Industrial Internet of Things edge computing unloading method and system based on distributed deep neural network

Abstract

The invention discloses an industrial Internet of Things edge computing unloading method and system based on a distributed deep neural network, and the method comprises the following steps: firstly, introducing a feature matrix of an attention mechanism optimization task, transmitting the optimized task features to a multi-layer deep neural network DNN, and generating an unloading decision set in parallel; allocating resources for each task; after the to-be-unloaded task is preliminarily selected, computing resources of the high-priority task are preferentially guaranteed in the resource allocation stage according to the priority and deadline of the task. For tasks with relatively large calculation amount or relatively high priorities, decision storage is performed by using a cache strategy to optimize unloading and resource allocation processes of subsequent similar tasks, so that efficient and low-delay task unloading in a large-scale and multi-task environment is realized. Meanwhile, a distributed cache updating mechanism is utilized in each unloading time slot, the unloading decision precision is continuously optimized, and the unloading efficiency and the overall stability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things and mobile edge computing, and in particular to an industrial Internet of Things edge computing offloading method and system based on a distributed deep neural network. Background Art

[0002] With the rapid development of the Internet of Things, the Industrial Internet of Things (IIoT) has become an important part of modern industrial automation, smart manufacturing and intelligent control. The Industrial Internet of Things connects devices, sensors and systems to achieve real-time collection, transmission and analysis of industrial data to optimize production processes and improve efficiency. However, since Industrial Internet of Things devices are usually deployed in resource-constrained environments, the computing power and power supply of a single device are insufficient, resulting in great challenges in processing high-frequency, low-latency and high-data volume tasks. In addition, Industrial Internet of Things devices are widely distributed, the task types are complex and diverse, and the delay and energy consumption problems are particularly prominent. Therefore, there is an urgent need for an efficient computing architecture to meet the real-time and high-efficiency requirements of the Industrial Internet of Things.

[0003] Mobile edge computing (MEC) is an emerging computing model that places computing and storage resources at the edge of the network, which can significantly reduce data processing latency and network bandwidth pressure. Through mobile edge computing, industrial IoT devices can offload computing-intensive and latency-sensitive tasks to edge servers or cloud resources that are closer to them, thereby reducing the computing burden and power consumption of terminal devices. Mobile edge computing has obvious advantages in industrial IoT applications, but it also brings about the complexity of task offloading decisions and resource allocation. In particular, with the increase in the number of industrial IoT terminal devices and the increase in task complexity, existing computing offloading algorithms often cannot effectively optimize system performance.

[0004] Computational offloading technology plays a core role in mobile edge computing. By offloading computing tasks from resource-constrained terminal devices to edge servers or remote clouds, it not only optimizes system performance but also reduces device energy consumption. The formulation of computational offloading strategies requires comprehensive consideration of multiple factors such as task complexity, latency requirements, bandwidth resources, and device status. However, existing computational offloading algorithms have obvious shortcomings when dealing with the high complexity and multi-task environment in IIoT. For example, traditional offloading algorithms are usually based on static strategies and fail to dynamically adapt to changes in tasks and network environments in real time; although some algorithms can allocate tasks in a multi-device environment, the efficiency and accuracy of the algorithms are greatly reduced in the face of large-scale equipment and complex task scenarios, resulting in increased task response delays or waste of resources.

[0005] Therefore, it is particularly important to develop new offloading algorithms to adapt to the complexity of devices and tasks in the industrial Internet of Things environment. These algorithms not only need to consider the balance between latency and energy consumption, but also need to be flexible to adapt to changing network conditions and device status. The necessity of research in this area is obvious because it will directly affect the application efficiency and sustainable development of the industrial Internet of Things. Summary of the invention

[0006] 1. Technical problem to be solved by the invention

[0007] In order to solve the problems of inefficient resource allocation, high system latency, and insufficient energy consumption optimization faced in the process of computing offloading in the Industrial Internet of Things (IIoT), while taking into account the dynamic adaptability of the offloading strategy and the system stability, the present invention provides an Industrial Internet of Things edge computing offloading method and system based on a distributed deep neural network. The present invention combines a distributed deep neural network with a multi-task resource allocation strategy to efficiently implement resource allocation and optimize system performance in large-scale equipment and complex task environments, ensuring that computing-intensive tasks are efficiently offloaded and processed with low latency and low energy consumption.

