to edge nodes with richer resources. The quality of this decision directly affects the performance of the entire system and the user experience.

[0003] Existing IoT task offloading methods typically rely on static rules or decision models based on the current instantaneous state. For example, some methods simply decide whether to offload tasks based on their preset priorities or the device's current remaining battery power. Other methods, while considering more factors such as network conditions and server load, still rely on a snapshot of the system at the moment of decision, lacking the ability to predict future changes in the system's state. Furthermore, task priorities are usually fixed and cannot dynamically adapt to changing circumstances.

[0004] However, the aforementioned existing technical solutions have significant technical shortcomings. Due to the highly dynamic and uncertain nature of the IoT environment, network load and device resource conditions can change drastically at any time. Making decisions based solely on instantaneous states often leads to short-sighted and suboptimal offloading strategies. Furthermore, fixed task priorities cannot accurately reflect the actual urgency of tasks in specific contexts, easily causing resource misallocation. Simultaneously, most of these methods lack an effective feedback and learning mechanism, failing to learn from the execution effects of historical decisions, thus making it difficult to adapt to constantly changing environments and application requirements, resulting in overall system performance failing to reach its optimal level. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an IoT edge computing task offloading method based on a comprehensive gateway. This method generates a unified information view by fusing multi-source data, performs collaborative decision-making by combining dynamic priority evaluation and short-term forecast data, and continuously optimizes the model using feedback data, thereby achieving adaptive, forward-looking, and efficient task offloading.

[0006] The above objectives can be achieved through the following approach: A method for offloading IoT edge computing tasks based on a comprehensive gateway includes: acquiring and fusing task attribute data, device status data, and network environment data from IoT devices to generate a unified information view; calculating the priority of each task based on the unified information view using a dynamic priority evaluation mechanism to generate a dynamic priority list; collecting historical network load data and device resource usage data, and using a time series analysis model for prediction to generate short-term prediction data; combining the dynamic priority list and the short-term prediction data, performing multi-objective optimization through a collaborative decision-making model to generate a task offloading strategy; controlling task execution according to the task offloading strategy, and collecting actual latency and energy consumption during execution to generate feedback data; using the feedback data to generate an adjustment signal, and updating the dynamic priority evaluation mechanism and the collaborative decision-making model based on the adjustment signal.

[0007] Optionally, generating a unified information view includes: receiving a task generation request from an IoT device; performing compliance verification on the priority declaration field according to a predefined urgency level classification standard to generate a verified task request; extracting task type descriptors and task size parameters from the verified task request to generate task attribute data; collecting current device power data and processor performance indicators from real-time monitoring data to generate device status data; measuring the transmission rate and response time parameters of the network channel from real-time monitoring data to generate network environment data; and performing standardized transformation and data association mapping on the task attribute data, the device status data, and the network environment data to generate a unified information view.

[0008] Optionally, generating the dynamic priority list includes: extracting the task attribute data and the device status data from the unified information view; performing a weighted calculation on the task attribute data and the device status data to generate an initial priority score; applying historical performance correction to the initial priority score to calculate a final priority score; and arranging all tasks in descending order according to the final priority score to generate a dynamic priority list.

[0009] Optionally, the step of correcting the initial priority score based on historical performance includes: retrieving historical execution efficiency data associated with the current task type and calculating a historical performance correction factor; multiplying the initial priority score by the historical performance correction factor to obtain the final priority score.

[0010] Optionally, generating short-term prediction data includes: periodically collecting network throughput samples and device CPU utilization samples from historical task execution records to construct a multidimensional time series dataset; performing missing value imputation and outlier smoothing on the multidimensional time series dataset to generate a standardized time series data stream; inputting the standardized time series data stream into a pre-trained time series analysis model, and outputting a network load prediction sequence and a device resource prediction sequence containing multiple future time steps through a multi-step rolling prediction method; integrating the network load prediction sequence and the device resource prediction sequence, and adding a prediction confidence score to generate short-term prediction data.

[0011] Optionally, the task offloading strategy includes: evaluating multiple alternative execution schemes for each task in the dynamic priority list, and calculating the predicted task latency and predicted energy consumption of each alternative execution scheme in combination with the short-term prediction data; applying a utility function for balancing performance to convert the predicted task latency and predicted energy consumption into a comprehensive utility score; selecting alternative execution schemes for each task according to the comprehensive utility score, and combining them to generate a task offloading strategy.

[0012] Optionally, converting the predicted task delay and predicted energy consumption into a comprehensive utility score includes: assigning dynamic weighting coefficients to the predicted task delay and the predicted energy consumption according to the position of the task in the dynamic priority list; and performing a weighted summation of the predicted task delay and the predicted energy consumption using the dynamic weighting coefficients to obtain a comprehensive utility score.

[0013] Optionally, generating feedback data includes: parsing the task unloading strategy to determine the task execution location of each task; sending control commands to IoT devices or edge servers based on the task execution location, and recording the actual latency and actual energy consumption during task execution to obtain performance logs; and integrating and processing the performance logs to generate feedback data.

