Method for realizing task unloading of industrial internet-of-things system based on width learning neural network
Through a width-learning neural network-based method, the problem of complexity of task offload decisions in the industrial Internet of Things is solved, the decision accuracy and response speed are improved, and the optimization and adaptability of system resources are achieved.
Patent Information
- Application Number
- CN202411290302.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2024-12-03
AI Technical Summary
In the industrial Internet of Things, as the number of devices increases, the cloud transmission of sensor data leads to increased bandwidth costs and network congestion, and the geographical location differences of equipment make task offload decisions complicated, and it is difficult for the existing technology to effectively consider device heterogeneity and network parameters. , resulting in low decision-making efficiency.
Using a width learning neural network-based approach, by collecting and labeling data, performing data normalization, and using a decision tree to extract features, training a width learning model to make task offload decisions, and optimizing system resource configuration and service quality.
Improves the accuracy of task offload decisions and system response speed, allows incremental updates to adapt to environmental changes, reduces the time cost of model adjustment, and improves system adaptability and performance.
Smart Images

Figure CN119065744A_ABST
Abstract
Description
Technical Field
[0001] The present invention mainly relates to the field of industrial Internet of Things task scheduling; in particular, to a method for unloading tasks of an industrial Internet of Things system based on a width learning neural network. Background Art
[0002] The Industrial Internet of Things supports the integration of heterogeneous industrial devices through wired or wireless communication networks, thereby realizing the collection of sensor data and decision optimization based on data analysis. As the number of Industrial Internet of Things devices grows, transmitting all sensor data to the cloud platform will lead to increased bandwidth costs and core network congestion. Deploying edge computing servers near the data source can provide local computing services for tasks, thereby reducing network bandwidth costs and traffic congestion and improving decision efficiency. Due to the geographical differences of devices, a device may have multiple edge servers as potential candidates for task offloading. Considering the benefits of offloading tasks to different candidates becomes a key factor in task offloading decisions. In order to achieve the optimal task offloading decision, it is necessary to comprehensively consider the heterogeneity of devices and the diversity of network parameters, and design corresponding algorithms. Machine learning is expected to make high-profit task offloading decisions in the complex network of the Industrial Internet of Things by extracting valuable information from large data sets and building models. Summary of the invention
[0003] In view of the difficulties of the existing technology, the present invention comprehensively considers the relevant parameters such as the location, computing power, channel status, etc. of devices and edge servers in the complex system of the Industrial Internet of Things, and proposes an Industrial Internet of Things task offloading solution based on a width learning model. The solution collects task and system data from the Industrial Internet of Things system, labels the data using a solver, and performs data normalization operations. The width learning model is trained using labeled data, and the feature weights calculated based on a decision tree are used to enhance the feature extraction capability of the model. The trained width learning model is used to make task offloading decisions, so that the system resource allocation is more reasonable and the service quality is significantly improved.
[0004] In order to solve the problems of the prior art, the present invention adopts the following technical solutions:
[0005] A method for offloading tasks of an industrial Internet of Things system based on a width learning neural network comprises the following steps:
[0006] Step 1: Build an initial task dataset by collecting information about servers, devices, and tasks in the IIoT Each row in the initial task data set corresponds to a task, and each column represents a specific task feature; the task feature includes information about the server, device, and task;
[0007] Step 2: Classify the tasks in the initial task dataset according to the task creation time to obtain task samples X in different time slotsm ;
[0008] Step 3: Use the Gurobi solver to solve the task samples X for different time slots m Select the edge server label for optimal task offloading and obtain task samples with edge server labels;
[0009] Step 4: Perform a normalization operation on the task samples with edge server labels to obtain a normalized and labeled dataset;
[0010]
[0011] Where: m is the row index, n is the column index, and Respectively represent the maximum and minimum values of each column;
[0012] Step 5: Extract important features of task samples from the normalized and labeled data set through the decision tree method, weight the task data samples, and obtain weighted input task data samples;
[0013] Step: 6: Train and update the task offloading neural network using weighted input task data samples.
