An Edge Collaborative Data Allocation Method Applied to Industrial Internet of Things

Abstract

The present invention relates to the field of industrial intelligent production, and specifically, it is an edge collaborative data distribution method applied to the industrial Internet of Things. It includes: deploying a pre-trained shallow neural network and a deep neural network at the local terminal device and the edge side at the industrial processing site respectively; building a system model closer to the actual application scenario considering the data processing capabilities and processing times of different devices; dynamically selecting the optimal data distribution method according to the system model. The present invention can customize appropriate data distribution decisions for machine learning tasks to shorten the system data processing time and can effectively adapt to the industrial production requirements.

Description

Technical Field

[0001] The present invention relates to the field of industrial intelligent production, and more particularly to an edge collaborative data distribution method applied to the industrial Internet of Things. Background Art

[0002] The industrial Internet of Things is an important technical foundation for driving the intelligent upgrade of industry. It realizes efficient intelligent production by interconnecting industrial intelligent devices with environmental perception, data processing, and transmission capabilities. With the rapid development of technologies such as artificial intelligence (AI) and edge computing, AI-enabled edge intelligence technology has been widely applied to the industrial Internet of Things architecture, enabling complex intelligent applications to be processed at the local edge side.

[0003] Currently, the methods for efficiently applying complex AI models to resource-constrained terminals and edge devices usually include: model splitting, model compression, early model exit, and edge task offloading, etc. However, the existing methods mainly focus on the deployment (offloading) methods of machine learning tasks at the terminal, edge, and cloud, and have not considered the impact of cloud, edge, and terminal collaborative data processing on machine learning tasks after task deployment. Summary of the Invention

[0004] Aiming at the problem of reducing system latency when efficiently applying complex AI models to the local and edge sides close to users to meet the real-time requirements of time-sensitive tasks, the present invention provides an edge collaborative data distribution method applied to the industrial Internet of Things.

[0005] The technical solution adopted by the present invention to achieve the above object is:

[0006] An edge collaborative data distribution method applied to the industrial Internet of Things, comprising the following steps:

[0007] Obtain the system characteristics in the workpiece sorting system based on the distributed deep neural network;

[0008] Construct an optimization model that minimizes the total data processing latency based on the system characteristics;

[0009] Analyze the data division ratio that minimizes the total latency according to the system characteristics;

[0010] Dynamically adjust the optimal data distribution method according to the influence of the system characteristics on the data division ratio.

[0011] The system characteristics include: the speed of the terminal device for processing data, the speed of the edge server for processing data, and the classification accuracy of the terminal device.

[0012] The workpiece sorting system based on the distributed deep neural network divides the collected raw data into two parts and assigns them to the terminal device or the edge server for data processing. α is the ratio of the raw data division. αm of the raw data is processed by the terminal device, and (1 - α)m of the raw data is assigned to the edge server for processing, where m is the total amount of raw data.

[0013] The terminal device deploys a pre-trained shallow neural network, and the edge server deploys a pre-trained deep neural network.

[0014] The aggregator calculates the information entropy of the shallow neural network and compares it with the set threshold T to determine whether the data processing result of the shallow neural network model is credible. If the information entropy is greater than the threshold, the classification result is not credible; otherwise, the classification result is credible. If it is credible, the data processing result is transmitted to the decision-making unit in the aggregator; if it is determined to be not credible, the raw data is sent to the edge server for data processing, where C represents the set of all possible classifications, x is the probability vector indicating the possibility of being assigned to each category, and i ∈ N*.

[0015] The optimization model for minimizing the total data processing delay is as follows:

[0016] The data processing delay of the terminal device is: τ local = αmdp l , where d is the size of a single data, and p l is the time taken to process 1 Mb of data locally;

[0017] The data processing time of the edge server is: where p e is the time taken for the edge server to process 1 Mb of data, M is the total amount of historical data processed by the terminal device, F is the total amount of data with untrusted classification results statistically obtained by the terminal device in M, and ξ is the penalty coefficient;

[0018] The total data processing delay can be expressed as: Γ = max{τ local , τ edge};

[0019] Then, the minimum total delay can be expressed as:

[0020]

[0021] s.t. C1: max{τ local , τ edge} < θ

[0022] C2: 0 < α ≤ 1

[0023] C3: ξ ≥ 0

[0024]

[0025] The specific data partitioning ratio that minimizes the total time delay according to system feature analysis is as follows:

[0026] When τ local ≥ τ edge At this time, At this time, Γ = τ local , and Γ increases monotonically with the increase of α; when τ local ≤ τ edge At this time, At this time, Γ = τ edge , and Γ decreases monotonically with the increase of α; when the data partitioning ratio that minimizes the total time delay At this time, the total time delay Γ is the smallest. At this time

[0027] The optimal data allocation method is as follows: Leave α * of the data for processing by the terminal device, and send (1 - α * ) of the data to the edge server for processing.

