An industrial edge collaborative scheduling optimization method and system based on federal digital twinning

Abstract

The present application relates to the technical field of industrial edge scheduling optimization, and more particularly to an industrial edge collaborative scheduling optimization method and system based on federated digital twinning. The method comprises extracting and compressing a feature vector by using local convolution and principal component analysis; mapping the feature vector into a digital twinning model of an edge node; calculating a scheduling weight score of each edge node based on the digital twinning model; constructing a global model based on a neural network, wherein the global model parameters and the learning rate are initialized based on the scheduling weight score; combining the hyperbolic tangent function and the collaborative trust score to aggregate the updated parameter gradient of each edge node and calculate a task migration score; and performing migration scheduling according to the task migration score.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial edge scheduling optimization, and in particular to an industrial edge collaborative scheduling optimization method and system based on federated digital twinning. BACKGROUND

[0002] With the rapid development of industrial internet and intelligent manufacturing, more and more industrial enterprises deploy edge computing nodes and digital twinning models to improve data acquisition efficiency, process visualization capability and distributed intelligent decision-making level. Digital twinning technology can make predictions, optimizations and controls in virtual space through real-time modeling and mapping of physical objects, and is an important infrastructure for industrial intelligence. At the same time, edge computing reduces transmission delay and enhances system response capability by sinking computing tasks to edge nodes close to data sources. In recent years, in order to solve the model synchronization and privacy protection problems between multiple edge nodes, federated learning has been gradually introduced into the industrial edge scene for collaborative training of shared models without direct exchange of raw data.

[0003] However, the above-mentioned existing industrial edge collaborative computing method still has many problems. First, in a large-scale distributed environment, the digital twinning node has a huge time and space data synchronization pressure, and traditional compression algorithms are difficult to balance data fidelity and transmission efficiency, resulting in lagging and distorted twinning models. Second, the computing power of edge nodes is heterogeneous and the tasks change frequently, and the current federated training algorithm mostly uses static weighted average method, which cannot fully reflect the difference in node capability and state change, seriously affecting the model convergence quality. Third, the existing computing flow scheduling strategy is mostly based on fixed rules or single-factor scoring, lacking comprehensive perception of network bandwidth fluctuation, task urgency and node stability, which easily leads to uneven allocation of computing resources and unbalanced system load.

[0004] Therefore, an innovative collaborative computing system that integrates efficient compression modeling, adaptive federated training and intelligent task migration mechanism is urgently needed to meet the urgent needs of low delay, high throughput and strong robustness in large-scale industrial edge scenarios. SUMMARY

[0005] In order to solve the above-mentioned problems, the present application provides an industrial edge collaborative scheduling optimization method and system based on federated digital twinning. To solve the technical problems of low time and space data synchronization efficiency, rigid edge collaborative optimization mechanism and lack of intelligence in task migration in multi-node industrial scenarios.

[0006] In a first aspect, the present application provides an industrial edge collaborative scheduling optimization method based on federated digital twinning, which adopts the following technical solution:

[0007] An industrial edge collaborative scheduling optimization method based on federated digital twinning, comprising:

[0008] obtain industrial time series data of the edge nodes;

[0009] extract compressed feature vectors by using local convolution and principal component analysis;

[0010] map the feature vectors into digital twin models of the edge nodes;

[0011] calculate scheduling weight scores of each edge node based on the digital twin models;

[0012] construct a global model based on a neural network, wherein the global model parameters and learning rate are initialized based on the scheduling weight scores; combine the updated parameter gradients of each edge node by using a hyperbolic tangent function and a collaborative trust score to obtain a task migration score;

[0013] perform migration scheduling according to the task migration score.

[0014] Further, the extracting compressed feature vectors by using local convolution and principal component analysis comprises: firstly applying a local convolution operation to the original data to extract local spatial features; and then performing principal component analysis (PCA) on the extracted local spatial features to retain only part of the principal components, thereby finally generating compressed feature vectors of the edge nodes.

[0015] ,

[0016] wherein, represents the compressed feature vector of the edge node i at time ; is the original data, and represents the industrial time series data of the edge node i at time ; is the size of the window in the local convolution operation; represents the number of retained principal components; represents the convolution processing on the original data of the edge node i at time , represents the size of the convolution kernel; represents the principal component analysis on the local spatial features extracted by convolution, and the most representative principal components are retained.