[0014] Optionally, updating the dynamic priority evaluation mechanism and the collaborative decision-making model according to the adjustment signal includes: extracting the actual delay and actual energy consumption from the feedback data, and comparing them with the predicted task delay and the predicted energy consumption to generate a deviation value; generating an adjustment signal containing the adjustment direction and magnitude according to the magnitude and direction of the deviation value; and iteratively updating the weight rules of the dynamic priority evaluation mechanism and the multi-objective optimization parameters of the collaborative decision-making model using the adjustment signal.

[0015] Based on the same inventive concept, this invention also provides an IoT edge computing task offloading system based on a comprehensive gateway. The system includes: a data fusion module for acquiring and fusing task attribute data, device status data, and network environment data from IoT devices to generate a unified information view; a priority evaluation module for calculating the priority of each task based on the unified information view using a dynamic priority evaluation mechanism to generate a dynamic priority list; a prediction data generation module for collecting historical network load data and device resource usage data, applying a time series analysis model for prediction, and generating short-term prediction data; a collaborative decision-making module for combining the dynamic priority list and the short-term prediction data, performing multi-objective optimization through a collaborative decision-making model, and generating a task offloading strategy; an execution control module for controlling task execution according to the task offloading strategy, collecting actual latency and energy consumption during execution, and generating feedback data; and a feedback optimization module for generating adjustment signals using the feedback data, and updating the dynamic priority evaluation mechanism and the collaborative decision-making model based on the adjustment signals.

[0016] Compared with the prior art, the present invention has the following advantages: This invention generates a unified information view by integrating multi-dimensional data such as task attributes, device status, and network environment, providing comprehensive and standardized contextual information for task offloading decisions. This overcomes the decision-making blindness caused by traditional methods relying on single or partial information, enabling offloading strategies to be based on a precise perception of the overall system status, thereby improving the accuracy and reliability of decisions and ensuring that resources are allocated according to actual needs.

[0017] This invention introduces a collaborative decision-making model that combines a dynamic priority evaluation mechanism with short-term forecast data. This mechanism not only assesses the immediate urgency of tasks in real time but also anticipates future network and equipment resource availability trends. This decision-making approach, which combines immediate needs with future states, achieves a shift from reactive scheduling to predictive planning, effectively avoiding potential performance bottlenecks and making task offloading strategies more forward-looking and timely.

[0018] This invention constructs a closed-loop feedback system from decision execution to model optimization. By continuously collecting actual performance data of task execution and comparing it with predicted values ​​to generate adjustment signals, the system can autonomously learn and iteratively update its core priority evaluation and collaborative decision-making models. This adaptive optimization capability enables the entire method to continuously adapt to dynamically changing environments and task loads, ensuring high efficiency and robustness in long-term operation and improving the system's intelligence level.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an IoT edge computing task offloading method based on an integrated gateway, according to an embodiment of the present invention.

[0022] Figure 2 This is a unified information view radar chart according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the dynamic priority score correction process in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of multi-dimensional time-series resource load prediction according to an embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of an IoT edge computing task offloading system based on a comprehensive gateway, according to an embodiment of the present invention. Detailed Implementation

[0026] 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 some embodiments of the present invention, not all embodiments. 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.

[0027] Reference Figure 1 One embodiment of the present invention proposes an IoT edge computing task offloading method based on a comprehensive gateway. It adopts the method of fusing multi-source data to generate a unified information view, combines dynamic priority evaluation and short-term prediction data for collaborative decision-making, and uses feedback data to continuously optimize the model, thereby achieving adaptive, forward-looking and efficient task offloading.

[0028] The method described in this embodiment specifically includes: S1. Acquire task attribute data, device status data, and network environment data of IoT devices and integrate them to generate a unified information view; Optionally, generating a unified information view includes: Receive task generation requests from IoT devices, perform compliance verification on the priority declaration field according to predefined urgency level classification standards, and generate verified task requests; Extract the task type descriptor and task size parameter from the verified task request to generate task attribute data; Collect current device power data and processor performance indicators from real-time monitoring data to generate device status data; Network environment data is generated by measuring the transmission rate and response time parameters of the network channel from real-time monitoring data. The task attribute data, device status data, and network environment data are standardized, transformed, and mapped to generate a unified information view.

[0029] Specifically, the process begins with the integrated gateway receiving a task generation request from an IoT device via a standard IoT communication protocol. This request includes a priority declaration field, which the integrated gateway first verifies for compliance. This verification is based on a predefined urgency grading standard, such as dividing priorities into several levels, and checking whether the declared value in the request is within a valid range. Only task requests that pass verification are considered verified and proceed to the next processing step. Next, key task attribute data is parsed and extracted from the data payload of the verified task request, primarily including a task type descriptor to identify the task computation type and task size parameters to quantify the task computation load. Simultaneously, the integrated gateway acquires status information related to the task source device by monitoring the data stream in real time. This process involves actively querying or passively receiving telemetry data from the device, collecting current device power data and processor performance indicators such as CPU utilization or available memory, which together constitute device status data. Furthermore, to assess the feasibility of data transmission, network channel parameters connecting the IoT device and the integrated gateway are measured from the real-time monitoring data. The current transmission rate and response time parameters are determined by sending probe packets or analyzing historical transmission records; these parameters constitute network environment data. The final step is to fuse the three types of heterogeneous data: task attribute data, device status data, and network environment data. To address the issue of inconsistent data units, the data is first standardized. This standardization can be achieved by applying the following min-max normalization formula to map the original values ​​to a uniform range, typically between 0 and 1, thus transforming them into dimensionless scalars: , in, This represents the standardized value; These are the raw measurements collected, such as specific battery percentages or network speed values; and These are the theoretical minimum and maximum values ​​for this data item, preset according to the parameter type. After standardization transformation, these standardized task attribute data, device status data, and network environment data are mapped and associated, that is, all information related to the same task request is organized into a structured data record, forming the final unified information view. For example... Figure 2 As shown, the unified information view can be visualized using a radar chart. Each coordinate axis in the chart represents a standardized information dimension, and each closed region represents a comprehensive status profile of a specific task across all dimensions.