[0014] Furthermore, the important feature process of extracting the task samples from the normalized and labeled data set by the decision tree method in step 5 includes the following steps:
[0015] Step 501: Based on the initial task data set The task characteristics S of the samples in n Calculate each task feature S according to the following formula n Gini index
[0016]
[0017] Where: z is the number of branches into which the dataset is divided, S n,l YesS n The lth sub-branch of ;
[0018] Step 502: Based on each task feature S n The maximum and minimum values of the Gini index are calculated according to the following formula to calculate the task feature importance:
[0019]
[0020] Among them: The maximum and minimum values of the Gini index of task characteristics are and
[0021] Step 503: Using the input task sample X m and feature importance Constructing weighted task data samples Generate training datasets for task offloading neural networks.
[0022] Furthermore, in step 6, the training and updating of the task offloading neural network is performed using the weighted input task data samples, including the following steps:
[0023] Step 601: Calculate the data weighted value using the following formula: The value Z mapped to the task feature node i :
[0024]
[0025] Where: Φ i (·) is the mapping function, is the random weight and is the bias term;
[0026] Step 602: Calculate the characteristic node value Z by the following formula: i The value H mapped to the enhanced node j :
[0027]
[0028] Where: j (·) is the mapping function;
[0029] Step 603: The combined matrix of the relationship between the data label Y and the feature nodes and the enhanced nodes is expressed by the following formula:
[0030]
[0031] Step 604: Solve the combination matrix [Z i ∣H j ] of the pseudo-inverse matrix [Z k ∣H u ] + ;
[0032] Step 605: According to the pseudo-inverse matrix [Z k ∣H u ] + Update the task offloading neural network weights W * =[Z k ∣H u ] + Y;
[0033] Step 606: By judging whether the task offloading neural network weight reaches the expected accuracy or new task data is continuously generated, if the conditions are met, then end, otherwise proceed to the next step;
[0034] Step 607: Calculate the incremental features and enhanced nodes caused by the introduction of new task data, and add the incremental nodes to the feature node and enhanced node matrix of the initial task data set;
[0035] Step 608: Calculate the weight update caused by the increment, obtain the updated weight, and save it to complete
[0036] Task offloading of neural network updates.
[0037] Beneficial Effects
[0038] Compared with the traditional technical solution, the beneficial effects brought by the present invention are:
[0039] 1. The solution proposed in the present invention has an efficient network structure, which is helpful for rapid model training, reduces the large amount of training time usually required by machine learning technology, and improves the real-time performance and response speed of the system. At the same time, the present invention adopts a feature extraction method based on decision tree theory to ensure that the model can consider influential features that may be ignored by traditional optimization schemes, thereby improving the accuracy of task offloading decisions.
[0040] 2. The solution proposed in this invention allows the system to be incrementally updated when integrating new equipment or generating new task data, thereby maintaining adaptability to environmental changes. If the accuracy of the model does not meet the requirements, it can be improved by adding feature nodes and enhanced nodes in the model. This feature ensures that the system can adapt quickly when faced with new inputs, reducing the time cost of model adjustment, while allowing the model to be flexibly adjusted according to actual needs, improving the adaptability and performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 This is a data processing flow chart of an industrial Internet of Things task offloading method based on a width learning system of the present invention;
[0042] Figure 2 This is an update flow chart of an industrial Internet of Things task offloading method based on a width learning system of the present invention. DETAILED DESCRIPTION
[0043] The following is combined with Figure 1 -Attached Figure 2 The present invention is described as follows:
[0044] like Figure 1As shown, the present invention provides a method for realizing industrial Internet of Things task unloading based on a width learning neural network, including: industrial Internet of Things task data preprocessing, industrial Internet of Things training and industrial Internet of Things updating, wherein:
[0045] The task data preprocessing includes the following steps 1-2:
[0046] Step 1: Collect task information in the environment from the industrial IoT system, synthesize it into an original data set, and perform preprocessing operations such as labeling and normalization of the data set. The specific process is as follows:
[0047] [1.1] Collect environmental information to build an initial task dataset for the Industrial Internet of Things. Each row in the initial task dataset corresponds to a task, and each column represents a specific task feature. The task feature contains information about the server, device, and task. Tasks are classified into different time slots based on the time when they were created. The time slot division is consistent with the Industrial Internet of Things system.