[0017] Further, the mapping of the feature vector into the digital twin model of the edge node includes preprocessing the real-time sensor data and device state information data collected by the edge node, including removing noise, data normalization, and feature extraction, combining the preprocessed data with the compressed feature vector and the resource capability vector of the edge node, and inputting them into the digital twin model, wherein the real-time sensor data and device state information are directly mapped by using an API interface to directly write the corresponding parameters of the existing digital twin model; the compressed feature vector and the resource capability vector of the edge node are mapped by using a rule engine, which is converted into executable instructions of the digital twin model by a preset rule, thereby forming a digital twin model of each edge node; the digital twin model of each edge node represents the current state, resource capability, and task execution of the edge node.

[0018] Further, the calculation of the scheduling weight score of each edge node based on the digital twin model includes calculating the scheduling weight score of each edge node at the current time based on the compressed feature vector and the resource capability vector in the digital twin model of each edge node, wherein the two-norm of the compressed feature vector in the digital twin model of the edge node is first calculated to measure the importance of the compressed feature vector of the edge node in its digital twin model; then the two-norm of the resource capability vector is calculated to represent the comprehensive computing power level of the edge node; finally, the feature strength is divided by the two-norm of the resource capability vector in the digital twin model of each edge node plus a positive number 𝜖, thereby obtaining the scheduling weight score of the edge node i in the node feature model at time

[0019]

[0020] wherein, represents a positive number to ensure that the denominator in the calculation is not zero, and is a constant; represents the two-norm of the compressed feature vector represents the two-norm of the resource capability vector.

[0021] Further, the initialization of the global model parameters and the learning rate based on the scheduling weight score includes defining a global model based on a neural network, initializing the current global model parameters including the weight matrix, the bias vector, and the learning rate, using the hyperbolic tangent function and the collaborative trust score to aggregate the updated parameter gradients uploaded by each edge node, and then performing federated optimization iteration:

[0022]

[0023] wherein, represents the current ​​​​​global model parameters at time t; updated global model parameters at time t+1; denotes a learning rate, denotes the total number of edge nodes participating in federated learning; denotes the index of an edge node, denotes an edge node updated parameter gradient uploaded to the global model; denotes a hyperbolic tangent function, denotes an edge node at time t collaborative trust score at time t.

[0024] Further, the aggregation of the updated parameter gradient of each edge node in combination with the hyperbolic tangent function and the collaborative trust score includes first taking the edge node the scheduling weight score at the current time t as a whole reflection of the resources and data quality of the edge node the gradient of the compressed feature vector in the local digital twin model of the current global model parameter , which represents the fitting direction of the compressed feature vector in the local digital twin model of the edge node to the current global model parameter, and the direction and amplitude of the parameter update of the global model between the last time t and the current time t+1 perform a cosine similarity calculation to measure the degree of fit of the edge node to the change trend of the current global model, and finally multiply the scheduling weight score by the cosine value to obtain the edge node collaborative trust score at the current time t reflects the collaborative effectiveness and trust level of the edge node in this round;

[0025] ,

[0026] wherein, denotes the direction and amplitude of the global parameter update between the last time t and the current time t+1, i.e., the change trend of the global model.

[0027] Further, the aggregation of the updated parameter gradient of each edge node in combination with the hyperbolic tangent function and the collaborative trust score further includes evaluating whether to migrate the computing task to other nodes according to the node state and resource capability information in the global digital twin model; first, taking the edge node the collaborative trust score in the training at the current time t as a basic factor, introducing a task urgency function ​, the task urgency function combines the resource intensity required by the computing task on the edge node i and the task waiting time , by taking the logarithm of the product of the two after adding one, it ensures that urgent tasks are prioritized during migration; finally, considering the current bandwidth state of the target node and the network fluctuation degree in the past time window , where represents the variance or standard deviation of the bandwidth in the past fixed time window, the fluctuation degree term is adjusted by the bandwidth fluctuation sensitivity factor , and finally the overall calculation is nonlinearly enhanced by the power index , the higher the migration score, the more suitable the target node is as the migration target of the current computing flow, the formula is:

[0028] ,

[0029] where, represents the computing task migration score, the overall adaptability of the computing task from node to node at the current time; represents the resource intensity requirement of task i.

[0030] Secondly, an industrial edge collaborative scheduling optimization system based on federal digital twinning includes:

[0031] The data acquisition module is configured to acquire industrial time series data of edge nodes;

[0032] The feature module is configured to extract compressed feature vectors using local convolution and principal component analysis;

[0033] The mapping module is configured to map the feature vectors into the digital twin model of the edge node;

[0034] The scheduling module is configured to calculate the scheduling weight score of each edge node based on the digital twin model;

[0035] The global module is configured to build a global model based on a neural network, wherein the global model parameters and learning rate are initialized based on the scheduling weight score; the updated parameter gradient of each edge node is aggregated by combining the hyperbolic tangent function and the collaborative trust score, and the task migration score is calculated;

[0036] The migration module is configured to perform migration scheduling according to the task migration score.