[0030] For example, in a smart security scenario, a high-definition camera (IoT device) deployed at the entrance generates a "face recognition" task after detecting human movement. The camera sends a task generation request to the integrated gateway via the Message Queuing Telemetry Transport Protocol (MQTT), with the priority declaration field in the request set to "urgent." The integrated gateway's urgency level classification includes "urgent," so the verification passes, and the task becomes a verified task request. Subsequently, the system parses the task payload, extracting task attribute data, including a task type descriptor of "FaceRecognition-HD" and a task size parameter of "5MB." Simultaneously, the integrated gateway actively queries the camera's status via telemetry, obtaining device status data such as a battery level of "20%" and a processor performance indicator of "CPU utilization of 85%." The system then analyzes recent communication records with the camera to obtain network environment data such as a transmission rate of "5Mbps" and a response time parameter of "80ms." Next, the system performs standardization transformations on these raw values, such as the battery level data... It is 20. =0, Given 100, the calculation yields... The value is 0.2. Similarly, all parameters such as CPU utilization, task size, and network speed are standardized in this way to obtain a set of dimensionless values. Finally, the system integrates these standardized task attributes, device status, and network environment data into a structured data record, namely {Task ID: T001, Task type standard value: 0.9, Task size standard value: 0.6, Device battery standard value: 0.2, CPU utilization standard value: 0.85,...}, constituting a unified information view for the face recognition task. This method transforms fragmented, multi-dimensional, and dimensionally diverse raw state information into a standardized and structurally complete decision-making basis, solving the problems of information inconsistency and incompleteness in the IoT environment. It provides a data foundation for subsequent dynamic priority evaluation and offloading decisions, ensuring the accuracy of decisions and global perception capabilities, thereby improving the intelligence level of the entire task offloading method and the reliability of system operation.

[0031] S2. Based on the unified information view, apply a dynamic priority evaluation mechanism to calculate the priority of each task and generate a dynamic priority list; Optionally, generating the dynamic priority list includes: Extract the task attribute data and the device status data from the unified information view; The task attribute data and the device status data are weighted and calculated to generate an initial priority score; The initial priority score is corrected for historical performance, and the final priority score is calculated. All tasks are sorted in descending order according to their final priority scores to generate a dynamic priority list.

[0032] Specifically, the process first extracts relevant task attribute data and device status data for each task from the generated unified information view. The task attribute data includes intrinsic characteristics such as the task's urgency and computational complexity, while the device status data reflects the current operating status of the source device, such as remaining battery power and processor load. Next, these standardized data are weighted to generate an initial priority score. This calculation is implemented using a multi-factor weighted model that combines the inherent importance of the task with the urgency of the device's current status. , in, Represents the initial priority score; This represents a comprehensive task attribute component composed of a weighted average of multiple standardized task attribute data. It represents a comprehensive equipment status component composed of multiple standardized equipment status data weighted together; and These are preset weighting coefficients, representing the relative importance of task attributes and device status in priority evaluation. These coefficients are pre-configured by the system based on the overall strategy or dynamically adjusted through learning optimization, and are all dimensionless values. After obtaining the initial priority score, it is further corrected based on historical performance. This correction aims to incorporate the task's historical execution experience, making the priority evaluation more forward-looking. Historical execution efficiency data associated with the current task type is retrieved, and a historical performance correction factor is calculated. The final priority score is obtained by multiplying the initial priority score by this correction factor. Finally, all tasks to be processed are sorted in descending order according to their calculated final priority scores, forming an ordered queue, which is the dynamic priority list. This list will serve as the primary basis for subsequent task unloading decisions.

[0033] For example, there are two tasks to be processed in the system. Task A is the "face recognition" task, and its standardized task attribute composite component is extracted from the unified information view. The value is 0.8, representing the overall component of equipment status. The value is 0.9. Task B is a routine task of "reporting ambient temperature and humidity," originating from a sensor. It is 0.2. The weight is 0.1. This is the system's preset weighting coefficient. It is 0.5. The initial priority score is 0.5. First, calculate the initial priority score. ; Next, the system checked historical records and found that the "face recognition" task had historically had low execution efficiency, with a historical performance correction factor of 1.2; while the "temperature and humidity reporting" task performed stably, with a correction factor of 1.0. The final priority score was... ; Finally, the system compares the two final scores and ranks task A before task B, forming a dynamic priority list. This method achieves a multi-dimensional dynamic priority evaluation mechanism by comprehensively considering the static attributes of tasks, the dynamic state of devices, and historical execution performance. It breaks away from the traditional static priority division method based on fixed rules or single declarations, and can adjust the processing order of tasks in real time and intelligently.