[0008] 2. Technical solution

[0009] In order to achieve the above object, the technical solution provided by the present invention is:

[0010] The present invention provides an industrial Internet of Things edge computing offloading method based on a distributed deep neural network, comprising the following steps:

[0011] Step 1: Collect the computing task data of each terminal device in the industrial Internet of Things environment, save the task data in the form of a matrix, and use it as a data input set for offloading decisions;

[0012] Step 2: Calculate the priority of each task of each device according to the task deadline and task waiting time;

[0013] Step 3: Before the task data is input into the deep neural network DNN, the task features are dynamically weighted optimized through the attention mechanism, and the optimized task data is input into multiple parallel deep neural networks DNN, and each deep neural network DNN is used to independently generate task offloading decisions;

[0014] Step 4: Based on the offloading decision set generated in step 3, calculate and optimize the bandwidth and computing resource allocation of each device task to obtain a candidate offloading strategy;

[0015] Step 5: Calculate the system utility of the candidate unloading strategies obtained in step 4, select the unloading strategy with the lowest system utility value as the optimal strategy, and store the optimal unloading strategy in the cache as training data for the deep neural network for subsequent iterative updates and model optimization.

[0016] Furthermore, the task data described in step one includes task size, task computation amount and task deadline.

[0017] Furthermore, the priority calculation formula of step 2 is as follows:

[0018]

[0019] In the formula, P represents the priority of each task, T deadline Indicates the time when the task should be completed, T wait Indicates the time a task waits to be processed in the system; the calculation result of P is between 0 and 1, and the smaller the P value, the higher the task priority.

[0020] Furthermore, in step three, an attention mechanism module is introduced to dynamically weighted optimize task features. The offloading strategy of the optimized task generated by the deep neural network DNN is a floating-point numerical vector, which is converted into a binary decision vector [0,1]. 0 means that the task is not offloaded to the edge server and is processed locally on the device; 1 means that the task will be offloaded to the edge server.

[0021] Furthermore, step four constructs a resource allocation model, the goal of which is to maximize the utility of the system, that is, to minimize the weighted sum of task delay and energy consumption on the basis of satisfying task priority and offloading requirements; the offloading decision is passed as input to the resource allocation model for resource allocation, and after the resource allocation is completed, the allocation result is adjusted to ensure that high-priority tasks obtain sufficient computing resources.

[0022] Furthermore, the adjustment of resource allocation includes dynamic bandwidth allocation adjustment and computing resource allocation adjustment; the dynamic bandwidth allocation obtains the actual bandwidth allocation amount based on the dynamic allocation formula of bandwidth resources according to the priority and bandwidth resources of the task; the computing resource allocation obtains the actual computing resource allocation amount based on the allocation formula of computing resources according to the priority and unloading decision of the task.

[0023] Furthermore, in step 5, the optimal strategy selection is realized through the system utility calculation function, and the calculation formula is:

[0024] U=αD+βE

[0025] Among them, U is the system utility value, D is the total delay of the task, E is the energy consumption of the task, and α and β are weight coefficients used to adjust the impact of delay and energy consumption in the utility function.

[0026] An industrial Internet of Things edge computing offloading system based on a distributed deep neural network of the present invention executes the edge computing offloading method, specifically comprising:

[0027] Task initialization module: used to collect and process task data of industrial IoT devices and generate a standardized data matrix;

[0028] Priority calculation module: used to calculate the priority of tasks based on task deadlines and waiting times;

[0029] Distributed offloading decision module: Introduces the attention mechanism, dynamically weights and optimizes task features. The optimized tasks are based on multiple parallel deep neural network (DNN) models to generate a variety of offloading decisions.

[0030] Resource allocation module: dynamically allocates device bandwidth and edge server resources based on offloading decisions;

[0031] Optimal strategy selection module: performs system utility calculations for different unloading strategies, selects the current optimal strategy to execute the task, and writes the optimal strategy results into the cache for decision optimization of future tasks.