[0037] In a third aspect, the present application provides a computer-readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and to execute the industrial edge collaborative scheduling optimization method based on federated digital twinning.

[0038] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer-readable storage medium, the processor being configured to implement instructions, and the computer-readable storage medium being configured to store a plurality of instructions, the instructions being adapted to be loaded by the processor and to execute the industrial edge collaborative scheduling optimization method based on federated digital twinning.

[0039] In summary, the present application has the following beneficial technical effects:

[0040] 1. Improving the dynamic scheduling and load balancing capability of computing flow: through the combination of federated learning and edge computing, the task allocation and computing flow migration can be adjusted in real time, ensuring the load balancing of each edge node and improving the computing efficiency.

[0041] 2. Enhancing the adaptive capability of the system: by utilizing the digital twinning model of the edge node, the migration of the computing flow can be automatically evaluated and optimized according to the resource status of the node, the urgency of the computing task, the bandwidth state and other factors, ensuring the execution of the task on the appropriate node.

[0042] 3. Improving the execution efficiency of the computing task: through the computing task migration score and the collaborative trust score, the optimal target node can be selected for the migration of the computing flow, reducing the delay of task execution and improving the overall performance.

[0043] 4. Optimizing resource usage: by accurately evaluating the resource capability of each edge node and the demand of the computing task, efficient resource allocation can be achieved, avoiding node overload or resource waste and improving the overall operation efficiency of the system.

[0044] 5. Supporting efficient operation of large-scale distributed systems: suitable for scenarios where multiple edge nodes work collaboratively, efficient computing resource scheduling and task management can be achieved without increasing the burden of the central server, adapting to complex distributed computing tasks. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 A structure diagram of the industrial edge collaborative scheduling optimization method based on federated digital twinning is provided. DETAILED DESCRIPTION

[0046] The present application will be further described in detail below with reference to the accompanying drawings.

[0047] Example 1

[0048] Reference Figure 1The industrial edge collaborative scheduling optimization method based on federal digital twinning of the embodiment comprises:

[0049] Obtaining industrial time series data of the edge node;

[0050] Extracting compressed feature vectors by local convolution and principal component analysis;

[0051] Mapping the feature vectors into the digital twin model of the edge node;

[0052] Calculating the scheduling weight score of each edge node based on the digital twin model;

[0053] Constructing a global model based on a neural network, wherein the global model parameters and the learning rate are initialized based on the scheduling weight score; combining the tangent hyperbolic function and the collaborative trust score to aggregate the updated parameter gradient of each edge node and calculate the task migration score;

[0054] Migrating and scheduling according to the task migration score.

[0055] Specifically,

[0056] S1. Data acquisition

[0057] Among them, at each edge node, data acquisition is performed by data acquisition equipment such as sensors to obtain industrial time series data, including motor current, temperature data, vibration data, rotational speed / frequency, load power, etc.

[0058] S2. Multi-node twin modeling and capability evaluation: collect and compress the industrial time series data of the edge node, extract compressed feature vectors by local convolution and principal component analysis, and map the three-dimensional resource capability vector, the compressed feature vector of each edge node, the real-time sensor data and device state information data collected by the edge node into the digital twin model of each edge node; calculate the scheduling weight score of the edge node to evaluate its comprehensive computing power level;

[0059] Among them, the feature extraction includes extracting compressed feature vectors by local convolution and principal component analysis, and the specific formula is:

[0060] ,

[0061] The constructed three-dimensional resource capability vector includes: Defining a three-dimensional resource capability vector , respectively representing the current time computing processing capability (such as CPU operation frequency or core number) of the edge node, the memory capacity (such as the available memory space size), and the network bandwidth condition (i.e. data uplink / downlink rate).

[0062] The three-dimensional resource capability vector, the compressed feature vector of each edge node, real-time sensor data and device state information data collected by the edge node are mapped into the digital twin model of each edge node. The contents are described below:

[0063] The preprocessed data is combined with the compressed feature vector and the resource capability vector of the edge node and input into the existing digital twin model. Among them, the real-time sensor data and device state information are directly mapped, that is, through the API interface or database link, directly written into the corresponding parameters of the existing digital twin model. And the compressed feature vector and the resource capability vector of the edge node adopt the rule engine mapping method, that is, through the preset rule (such as formula or logical judgment) to convert into executable instructions of the digital twin model, and then form the digital twin model of each edge node.