[0034] Optionally, the step of correcting the initial priority score based on historical performance includes: Retrieve historical execution efficiency data associated with the current task type and calculate the historical performance correction factor; The initial priority score is multiplied by the historical performance correction factor to obtain the final priority score.

[0035] Specifically, this process first requires maintaining a historical task execution record database, which stores the type, execution parameters, and final performance metrics such as actual latency and resource consumption for each completed task. When the initial priority score of a new task needs to be adjusted, the task type descriptor is first extracted from that task. Using this descriptor as an index, the historical task execution record database is searched to find all historical execution records with the same type as the current task. Subsequently, a historical performance correction factor is calculated based on these retrieved historical records. The final priority score is obtained using the following formula: , in, This is the final priority score obtained after correction; The initial priority score is a dimensionless value calculated by weighting task attribute data and device status data. It is a historical performance correction factor, derived by analyzing historical data. Specifically, The historical performance correction factor can be determined by the ratio between the historical average execution time of a specific task type and the baseline execution time for that task type, or by the statistical average of other dimensionless performance metrics that reflect execution efficiency. For example, if the historical average execution time of a certain type of task is significantly higher than its theoretical baseline time, its historical performance correction factor will be greater than 1, and vice versa. Figure 3 As shown in the figure, the priority scores of four different tasks have changed. The initial priority score is calculated based on the current state, while the final priority score is obtained by multiplying the initial score by the respective historical performance correction factor.

[0036] For example, a new "VideoStream Transcoding" task with an initial priority score of 0.7 requires historical performance correction. The system extracts its task type descriptor "VideoStreamTranscoding" and finds 100 records of the same type of task in the historical task execution record database. From these records, the system calculates the historical average execution time of this type of task to be 300 milliseconds. Simultaneously, the system configuration stores a baseline execution time for this task type, measured under ideal conditions, which is 250 milliseconds. Based on this, the system calculates the historical performance correction factor. The initial priority score is calculated by dividing the historical average execution time by the baseline execution time, i.e., 300ms / 250ms, resulting in a dimensionless value of 1.2. Finally, the system multiplies the initial priority score by this factor to obtain the final priority score. This method adds a self-learning and experience feedback dimension to the priority evaluation mechanism, making priority evaluation more forward-looking and accurate, thereby optimizing the overall allocation efficiency of system resources, improving the intelligence level of task scheduling, and enhancing the operational stability of the entire system.

[0037] S3. Collect historical network load data and device resource usage data, apply time series analysis models to make predictions, and generate short-term forecast data. Optionally, generating short-term forecast data includes: A multidimensional time series dataset is constructed by periodically collecting network throughput samples and device CPU utilization samples from historical task execution records. Missing value imputation and outlier smoothing are performed on the multidimensional time series dataset to generate a standardized time series data stream; The standardized time-series data stream is input into a pre-trained time-series analysis model, and a network load prediction sequence and a device resource prediction sequence containing multiple future time steps are output through a multi-step rolling prediction method. The network load prediction sequence and the device resource prediction sequence are integrated, and a prediction confidence score is added to generate short-term prediction data.

[0038] Specifically, the process begins by periodically extracting key performance indicator samples from historical task execution records. Specifically, network throughput samples are collected to reflect historical network communication load, and CPU utilization samples are collected to reflect the historical resource usage intensity of edge devices and servers. These timestamped samples are organized to construct a multidimensional time-series dataset, where each data point contains information across multiple dimensions, including time, network load, and device load. Subsequently, due to potential interruptions or interference in the actual collected data, a data preprocessing stage is initiated. This stage performs missing value imputation on the multidimensional time-series dataset, for example, by using the mean or interpolation of neighboring points to fill data gaps, and outlier smoothing, such as applying moving average filters to weaken transient, unrepresentative data spikes. These cleaning operations generate a continuous and stable standardized time-series data stream. This standardized time-series data stream is then fed into a pre-trained time-series analysis model. This model, such as a Long Short-Term Memory (LSTM) network or an Autoregressive Integral Moving Average (ARIMA) model, has learned the inherent patterns of system load changes by studying historical data. The model employs a multi-step rolling forecasting method. After predicting data for the first future time step, this predicted value is used as new input to predict data for the second time step, and so on, ultimately outputting a network load forecast sequence and a device resource forecast sequence containing multiple consecutive future time steps. Finally, these two sets of forecast sequences are integrated and packaged into a structured data object. To enhance the reliability of the decision, a forecast confidence score is attached to each set of forecast sequences, quantifying the model's assessment of the accuracy of its predictions. The integrated data packet constitutes the final short-term forecast data, depicting the expected state of network and device resources over a future period for subsequent collaborative decision-making models. Figure 4 As shown, this process is carried out for multi-dimensional time series data. The upper and lower sub-graphs illustrate the generation process of the device resource prediction sequence and the network load prediction sequence, respectively. In each sub-graph, the solid line represents the collected historical samples, the dashed line after the prediction starting point represents the future prediction sequence output by the time series analysis model, and the shaded area surrounding the prediction sequence represents the prediction confidence level.