[0028] The dynamic adjustment of the optimal data allocation method according to the influence of system features on the data partitioning ratio is specifically as follows:

[0029] The optimal allocation ratio α * is positively correlated with the probability that the classification result of the shallow neural network of the terminal device is untrustworthy , and is negatively correlated with the ratio of the data processing speeds of the terminal device and the edge server ;

[0030] The penalty coefficient ξ is negatively correlated with and is negatively correlated with ; When the terminal device executes the allocation method, ξ is dynamically adjusted according to system features, so that the data partitioning ratio calculated by the terminal device tends to the data partitioning ratio that minimizes the total time delay, thereby improving the data processing efficiency of the entire workpiece sorting system.

[0031] The present invention has the following beneficial effects and advantages:

[0032] 1. Taking into account the data processing capabilities and processing times of different devices, a model that better conforms to the system features in the workpiece sorting system based on distributed deep neural networks is constructed.

[0033] 2. While improving the edge utilization rate, the system time delay is reduced.

[0034] 3. Using a dynamic method to give the optimal data allocation method, which can better meet the actual production needs. Description of the Drawings

[0035] Figure 1 This is the application scenario diagram of the present invention;

[0036] Figure 2 This is the system structure diagram of the present invention;

[0037] Figure 3 This is the data distribution method diagram of the present invention;

[0038] Figure 4 This is the method flow diagram of the present invention. Specific implementation manners

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] As Figure 4 shown, a data distribution method applied to edge collaboration in industrial Internet of Things includes the following steps:

[0041] Step 1: Obtain system features in a workpiece sorting system based on a distributed deep neural network;

[0042] Step 2: Construct an optimization model that minimizes the total data processing delay based on the system features;

[0043] Step 3: Analyze the data partitioning ratio that minimizes the total delay according to the system features;

[0044] Step 4: Dynamically adjust the optimal data distribution method according to the influence of the system features on the data partitioning ratio.

[0045] The said Step 1 includes:

[0046] In a workpiece sorting system based on a distributed deep neural network as Figure 1 shown, a pre-trained shallow neural network (SNN) is deployed on a terminal device, and a pre-trained deep neural network (DNN) is deployed on an edge server. The workpiece sorting system based on the distributed deep neural network divides the collected raw data into two parts, and gives them to the SNN model of the terminal device or the DNN model of the edge server for data processing. α is used as the data partitioning ratio, αm of the data is processed by the terminal device, and (1 - α)m of the data is allocated to the edge server for processing, where m is the total amount of data. The system features include: the data processing speed of the terminal device, the data processing speed of the edge server, and the classification accuracy of the terminal device. Figure 2 This is the system structure diagram.

[0047] The said Step 2 includes:

[0048] The data processing delay of the terminal device is: τ local = αmdp l, where α is the proportion of data allocated to the terminal device for processing, m is the total amount of data, d is the size of a single data, and p l is the time taken by the terminal device to process 1 Mb of data.

[0049] The data processing time of the edge server is: where p e is the time taken by the edge server to process 1 Mb of data, M is the total amount of historical data processed by the device, F is the total amount of data with untrustworthy local classification results statistically obtained from M, and ξ is the penalty coefficient.

[0050] The total delay of data processing can be expressed as: Γ = max{τ local , τ edge};

[0051] Taking the total delay of data processing as the objective, an optimization problem is constructed:

[0052]

[0053] s.t. C1: max{τ lical , τ edge} < θ

[0054] C2: 0 < α ≤ 1

[0055] C3: ξ ≥ 0

[0056]

[0057]

[0058] The said step 3 includes:

[0059] Due to the limited computing power of the SNN model deployed on the terminal device, classification errors may occur. Therefore, information entropy is introduced and compared with the set threshold T to determine whether the classification result of the terminal device is trustworthy. If the information entropy is greater than the threshold, the classification result is untrustworthy; otherwise, the classification result is trustworthy. If it is trustworthy, the data processing result is transmitted to the decision-making unit in the aggregator; if it is untrustworthy, the data is sent to the edge server for re-data processing, where C represents the set of all possible classifications, x is a probability vector indicating the possibility of being assigned to each category, and i ∈ N*.