[0032] 3. Beneficial effects

[0033] Compared with the existing known technologies, the technical solution provided by the present invention has the following significant effects:

[0034] The present invention proposes an industrial Internet of Things (IIoT) computing offloading method based on a distributed deep neural network. Aiming at the characteristics of the IIoT environment with many devices, complex tasks and delay sensitivity, the method designs the following offloading scheme: first, an attention mechanism is introduced to optimize the feature matrix of the task, and the optimized task features are transmitted to a multi-layer deep neural network (DNN) to generate an offloading decision set in parallel, and allocate resources for each task; after the preliminarily selected tasks to be offloaded, the computing resources of high-priority tasks are prioritized in the resource allocation stage according to the task priority and deadline. For tasks with large computational volume or high priority, the present invention uses a cache strategy for decision storage to optimize the unloading and resource allocation process of subsequent similar tasks, thereby realizing efficient and low-latency task offloading in a large-scale, multi-task environment. At the same time, the present invention uses a distributed cache update mechanism in each offloading time slot to continuously optimize the offloading decision accuracy, thereby improving the offloading efficiency and the overall stability of the system.

[0035] (1) In the IIoT scenario, multiple devices and complex tasks coexist. Aiming at this feature, the present invention designs a distributed DNN model to generate unloading decisions in parallel. By comprehensively considering the priority and deadline of the task, resources are allocated to key tasks first. The distributed DNN can generate unloading decisions in parallel on different edge servers, greatly improving the efficiency of the unloading process, ensuring the rational use of system resources in multi-tasking scenarios, and avoiding resource waste and increased unloading delays.

[0036] (2) In the IIoT environment, real-time performance and energy consumption optimization are particularly important. To this end, the present invention introduces task priority and deadline filtering mechanisms in the generation of offloading decisions. Tasks with high real-time performance are given priority in allocating bandwidth resources, and tasks that are sensitive to energy consumption select low-power edge nodes, thereby ensuring the overall response speed of the system and reducing device energy consumption. Through the cache structure, the present invention continuously records and optimizes offloading decisions, so that subsequent similar tasks can be quickly scheduled, improving overall computing efficiency and extending the operating life of the device.

[0037] (3) In the face of IIoT application scenarios with diverse tasks and large amounts of data, the present invention avoids the overload accumulation of complex tasks in the process of offloading strategy generation and resource allocation optimization. The task features are dynamically weighted through the attention mechanism, making the offloading decision generation process more adaptable and accurate. Compared with traditional DNN, the attention mechanism can effectively capture key task features and significantly improve the system's responsiveness to complex task scenarios. The parallel processing and cache reuse mechanism of distributed DNN ensures the dynamic adaptability of the offloading strategy, improves the accuracy and response speed of offloading decisions, and ensures the smooth and reliable operation of large-scale computing tasks in the IIoT system.

[0038] Through the technical solution proposed in the present invention, the distributed deep learning method realizes efficient and low-latency computing offloading in large-scale IIoT scenarios, effectively optimizes resource utilization and task response performance, thereby significantly improving the stability and resource management efficiency of the industrial Internet of Things system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of an industrial Internet of Things edge computing offloading method based on a distributed deep neural network of the present invention.

[0040] Figure 2 It is a structural flow chart of the deep neural network in the present invention. DETAILED DESCRIPTION

[0041] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments.

[0042] Example 1

[0043] like Figure 1 As shown, an industrial Internet of Things edge computing offloading method based on a distributed deep neural network in this embodiment includes the following processes:

[0044] Step 1: Initialize device task data: The task initialization module collects the computing task data of each terminal device in the IIoT environment, including parameter information such as task size, computing amount, and deadline. The initialization of device task data includes the normalization of task data format, ensuring that all task data are input in a standardized format such as size, computing amount, and deadline, so as to facilitate efficient processing of the offloading model. The task data is stored in the form of a standard matrix to provide input data for subsequent offloading decisions.

[0045] Step 2: Task priority calculation: In the priority calculation module, the task priority of each device is calculated according to the task deadline and task waiting time. deadline Indicates the time when the task should be completed, T wait Indicates the time a task waits in the system to be processed, using the formula:

[0046]

[0047] The calculation result is between 0 and 1. The smaller the P value, the higher the task priority. When the deadline and waiting time of the task are equal, the priority P is 0, indicating that the task priority is the highest. When the task just starts waiting (the waiting time is 0), the priority P is 1, indicating that the task priority is the lowest.

[0048] In this way, the importance of different tasks is ranked to ensure that tasks with higher priority are given priority and resource allocation for critical tasks is guaranteed.