[0039] For example, the integrated gateway needs to predict its own CPU load and the network load of its local area network (LAN) over the next 15 minutes. The system first retrieves minute-by-minute CPU utilization and network throughput samples from the past 6 hours of historical data, aligning these two sets of data by timestamp to form a two-dimensional time-series dataset containing 360 data points. During data preprocessing, the system detects that due to a brief system restart, two data points are missing for both CPU utilization and network throughput; these are filled by averaging the two missing data points. Next, the system applies a moving average filter to smooth out outliers in the entire multi-dimensional sequence, eliminating irrelevant instantaneous spikes. The processed, normalized time-series data stream is then input into a pre-trained Long Short-Term Memory (LSTM) network model using historical data. This model has learned the gateway's overall load pattern, such as periodic CPU spikes at each hour and consistently high network traffic during peak hours. The model employs a multi-step rolling prediction method, simultaneously outputting two sets of prediction sequences covering 15 future time steps: one set predicts device resource utilization (e.g., predicting CPU utilization to reach 90% at the 10th minute); the other set predicts network load (e.g., predicting network connectivity for the next 5 minutes, but a sharp drop in throughput after the 8th minute). Finally, the system integrates and packages these two sets of prediction sequences, assigning them prediction confidence scores (e.g., 95% confidence for CPU and 92% confidence for network) to form the final short-term prediction data. This method shifts offloading decisions from solely relying on the instantaneous, current system state to comprehensively considering expected load conditions over a future period, thus transforming reactive to predictive decision-making. This forward-looking capability allows the system to proactively avoid impending network congestion or device overload, selecting the offloading scheme with the best overall performance over the future. This improves the timeliness and robustness of task offloading strategies, ensuring the continued effectiveness of decisions over a future period, thereby enhancing the performance stability and resource utilization efficiency of the entire IoT edge computing system.

[0040] S4. Combining the dynamic priority list and the short-term forecast data, multi-objective optimization is performed through a collaborative decision-making model to generate a task unloading strategy. Optionally, the generation task unloading strategy includes: For each task in the dynamic priority list, multiple alternative execution schemes are evaluated, and the predicted task delay and predicted energy consumption of each alternative execution scheme are calculated in combination with the short-term forecast data. A utility function for balancing performance is applied to convert the predicted task delay and predicted energy consumption into a comprehensive utility score; Based on the overall utility score, alternative execution plans are selected for each task and combined to generate a task unloading strategy.

[0041] Specifically, for each task in the list, all possible alternative execution schemes are evaluated first. An alternative execution scheme specifies the task's execution location, such as execution locally on the IoT device, offloading to a unified gateway, or offloading to an edge server in the network. For each alternative execution scheme, a detailed performance forecast is performed using short-term prediction data. Specifically, the predicted task latency required for data transmission is calculated using the predicted network load forecast sequence, and the time consumed in the task computation stage is estimated using the predicted device resource forecast sequence; simultaneously, the corresponding predicted energy consumption is calculated based on the device's energy consumption model and the predicted execution time. Thus, each alternative execution scheme is associated with a set of predicted performance metrics, namely predicted task latency and predicted energy consumption. Next, to balance the conflicting objectives of latency and energy consumption, a utility function is applied. This function aims to integrate multi-dimensional performance metrics into a single scalar value, namely a comprehensive utility score, allowing for direct comparison of the merits of different schemes. After calculating the comprehensive utility score for all alternative execution schemes for each task, the scheme with the best score is selected for that task, for example, the scheme with the lowest score. By combining the optimal choices of all tasks in the dynamic priority list, a complete task unloading strategy is formed that includes all task allocation decisions.

[0042] For example, for the high-priority task "face recognition" in the dynamic priority list, the system evaluates three alternative execution schemes. Scheme 1: execute locally on the camera when its battery is at 20%; Scheme 2: offload to the integrated gateway; Scheme 3: offload to a remote edge server. The system uses short-term prediction data to estimate performance. Scheme 1: predicted task latency of 400ms, predicted energy consumption of 5% of the device battery. Scheme 2: predicted network and computing resources of the gateway will be relatively idle in the near future, predicted task latency of 150ms, predicted energy consumption of 0.1% of the gateway. Scheme 3: predicted network path congestion, predicted task latency as high as 500ms, predicted energy consumption of 0.05% of the server. Next, the system applies a utility function, which assigns a very high weight to latency and a low weight to energy consumption based on the high priority of the task, and calculates the comprehensive utility score of the three schemes. Ultimately, the system finds that the score of Scheme 2 is significantly lower than the other two, so the system selects Scheme 2, which has the lowest score, for the task, i.e., offloading to the integrated gateway. After repeating this process for all tasks in the list, all optimal choices are aggregated to form the final task offloading strategy. This method constructs a systematic and forward-looking decision-making framework to generate task offloading strategies. By evaluating multiple alternative execution plans and utilizing short-term forecast data, decisions are no longer limited to the current system state but can anticipate and adapt to future environmental changes, thereby avoiding performance bottlenecks caused by short-sighted decisions.

[0043] Optionally, converting the predicted task delay and predicted energy consumption into a comprehensive utility score includes: Based on the position of the task in the dynamic priority list, assign dynamic weighting coefficients to the predicted task delay and the predicted energy consumption; The comprehensive utility score is obtained by weighting and summing the prediction task delay and the prediction energy consumption using the dynamic weighting coefficients.