[0048] 【1.2】Perform label processing on each task data in the task data set, that is, label the data set samples. m The task in each time slot is divided into time slots, and the tasks in each time slot are unloaded to the edge server in batches. The unloading decision selects the best edge server for each task. The unloading decision for each time slot is to minimize the sum of the total processing time of all tasks. This optimization problem is formally defined, and the Gurobi optimization solver is used to find the optimal solution. The corresponding solution result Y m Used as task sample X m Labels, where: Label Y m Indicates the edge server number.
[0049] [1.3] Data normalization is a key step for neural networks to effectively extract task data features. The present invention applies min-max normalization to each column in the task data set.
[0050] Let m be the row index and n be the column index, and Represents the maximum and minimum values of each column respectively, then element a in the data set mn The normalized value of is calculated as:
[0051]
[0052] Step 2: Calculate the importance of each task feature in the data set and obtain the important feature matrix. Specifically include:
[0053] 【2.1】For the dataset The task features of the samples in the task are calculated for each task feature S n Gini index To measure the importance of different task features. Assume that the branch is from the task feature S n Initially, z is the number of branches into which the dataset is divided, S n,l YesS n The task feature S is calculated using the following formula: n The Gini index is:
[0054]
[0055] 【2.2】Use this formula to calculate S n,l The Gini index of the decision tree is obtained until the last layer. Assume that the last layer task feature is S n,l , p(x l ) is based on the feature S n,l The proportion of classified samples to the total samples, and the Gini index calculation formula is:
[0056]
[0057] 【2.3】Assume that the maximum and minimum values of the Gini index of all task features are and The importance of task feature Sn The calculation is as follows:
[0058]
[0059] [2.4] Obtain the feature importance matrix and calculate the weighted value of the input data By task sample X m It is obtained by multiplying it with the feature importance matrix.
[0060] like Figure 2 As shown, the task offloading neural network training and updating includes the following steps 1 to 2:
[0061] Step 1: Use the weighted input data samples to complete the model training process, which includes:
[0062] 【1.1】Define the mapping function Φ i (·), random weight and the bias term The data weight value is calculated using the following formula By the mapping function Φ i (·) is mapped to the characteristic node value Z i :
[0063]
[0064] 【1.2】Define the mapping function ξ j (·), random weight and the bias term The characteristic node value Z is calculated using the following formula: i By mapping function ξ j (·) is mapped to the enhanced node value H j :
[0065]
[0066] [1.3] Combine the feature nodes and the enhancement nodes into a combination matrix and define the problem. Given Y and Z i , Φ i (·),ξ j Under the premise of , the following formula is used to express the relationship between data label Y and feature nodes and enhanced nodes. The goal is to solve the weight W * :
[0067]
[0068] 【1.4】Use ridge regression to solve the model [Z i ∣H j ] of the pseudo-inverse matrix [Z k ∣H u ] + , calculate the trained weight W * =[Z k ∣H u ] + Y, complete the task offloading model training process.
[0069] Step 2: If the trained model cannot achieve the expected accuracy or new data continues to be generated, it is necessary to incrementally update the trained model, including:
[0070] [2.1] Calculate the incremental features and enhanced nodes caused by the introduction of new data, and add the incremental nodes to the feature node and enhanced node matrix of the initial matrix. Then, calculate the weight update caused by the increment, obtain the updated weight, and save it to complete the update process of the training model.