[0049] Step 3: Distributed DNN generates offloading decisions: First, in order to improve the accuracy of offloading decisions, before the task features are input into DNN, a new attention mechanism module is added to dynamically weight and optimize the task features. The implementation process is as follows:

[0050] Task feature input: Input the task feature matrix (S), which contains information such as task size, computational effort, and deadline.

[0051] Feature decomposition: Generate query vector Q, key vector K, and value vector V through linear transformation:

[0052] Q=W q S,K=W k S,V=W v ·S

[0053] Among them, W q , W k , W v is the trainable weight matrix.

[0054] Calculate the attention weight: Use the dot product to calculate the correlation between Q and K, and normalize the generated weights through the Softmax function:

[0055]

[0056] Among them, d k is the dimension of the key vector, used for normalization.

[0057] Optimize feature generation: Use attention weights to weight the value vector and generate an optimized feature matrix:

[0058] Output=Attention(Q,K)·V

[0059] DNN generates offloading decisions: The optimized task feature Output is input into multiple parallel deep neural networks (DNNs), and each deep neural network DNN is used to independently generate task offloading decisions. The number of parallel DNNs is set to k, and k candidate offloading decision sets are obtained.

[0060] Among them, the specific process of the parallel deep neural network DNN using task data to generate an offloading strategy is as follows:

[0061] Each deep neural network DNN preprocesses the input task data S, calculates layer by layer through the weight matrix W and bias term b of each layer, and finally generates an independent offloading decision X. The formula is:

[0062] X=f(W,S)+b,

[0063] Where f represents the activation function (such as ReLU or Sigmoid), W is the weight matrix, and b is the bias term.

[0064] This calculation method extracts task features layer by layer. By adjusting weights and bias items, the deep neural network DNN can dynamically adapt to changes in task features and generate appropriate offloading decisions.

[0065] It is worth noting that, combined with Figure 2In a neural network, each neuron node receives the output value of the previous layer of neurons as the input value of the neuron, and passes the input value to the next layer. The input layer neuron node will directly pass the input attribute value to the next layer. In a multi-layer neural network, there is a functional relationship between the output of the upper layer node and the input of the lower layer node. This function is the activation function. The commonly used activation function is the nonlinear activation function. The use of nonlinear activation functions is to increase the nonlinear factors of the neural network model in order to make the network more powerful and increase its capabilities. The activation function f introduces nonlinear processing capabilities, enabling DNN to perform in-depth analysis of complex task features. The choice of activation function depends on the distribution of task features and the DNN model design. Common activation functions include ReLU (f(x) = max(0, x), etc., which are respectively applicable to sparse representation and probabilistic prediction.

[0066] The output of the deep neural network DNN layer-by-layer calculation contains different representations of task features. Through layer-by-layer accumulation and conversion, it finally generates an offloading strategy that comprehensively considers the task requirements.

[0067] After layer-by-layer calculation of the DNN model, the final generated offloading strategy X is a floating-point numerical vector, which is converted into a binary decision vector [0,1] to clarify whether each task needs to be offloaded. 0 means no offloading, and 1 means offloading, which is used to indicate whether the tasks of each device are offloaded to the edge server.

[0068] The specific process of this step is to set the decision threshold, set the value greater than or equal to the threshold in the offloading strategy X generated by the DNN model to 1, indicating that the task will be offloaded to the edge server; set the value less than the threshold to 0, indicating that the task will be processed locally on the device. Through the generated binary offloading decision vector, the present invention realizes the offloading judgment of each device task in a multi-task environment, effectively improving the accuracy of offloading decisions and the efficiency of task allocation.

[0069] Step 4: Resource allocation and optimization: For the execution of offloading decisions, the resource allocation module will dynamically optimize the allocation of computing resources and network bandwidth of edge servers based on the priority of device tasks and the offloading decision set generated by the DNN model in step 3. By ensuring that high-priority tasks receive priority resource support, it aims to effectively reduce task latency, improve system response speed, and optimize overall energy consumption. Specifically:

[0070] (1) Construction of resource allocation model: The goal of the resource allocation model is to maximize the utility of the system, that is, to minimize the weighted sum of task delay and energy consumption on the basis of satisfying task priority and offloading requirements. Assume that the system has the following resource parameters: B and C, where B represents the bandwidth of the edge node (i.e., task transmission resources) and C represents the computing resources of the edge server (i.e., resources for processing tasks).