[0044] Specifically, this method first requires dimensionless processing of performance indicators with different dimensions to ensure logical consistency of the calculation. Specifically, for each alternative execution plan, the predicted task delay and predicted energy consumption are compared with a set of benchmark values ​​or possible maximum and minimum values. Through normalization, such as min-max normalization, they are converted into dimensionless values ​​within a unified range. After normalization, a set of dynamic weight coefficients are assigned to the task based on its position in the dynamic priority list, i.e., its priority. This allocation mechanism follows a core principle: the higher the priority of a task, the higher its sensitivity to delay and the relatively lower its sensitivity to energy consumption. Therefore, for high-ranking tasks, a larger weight is assigned to the predicted task delay and a smaller weight to the predicted energy consumption. Conversely, for low-ranking tasks, a smaller weight is assigned to delay and a larger weight to energy consumption to encourage energy conservation. Finally, the final comprehensive utility score is calculated by weighted summing of the normalized predicted task delay and predicted energy consumption. , in, Representing the The overall utility score of each alternative implementation plan; and These represent the dynamic weighting coefficients assigned to latency and energy consumption, respectively. They are dimensionless values ​​determined based on the task's position in the dynamic priority list, and their sum is typically 1 to maintain consistency in the scoring scale. and These are the dimensionless predicted task delay and predicted energy consumption, respectively, after normalization. The formula linearly combines these two performance metrics from different dimensions using weights linked to task priority, thereby generating a scalar value that can uniformly measure the merits of a solution.

[0045] For example, taking the "face recognition" task, firstly, the system needs to normalize the prediction performance metrics. Assume the system's latency normalization range is 100ms to 600ms, and the energy consumption normalization range is 0.05% to 5%. For Scheme 1, the prediction task latency is 400ms, and its... ; Energy consumption is projected to be 5%, For Option 2, the prediction task latency is 150ms. The projected energy consumption is 0.1%, For Option 3, the prediction task latency is 500ms. The predicted energy consumption is 0.05%, Because the "face recognition" task is a high-priority task, the system assigns it a dynamic weight coefficient that emphasizes latency. , Now, the system uses a formula to calculate the overall utility score for each option. Option one's score is... Option 2 scores as follows: Option 3 received a score of [score missing]. By comparison, Option 2 has the lowest overall utility score, therefore the system will choose this option. This method provides an adaptive solution to the multi-objective trade-off problem in task offloading decisions, directly related to task importance. This flexible trade-off mechanism makes system resource allocation more intelligent and refined, and can improve the energy efficiency and sustainability of the entire system while meeting key performance requirements.

[0046] S5. Control task execution according to the task unloading strategy, collect actual delay and energy consumption during execution, and generate feedback data; Optionally, the generated feedback data includes: The task unloading strategy is analyzed to determine the task execution location for each task; Based on the task execution location, control commands are sent to IoT devices or edge servers, and the actual latency and energy consumption during task execution are recorded to obtain a performance log; The performance logs are integrated and processed to generate feedback data.

[0047] Specifically, the process first requires parsing the task offloading strategy. The task offloading strategy is essentially a mapping table that specifies the final execution location assigned to each task. Each entry in the strategy is read one by one to determine the execution location for each task—whether it executes locally on the IoT device or is offloaded to a specific edge server. Based on the determined execution location, control commands containing task code or task identifiers are sent to the target execution node (IoT device or edge server) via the appropriate communication protocol to initiate the actual execution of the task. Simultaneously with sending the commands, a monitoring and recording mechanism is initiated. This mechanism records the entire lifecycle of the task from start to completion. By setting timestamps at key nodes in the task flow, such as task issuance time, start time, completion time, and result return time, the actual latency of task execution is calculated. Simultaneously, if the execution node supports energy consumption monitoring, the actual energy consumption during task execution is obtained by querying the device's power management interface or energy consumption sensors. These raw records, containing timestamps, actual latency, and actual energy consumption, are collected to form a preliminary performance log. Finally, the scattered performance logs from different tasks are integrated and processed, and the relevant records are correlated and structured to form a unified dataset containing the real performance metrics of all executed tasks, which is the final feedback data.

[0048] For example, based on the generated task offloading strategy, the system determines to offload the "face recognition" task T001 to the integrated gateway for execution. The system sends an execution command to the integrated gateway at time t1 and records t1. After completing the task, the integrated gateway returns the result at time t2. Upon receiving the result, the system calculates the actual delay as t2 minus t1, which is 160ms. Simultaneously, the system queries the integrated gateway's power management unit to determine that the energy consumption increment during the execution of task T001 is 0.12 joules, which is the actual energy consumption. To compare this with the predicted relative value, the system further queries the gateway's rated power and execution time, or directly obtains the energy consumption percentage from the power management unit, calculating that the actual energy consumption of this task is equivalent to 0.11% of the device's total energy consumption. The system records this set of performance data: {Task ID: T001, Task Type: FaceRecognition-HD, Execution Location: Integrated Gateway, Actual Latency: 160ms, Actual Energy Consumption: 0.12J, Actual Energy Consumption_Relative Value: 0.11%}. Once a batch of tasks is completed, all similar records are integrated into a structured table or database to form the batch's feedback data for subsequent optimization. This method, by establishing a closed-loop monitoring and recording mechanism, ensures that the actual effect of task offloading decisions can be accurately quantified and tracked, providing a data foundation for the adaptive optimization of the entire system.