[0071] [2.2] When adding an enhancement node, merge the new enhancement node matrix into the matrix consisting of the feature nodes and enhancement nodes of the initial matrix, where: the newly added enhancement node is mapped from the feature node using randomly generated weights and biases. Then, calculate the pseudo-inverse of the matrix after adding the enhancement node to obtain the updated weight.
[0072] [2.3] When adding a feature node, obtain the new feature node and its corresponding enhancement node, and merge them with the matrix consisting of the feature nodes and enhancement nodes of the initial matrix. The updated weights can be calculated by the pseudo-inverse of the merged matrix.
[0073] Although the present invention has been described above, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many modifications without departing from the purpose of the present invention, which are all within the protection of the present invention.
Claims
1. A method for realizing task offloading of industrial IoT system based on wide learning neural network, It is characterized in that The steps include: Step 1: Build an initial task dataset by collecting information about servers, devices, and tasks in the IIoT Each row in the initial task data set corresponds to a task, and each column represents a specific task feature; the task feature includes information about the server, device, and task; Step 2: Classify the tasks in the initial task dataset according to the task creation time to obtain task samples X in different time slots m ; Step 3: Use the Gurobi solver to solve the task samples X for different time slots m Select the edge server label for optimal task offloading and obtain task samples with edge server labels; Step 4: Perform normalization on the task samples with edge server labels to obtain a normalized and labeled dataset; Where: m is the row index, n is the column index, and Respectively represent the maximum and minimum values of each column; Step 5: Extract important features of task samples in the normalized and labeled data set through the decision tree method, weight the task data samples, and obtain weighted input task data samples; Step 6: Train and update the task offloading neural network using weighted input task data samples.
2. According to claim 1, a method for realizing task offloading of industrial IoT system based on width learning neural network, It is characterized in that The process of extracting important features of task samples in the normalized and labeled data set by using the decision tree method in step 5 includes the following steps: Step 501: Based on the initial task data set The task characteristics S of the samples in n Calculate each task feature S according to the following formula n Gini index Where: z is the number of branches into which the dataset is divided, S n,l YesS n The lth sub-branch of ; Step 502: Based on each task feature S n The maximum and minimum values of the Gini index are calculated according to the following formula to calculate the task feature importance: Among them: The maximum and minimum values of the Gini index of task characteristics are and Step 503: Using the input task sample X m and feature importance Constructing weighted task data samples Generate training datasets for task offloading neural networks.
3. According to claim 1, a method for offloading tasks of an industrial IoT system based on a width learning neural network, It is characterized in that In step 6, the task offloading neural network is trained and updated by using the weighted input task data samples, including the following steps: Step 601: Calculate the data weighted value using the following formula: The value Z mapped to the task feature node i : Where: Φ i (·) is the mapping function, is the random weight and is the bias term; Step 602: Calculate the characteristic node value Z by the following formula i The value H mapped to the enhanced node j : Where: j (·) is the mapping function; Step 603: The combined matrix of the relationship between the data label Y and the feature nodes and the enhanced nodes is expressed by the following formula: Step 604: Solve the combination matrix [Z i ∣H j ] of the pseudo-inverse matrix [Z k ∣H u ] + ; Step 605: According to the pseudo-inverse matrix [Z k ∣H u ] + Update the task offloading neural network weights W * =[Z k ∣H u ] + Y; Step 606: By judging whether the task offloading neural network weight reaches the expected accuracy or new task data is continuously generated, if the conditions are met, then end, otherwise proceed to the next step; Step 607: Calculate the incremental features and enhanced nodes caused by the introduction of new task data, and add the incremental nodes to the feature node and enhanced node matrix of the initial task data set; Step 608: Calculate the weight update caused by the increment, obtain the updated weight, save it, and complete the task of unloading the neural network update.