[0060] For the optimization problem described in step 2, when τ local ≥ τ edge , at this time Γ = τ local , and Γ increases monotonically with the increase of α; when τ local ≤ τ edge , At this time, Γ = τ edge , Γ decreases monotonically as α increases. Therefore, the data allocation ratio that minimizes the total delay Γ is obtained: At this time, the system delay

[0061] Step 4 includes:

[0062] Figure 3 It is a data allocation method diagram. According to the classification error rates of different SNN models and the processing speeds of different devices, different penalty coefficients ξ are defined to make the data allocation ratio calculated by the terminal device tend to the data allocation ratio that minimizes the total delay. Among them, the penalty coefficient ξ is negatively correlated with negatively correlated, and with negatively correlated. The optimal data allocation ratio α * is positively correlated with the probability that the classification result of the shallow neural network of the terminal device is untrustworthy and negatively correlated with the ratio of the data processing speeds of the terminal device and the edge server negatively correlated. Based on this, the optimal data allocation method is dynamically given: Leave the data of α * to be processed by the terminal device, and send the data of (1 - α * ) to be processed by the edge server.

Claims

1. An edge collaborative data allocation method applied to the industrial Internet of Things, characterized in that It includes the following steps: Obtain the system characteristics in the workpiece sorting system based on the distributed deep neural network; Construct an optimization model that minimizes the total data processing delay based on the system characteristics; Analyze the data partitioning ratio that minimizes the total delay according to the system characteristics; Dynamically adjust the optimal data allocation method according to the influence of the system characteristics on the data partitioning ratio; The optimization model that minimizes the total data processing delay is: The data processing delay of the terminal device is: τ local = αmdp l , where d is the size of a single data, p l is the time taken to process 1 Mb of data locally, α is the proportion of the original data partition, and m is the total amount of the original data; The data processing time of the edge server is: where p e is the time taken by the edge server to process 1 Mb of data, M is the total amount of historical data processed by the terminal device, F is the total amount of data with untrustworthy classification results statistically obtained by the terminal device in M, and ξ is the penalty coefficient; The total time delay of data processing is expressed as: Γ = max{τ local , τ edge}; Then, the minimum total delay is expressed as:

2. The edge collaborative data distribution method applied to the industrial Internet of Things according to claim 1, wherein The system characteristics include: the speed of the terminal device for processing data, the speed of the edge server for processing data, and the classification accuracy of the terminal device.

3. The edge collaborative data distribution method applied to the industrial Internet of Things according to claim 1, wherein The workpiece sorting system based on the distributed deep neural network divides the collected original data into two parts and allocates them to the terminal device or the edge server for data processing. α is the ratio of the original data partitioning. αm of the original data is processed by the terminal device, and (1 - α)m of the original data is allocated to the edge server for processing, where m is the total amount of the original data.

4. An edge collaborative data distribution method applied to the industrial Internet of Things according to claim 3, characterized in that, The terminal device deploys a pre-trained shallow neural network, and the edge server deploys a pre-trained deep neural network.

5. The edge collaborative data distribution method applied to the industrial Internet of Things according to claim 4, characterized in that The aggregator calculates the information entropy of the shallow neural network and compares it with the set threshold T to determine whether the data processing result of the shallow neural network model is credible. If the information entropy is greater than the threshold, the classification result is not credible; otherwise, the classification result is credible. If it is credible, the data processing result is passed to the decision-making unit in the aggregator; If it is determined to be untrusted, the original data is sent to the edge server for data processing, where C represents the set of all possible classifications, x is a probability vector indicating the possibility of being assigned to each category, and i ∈ N*.

6. The edge collaborative data distribution method applied to the industrial Internet of Things according to claim 1, wherein, Specifically, analyzing the data partitioning ratio that minimizes the total delay according to the system characteristics is: When τ local ≥ τ edge , at this time Γ = τ local , Γ increases monotonically with the increase of α; when τ local ≤ τ edge , at this time Γ = τ edge , Γ decreases monotonically with the increase of α; when the data partitioning ratio that minimizes the total delay , the total delay Γ is the smallest, and at this time 7. The edge collaborative data distribution method applied to the industrial Internet of Things according to claim 6, characterized in that, The optimal data allocation method is: leave the data of α * at the terminal device for processing, and send the data of (1 - α * ) to the edge server for processing.

8. The edge collaborative data distribution method applied to the industrial Internet of Things according to claim 7, characterized in that, Specifically, dynamically adjusting the optimal data allocation method according to the influence of the system characteristics on the data partitioning ratio is: Optimal allocation ratio α * is positively correlated with the probability that the classification result of the shallow neural network of the terminal device is untrustworthy and is negatively correlated with the ratio of the data processing speeds of the terminal device and the edge server ; The penalty coefficient ξ is negatively correlated with and negatively correlated with When the allocation method is executed on the terminal device, ξ is dynamically adjusted according to the system characteristics, so that the data division ratio calculated by the terminal device tends to the data division ratio with the minimum total delay, thereby improving the data processing efficiency of the entire workpiece sorting system.

Images