[0049] S6. Generate an adjustment signal using the feedback data, and update the dynamic priority evaluation mechanism and the collaborative decision-making model based on the adjustment signal.

[0050] Optionally, updating the dynamic priority evaluation mechanism and the collaborative decision-making model based on the adjustment signal includes: The actual delay and actual energy consumption are extracted from the feedback data and compared with the predicted task delay and predicted energy consumption to generate a deviation value. Based on the magnitude and direction of the deviation value, an adjustment signal containing the adjustment direction and amplitude is generated; The weight rules of the dynamic priority evaluation mechanism and the multi-objective optimization parameters of the collaborative decision-making model are iteratively updated using the adjustment signal.

[0051] Specifically, the process first extracts the measured performance values—actual latency and actual energy consumption—from the feedback data for each completed task. Simultaneously, it retrieves the predicted values ​​used when generating the task offloading strategy from historical decision records, namely, the predicted task latency and predicted energy consumption. By subtracting the predicted values ​​from the actual values, two core deviation values ​​are calculated: latency deviation and energy consumption deviation, as shown in the following formula: in, Represents delay bias, Represents energy consumption deviation; and These are the actual delays obtained from the feedback data and the forecasting task delays used in decision-making, respectively. and These are the actual energy consumption and the predicted energy consumption, respectively. This set of deviation values ​​forms the core of the adjustment signal. The magnitude and sign of these two deviation values ​​directly reflect the accuracy of the prediction model and the gap between the decision and reality. Next, based on the magnitude and direction of these two deviation values, a structured adjustment signal containing the adjustment direction and magnitude is generated. For example, a large positive delay deviation indicates that the model's prediction is too optimistic, and the actual delay far exceeds expectations. The direction of the adjustment signal will be to increase the penalty for delay, and its magnitude will be proportional to the magnitude of the deviation. Finally, this adjustment signal is used to update the core model parameters iteratively. On one hand, it is used to update the weight rules in the dynamic priority evaluation mechanism. This is precisely the process of dynamically learning and adjusting the historical performance correction factors used to calculate the final priority score. For example, if a certain type of task repeatedly exhibits large delay deviations, the adjustment signal will increase the historical performance correction factor for that task type, thereby automatically assigning it a higher initial weight in future priority calculations. On the other hand, the adjustment signal is also used to update the multi-objective optimization parameters in the collaborative decision-making model. For example, if it is found that the actual latency of high-priority tasks is generally higher than expected, the adjustment signal will guide the system to adjust the dynamic weight coefficient allocation logic in the utility function, further increasing the weight of latency factors in future high-priority task decisions, i.e., adjustment. and The generation rules.

[0052] For example, the system processes "face recognition" task records from feedback data. Its measured performance value is an actual latency of 160ms, while the predicted task latency used in decision-making is 150ms. The system calculates a latency deviation of +10ms. Since this is a high-priority task, the system considers this deviation significant. Therefore, the system generates an adjustment signal indicating that higher attention needs to be given to the "FaceRecognition-HD" task type. This adjustment signal is used to update the dynamic priority evaluation mechanism, specifically by fine-tuning the historical performance correction factor for this task type from 1.2 to 1.22. Simultaneously, this adjustment signal also applies to the collaborative decision-making model. Because the actual latency of high-priority tasks is generally higher than expected, the adjustment signal guides the system to fine-tune the allocation logic of the dynamic weight coefficients in the utility function. For example, for tasks with the top 10% priority, their latency weights... One of the baseline parameters in the calculation formula is increased, so that latency will play a greater role in future decision-making for such tasks. This method establishes a closed-loop feedback and self-learning mechanism from execution to decision-making, enabling the dynamic priority assessment mechanism to more accurately reflect the true urgency of the task, and the collaborative decision-making model to more effectively balance latency and energy consumption. The system is therefore no longer a static, rule-based decision-making entity, but an intelligent system capable of learning from experience and dynamically adapting to environmental changes, improving the long-term accuracy of decisions and the overall robustness of the system.

[0053] Based on the same inventive concept, such as Figure 5 As shown, the present invention also provides an IoT edge computing task offloading system based on an integrated gateway, the system comprising: The data fusion module is used to acquire and fuse task attribute data, device status data, and network environment data of IoT devices to generate a unified information view. The priority evaluation module is used to calculate the priority of each task based on the unified information view and apply a dynamic priority evaluation mechanism to generate a dynamic priority list. The predictive data generation module is used to collect historical network load data and device resource usage data, apply time series analysis models to make predictions, and generate short-term predictive data. The collaborative decision-making module is used to combine the dynamic priority list and the short-term forecast data, and perform multi-objective optimization through the collaborative decision-making model to generate a task unloading strategy. The execution control module is used to control task execution according to the task offloading strategy, collect actual delays and energy consumption during the execution process, and generate feedback data. The feedback optimization module is used to generate adjustment signals using the feedback data, and to update the dynamic priority evaluation mechanism and the collaborative decision-making model based on the adjustment signals.