[0064] S3. Feature extraction

[0065] At each edge node, the industrial time series data of the current edge node at the current time is collected as raw data. In order to improve the synchronization efficiency of the raw data and reduce the transmission load, the raw data is compressed and extracted. Specifically, first, a local convolution operation is applied to the raw data to extract local spatial features; next, the extracted local spatial features are processed by principal component analysis (PCA) to retain only the most representative principal components to achieve data dimension compression. Finally, the compressed feature vector of the corresponding edge node is generated, and the formula is as follows:

[0066]

[0067] represents the compressed feature vector of the edge node i at time is the raw data, which represents the industrial time series data of the edge node i at time is the size of the window in the local convolution operation, which is derived from the actual task demand, and the value range is [1, 10]; represents the number of retained principal components; represents the convolution processing of the raw data of the edge node i at time represents the size of the convolution kernel; represents the principal component analysis (PCA) of the local spatial features extracted by convolution, and the most representative principal components are retained.

[0068] S4. Digital twin model

[0069] ​​​​​​After obtaining the compressed feature vector of edge node i, the computing resources of the edge node also need to be quantified and represented. For each edge node A three-dimensional resource capability vector is defined , respectively representing the current time computing processing capability (such as CPU operation frequency or core number) of the edge node, the memory capacity (for example, the size of available memory space), and the network bandwidth status (i.e. data uplink / downlink rate).

[0070] The real-time sensor data (such as temperature, humidity, pressure, etc.) and device status information data (such as power consumption, running status) collected by the edge node are preprocessed, including removing noise, data normalization (ensuring data in the same scale range) and feature extraction (extracting key features) and other steps to ensure data quality and availability. The above preprocessed data, combined with the compressed feature vector and the resource capability vector of the edge node, are respectively input into the existing digital twin model. Among them, the real-time sensor data and device status information use direct mapping, that is, through API interface or database link, directly write into the corresponding parameters of the existing digital twin model; such as temperature data directly writes into the environmental temperature parameter in the model, and device power consumption directly updates the energy consumption parameter in the model. The compressed feature vector and the resource capability vector of the edge node use rule engine mapping, that is, by pre-setting logic according to the specific application scene demand, the compressed feature vector is converted into executable instructions of the digital twin model, such as when the bandwidth element in the resource capability vector is lower than the preset value, the model automatically updates its network availability parameter; and then the digital twin model of each edge node is formed.

[0071] The digital twin model of each edge node represents the current state, resource capability and task execution of the edge node. Since the state and resource capability of the edge node are constantly changing (for example, industrial time series changes over time), the digital twin model of the edge node needs to be updated regularly, and the updated parameters (such as Gradient is used for subsequent federated training.

[0072] S5. Schedule weight score

[0073] Next, based on the compressed feature vector and the resource capability vector in the digital twin model of each edge node, the scheduling weight score of each edge node at the current time is calculated: first, the two-norm of the compressed feature vector in the edge node digital twin model (i.e. feature intensity) is calculated to measure the importance of the compressed feature vector of the edge node in its digital twin model; then the two-norm of the resource capability vector is calculated, representing the comprehensive computing power level of the edge node. Finally, the feature intensity is divided by the two-norm of the resource capability vector in the digital twin model of each edge node plus a positive number 𝜖, thereby obtaining the scheduling weight score of edge node i in the digital twin model of the node feature model at time the scheduling weight score :

[0074] ,

[0075] wherein, is a positive number, ensuring that the denominator in the calculation is not zero, is a constant (such as ), to improve the numerical stability of the formula; represents the two-norm of the compressed feature vector ; represents the two-norm of the resource capability vector.

[0076] S6. Federated training and computing flow migration collaborative execution: initialize global model parameters and learning rate; combine the hyperbolic tangent function and the collaborative trust score to aggregate the updated parameter gradient uploaded by each edge node; calculate the task migration score; finally select the target node with the maximum migration score to perform computing flow migration, ensuring load balancing and dynamic scheduling.

[0077] S6.1 Define a global model based on a neural network for updating the parameter gradient of the digital twin model of the edge node in the federated learning environment, support task scheduling and load balancing of edge computing, including weight matrix, bias vector, and other global model parameters.

[0078] S6.2 Federated training and computing flow migration collaborative execution: use the hyperbolic tangent function to perform nonlinear transformation and smoothing processing on the parameter gradient uploaded by each edge node, and update the smoothed parameter gradient based on the collaborative trust score calculated based on the scheduling weight score; At the same time, based on the collaborative trust score, calculate the task migration score; finally select the target node with the maximum migration score to perform computing flow migration, realize load balancing and dynamic scheduling.