[0071] The resource allocation model aims to optimize based on the current device bandwidth and server computing resources. Assuming the bandwidth is B and the server computing resources are C, the resource allocation model can be expressed as:

[0072] maximize U=f(B,C)

[0073] Among them, U represents the overall utility of the system, and f is a function that represents the relationship between device and server resources.

[0074] The resource allocation objective can be defined as the following utility function:

[0075]

[0076] Where: D i represents the processing delay of task i, E i represents the energy consumption of task i, α i and β i is a weight parameter used to adjust the relative priority between delay and energy consumption.

[0077] The optimization goal of the system is to minimize U, that is:

[0078]

[0079] After the resource allocation model is built, the offloading decision needs to be passed as input to the resource allocation model for resource allocation. After the resource allocation is completed, the allocation results need to be adjusted to ensure that high-priority tasks obtain sufficient computing resources. Specifically:

[0080] Dynamic bandwidth allocation: According to the priority of the task P i , where i represents the index of the task, and bandwidth resource B, allocates more bandwidth to high priority tasks to ensure low latency in the transmission process. The dynamic allocation formula of bandwidth resources is as follows:

[0081] B′ i =B i +γP i

[0082] Where: B′ i is the actual bandwidth allocation for task i, B iis the initial bandwidth resource allocation value, γ is the bandwidth adjustment factor, which is used to control the priority P i Impact on bandwidth allocation.

[0083] Computing resource allocation: Similarly, the computing resource allocation of edge servers is adjusted based on task priority and offloading decisions. The computing resource allocation formula is:

[0084]

[0085] Where: C′ i is the actual computing resource allocation for task i, C i is the initial computing resource allocation value, is the computing resource adjustment factor, controlling the task priority P i The weight in resource allocation, X i Whether task i is offloaded to the edge server.

[0086] Step 5, optimal strategy selection and update: In the optimal strategy selection module, the candidate offloading strategies obtained in step 4 are calculated for system utility, including task delay and energy consumption, and the offloading strategy with the lowest system utility value is selected as the optimal strategy. The optimal offloading strategy is stored in the cache as training data for the deep neural network for subsequent iterative updates and model optimization, so that the deep neural network can automatically update the decision model in subsequent task offloading.

[0087] The optimal strategy selection is realized through the system utility calculation function, and the calculation formula is:

[0088] U=αD+βE

[0089] Among them, U is the system utility value, which represents the comprehensive score of the strategy. The lower the utility value, the better the strategy. D is the total delay of the task, which refers to the total time from unloading to completion of the task. E is the energy consumption of the task, which refers to the total energy consumed during the execution of the task. The α and β weight coefficients are used to adjust the impact of delay and energy consumption in the utility function. Finally, the unloading decision with the smallest system utility is selected as the output.

[0090] Example 2

[0091] A computing offloading system based on a distributed deep neural network in this embodiment includes:

[0092] Task initialization module: used to collect and process task data of IIoT devices and generate a standardized data matrix.

[0093] Priority calculation module: used to calculate the priority of a task based on its deadline and waiting time.

[0094] Distributed offloading decision module: A new attention mechanism module is added to dynamically weight and optimize the task data features, and then generate multiple offloading decisions based on k parallel deep neural network DNN models.

[0095] Resource allocation module: Dynamically allocates device bandwidth and edge server resources based on offloading decisions.

[0096] Optimal strategy selection module: performs system utility calculations for different unloading strategies, selects the current optimal strategy to execute the task, and writes the optimal strategy results into the cache for decision optimization of future tasks.

[0097] The present invention proposes an industrial Internet of Things edge computing offloading method and system based on distributed deep neural network. Aiming at the characteristics of diverse equipment, complex data and high demand for task offloading in industrial Internet of Things scenarios, the coordinated mechanism of distributed deep neural network and edge computing node is designed to optimize the task offloading decision and computing resource allocation strategy. The system distributes large-scale computing tasks to each edge node by collaboratively performing deep learning calculations on edge nodes and the cloud, effectively reducing the computing load of the central server and improving the real-time performance of data processing; and through dynamic strategy adjustment based on distributed learning, efficient task offloading is achieved, and resource utilization efficiency in multi-device and multi-task scenarios is optimized. The method greatly improves the response speed of complex tasks in industrial Internet of Things scenarios, ensures the load balancing of edge devices and cloud computing resources, and reduces the risk of response delay and waste of system resources caused by task accumulation, thereby ensuring the stability and reliability of high-concurrency and low-latency tasks in industrial Internet of Things environments.