[0054] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0055] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for gateway-based Internet of Things (IoT) edge computing task offloading, the method comprising: The method comprises: Obtain task attribute data, device state data and network environment data of Internet of Things devices and fuse them to generate a unified information view; Based on the unified information view, apply a dynamic priority evaluation mechanism to calculate the priority of each task and generate a dynamic priority list; Collect historical network load data and device resource usage data, apply a time series analysis model for prediction, and generate short-term prediction data; Combine the dynamic priority list and the short-term prediction data, and use a collaborative decision-making model for multi-objective optimization to generate a task offloading strategy; Control task execution according to the task offloading strategy, and collect actual delay and energy consumption during execution to generate feedback data; Use the feedback data to generate an adjustment signal, and update the dynamic priority evaluation mechanism and the collaborative decision-making model according to the adjustment signal. 2.The integrated gateway-based task offloading method for edge computing in IoT according to claim 1, wherein, The generation of the unified information view comprises: Receive a task generation request from an Internet of Things device, perform compliance verification on the priority declaration field according to a predefined emergency level grading standard, and generate a verified task request; Extract task type descriptors and task size parameters from the verified task request to generate task attribute data; Collect current device power data and processor performance indicators from real-time monitoring data to generate device state data; Measure network channel transmission rate and response time parameters from real-time monitoring data to generate network environment data; Perform standardization conversion and data correlation mapping on the task attribute data, the device state data and the network environment data to generate a unified information view. 3.The integrated gateway-based IoT edge computing task offloading method of claim 1, wherein, The generation of the dynamic priority list comprises: Extract the task attribute data and the device state data from the unified information view; Perform weighted calculation on the task attribute data and the device state data to generate an initial priority score; Perform historical performance correction on the initial priority score to calculate a final priority score; Arrange all tasks in descending order according to the final priority score to generate a dynamic priority list. 4.The method of claim 3, wherein, The historical performance correction of the initial priority score comprises: Retrieve historical execution efficiency data associated with the current task type to calculate a historical performance correction factor; Multiply the initial priority score by the historical performance correction factor to obtain the final priority score.

5. The method of claim 1, wherein, Generating short-term prediction data comprises: Periodically collect network throughput samples and device CPU utilization samples from historical task execution records to build a multi-dimensional time series data set; Perform missing value filling and outlier smoothing on the multi-dimensional time series data set to generate a standardized time series data stream; Input the standardized time series data stream into a pre-trained time series analysis model to output a network load prediction sequence and a device resource prediction sequence containing multiple future time steps through a multi-step rolling prediction method; Integrate the network load prediction sequence and the device resource prediction sequence and add a prediction confidence score to generate short-term prediction data. 6.The method of claim 1, wherein, The generation of the task offloading strategy comprises: evaluating a plurality of alternative execution schemes for each task in the dynamic priority list and calculating a predicted task delay and a predicted energy consumption for each alternative execution scheme in combination with the short-term prediction data; applying a utility function for balancing performance to convert the predicted task delay and the predicted energy consumption into a comprehensive utility score; selecting an alternative execution scheme for each task according to the comprehensive utility score and combining to generate a task offloading strategy.

7. The method of claim 6, wherein the method further comprises: The converting the predicted task delay and the predicted energy consumption into a comprehensive utility score comprises: assigning a dynamic weight coefficient to the predicted task delay and the predicted energy consumption according to the position of the task in the dynamic priority list; performing a weighted sum of the predicted task delay and the predicted energy consumption by the dynamic weight coefficient to obtain a comprehensive utility score. 8.The integrated gateway-based task offloading method for edge computing in IoT of claim 1, wherein, The generating feedback data comprises: analyzing the task offloading strategy to determine a task execution location for each task; sending a control instruction to an Internet of Things device or an edge server according to the task execution location and recording an actual delay and an actual energy consumption in a task execution process to obtain a performance log; integrating the performance log to generate feedback data. 9.The integrated gateway-based task offloading method for edge computing in IoT of claim 1, wherein, The updating the dynamic priority evaluation mechanism and the collaborative decision model according to the adjustment signal comprises: extracting an actual delay and an actual energy consumption from the feedback data and comparing them with the predicted task delay and the predicted energy consumption to generate a deviation value; generating an adjustment signal containing an adjustment direction and an adjustment amplitude according to the size and direction of the deviation value; iteratively updating a weight rule of the dynamic priority evaluation mechanism and a multi-objective optimization parameter of the collaborative decision model using the adjustment signal.

10. A comprehensive gateway-based IoT edge computing task offloading system, applied to the comprehensive gateway-based IoT edge computing task offloading method of any one of claims 1-9, characterized in that, The system comprises: a data fusion module for obtaining and fusing task attribute data, device state data and network environment data of an Internet of Things device to generate a unified information view; a priority evaluation module for calculating a priority of each task based on the unified information view and generating a dynamic priority list by applying a dynamic priority evaluation mechanism; a prediction data generation module for collecting historical network load data and device resource usage data and generating short-term prediction data by applying a time series analysis model for prediction; a collaborative decision module for generating a task offloading strategy by performing multi-objective optimization through a collaborative decision model in combination with the dynamic priority list and the short-term prediction data; an execution control module for controlling task execution according to the task offloading strategy and collecting an actual delay and an energy consumption in an execution process to generate feedback data; a feedback optimization module for generating an adjustment signal using the feedback data and updating the dynamic priority evaluation mechanism and the collaborative decision model according to the adjustment signal.

Images