[0098] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. An industrial Internet of Things edge computing offloading method based on distributed deep neural network, characterized in that: The following steps are involved: Step 1: Collect the computing task data of each terminal device in the industrial Internet of Things environment, save the task data in the form of a matrix, and use it as a data input set for offloading decisions; Step 2: Calculate the priority of each task of each device according to the task deadline and task waiting time; Step 3: Before the task data is input into the deep neural network DNN, the task features are dynamically weighted optimized through the attention mechanism, and the optimized task data is input into multiple parallel deep neural networks DNN, and each deep neural network DNN is used to independently generate task offloading decisions; Step 4: Based on the offloading decision set generated in step 3, calculate and optimize the bandwidth and computing resource allocation of each device task to obtain a candidate offloading strategy; Step 5: Calculate the system utility of the candidate unloading strategies obtained in step 4, select the unloading strategy with the lowest system utility value as the optimal strategy, and store the optimal unloading strategy in the cache as training data for the deep neural network for subsequent iterative updates and model optimization.

2. According to claim 1, an industrial Internet of Things edge computing offloading method based on a distributed deep neural network is characterized in that: The task data described in step 1 includes task size, task computation amount and task deadline.

3. According to claim 1, an industrial Internet of Things edge computing offloading method based on a distributed deep neural network is characterized in that: The priority calculation formula for step 2 is as follows: In the formula, P represents the priority of each task, T deadline Indicates the time when the task should be completed, T wait Indicates the time a task waits to be processed in the system; the calculation result of P is between 0 and 1, and the smaller the P value, the higher the task priority.

4. According to claim 1, an industrial Internet of Things edge computing offloading method based on a distributed deep neural network is characterized in that: In step three, a new attention mechanism module is added to dynamically weighted optimize task features. The offloading strategy of the optimized task generated by the deep neural network DNN is a floating-point numerical vector, which is converted into a binary decision vector [0,1]. 0 means that the task is not offloaded to the edge server and is processed locally on the device; 1 means that the task will be offloaded to the edge server.

5. According to claim 1, an industrial Internet of Things edge computing offloading method based on a distributed deep neural network is characterized in that: Step 4: Build a resource allocation model. The goal of this resource allocation model is to maximize the utility of the system, that is, to minimize the weighted sum of task delay and energy consumption on the basis of satisfying task priority and offloading requirements. Pass the offloading decision as input to the resource allocation model for resource allocation. After the resource allocation is completed, adjust the allocation result to ensure that high-priority tasks obtain sufficient computing resources.

6. The industrial Internet of Things edge computing offloading method based on a distributed deep neural network according to claim 5 is characterized in that: The resource allocation adjustment includes dynamic bandwidth allocation adjustment and computing resource allocation adjustment; the dynamic bandwidth allocation obtains the actual bandwidth allocation amount based on the dynamic allocation formula of bandwidth resources according to the priority of the task and the bandwidth resources; The computing resource allocation is based on the priority and unloading decision of the task, and the actual computing resource allocation amount is obtained based on the computing resource allocation formula.

7. The industrial Internet of Things edge computing offloading method based on a distributed deep neural network according to claim 1 is characterized in that: In step 5, the optimal strategy selection is realized through the system utility calculation function, and the calculation formula is: U=αD+βE Among them, U is the system utility value, D is the total delay of the task, E is the energy consumption of the task, and α and β are weight coefficients used to adjust the impact of delay and energy consumption in the utility function.

8. An industrial Internet of Things edge computing offloading system based on distributed deep neural network, characterized by: Executing the edge computing offloading method according to any one of claims 1 to 7, specifically comprising: Task initialization module: used to collect and process task data of industrial IoT devices and generate a standardized data matrix; Priority calculation module: used to calculate the priority of tasks based on task deadlines and waiting times; Distributed offloading decision module: A new attention mechanism module is added to dynamically weight and optimize task data features, and then generate multiple offloading decisions based on multiple parallel deep neural network DNN models; Resource allocation module: dynamically allocates device bandwidth and edge server resources based on offloading decisions; Optimal strategy selection module: performs system utility calculations for different unloading strategies, selects the current optimal strategy to execute the task, and writes the optimal strategy results into the cache for decision optimization of future tasks.

Images