[0079] The scheduling weight score of the edge node is used to evaluate the collaborative trustworthiness of the edge node. Specifically, first, take the scheduling weight score of the edge node at the current time as the overall reflection of the resource and data quality of the edge node . Further, calculate the gradient of the compressed feature vector of the edge node , which represents the fitting direction of the compressed feature vector of the current edge node , and the cosine similarity calculation is performed on the difference between the parameter gradient update and the parameter gradient update between the last time and the current time . Finally, multiply the scheduling weight score by the cosine value to obtain the collaborative trust score of the edge node at the current time, reflecting the collaborative effectiveness and trustworthiness of the edge node at the current time.

[0080] ,

[0081] in, Indicates the previous moment With the current moment The difference in parameter gradient updates between them, i.e., the trend of parameter gradient changes; The cosine similarity function is used to measure the angular similarity between two vectors, with a value range of [-1, 1]. The collaborative credibility score is used to adjust the contribution of each edge node in the parameter gradient update process. The higher the collaborative credibility score, the more consistent the gradient direction of the current edge node is with the global update direction, i.e., the greater the contribution of the current edge node.

[0082] S6.3 updates the parameter gradients of the digital twin model of edge nodes in a federated learning environment. Specifically, firstly, to improve the stability of the parameter gradient update process, a hyperbolic tangent function is used ( The parameter gradients uploaded by each edge node are smoothed using a nonlinear transformation to mitigate interference from extreme gradients. Next, the smoothed parameter gradients of each edge node are weighted and summed based on the collaborative credibility score of each edge node to form the aggregated parameter gradient. Furthermore, the updated parameter gradient θ(t+1) is obtained.

[0083] ,

[0084] in, Indicates the current The gradient of parameters at time step, i.e., the global model parameters; Indicates the next moment (the...) The updated parameter gradient at time (time); The learning rate is used to control the step size of each parameter gradient update. Its value is derived from expert experience or adaptive learning rate algorithm, and is set according to the actual situation. The value range is [0.001, 0.01]. This represents the total number of edge nodes participating in federated learning; Indicates the first The index of each edge node, ; Let be the loss function, representing the parameter gradient. At the edge node compressed feature vectors The prediction error used in the prediction process; Indicates the current time at the edge node Update direction; This represents the hyperbolic tangent function, which has the effect of smoothing and limiting extreme values, preventing the training process from becoming unstable due to excessive gradients. Represents edge nodes In the Time-based collaborative credibility scoring;

[0085] The updated parameter gradient θ(t+1) is fed back to each edge node until the change in the parameter gradient is less than the threshold preset by the expert experience method. The edge nodes continue to perform local training and updates based on the updated parameter gradient.

[0086] S6.4 Based on the node status and resource capacity information of the edge nodes, assess whether it is necessary to migrate the computing tasks to other nodes. Specifically, this includes: using edge nodes... Collaborative credibility scoring As a fundamental factor, a task urgency function is introduced. This is used to characterize the urgency of the computational tasks performed on edge node i. The task urgency function combines the resource intensity required by the computational tasks on edge node i. The task has been waiting time. By taking the logarithm of the product of adding one to each of the two, we ensure that urgent tasks are prioritized during the migration process. Finally, we consider the target node. Current bandwidth status Network volatility over the past time window ,in This represents the variance or standard deviation of bandwidth over a fixed time window, reflecting the degree of change and stability of the target node's network bandwidth. The volatility term is expressed through a bandwidth volatility sensitivity factor. Adjust the weights, and finally use the power exponent. The overall calculation above is enhanced nonlinearly to generate a transfer score. A higher transfer score indicates that the target section is more suitable as a transfer target for the current computational flow. The formula for calculating the task transfer score is:

[0087] ,

[0088] in, This represents the calculation of task transfer score at time [time]. Computational tasks from edge nodes Assigned to nodes The overall fit is indicated by a higher value, which means a higher degree of matching and a more suitable migration. The resource intensity requirements (such as CPU / GPU time usage, memory requirements, etc.) of the task on edge node i are normalized; This represents the waiting time from submission to the current state of the task on edge node i; The task urgency function, which represents the urgency of a task's execution, is defined as follows: ; bandwidth load status of the target node bandwidth load status of the target node network bandwidth utilization at time network bandwidth utilization at time bandwidth fluctuation degree of the target node j, which is the variance of the bandwidth of the target node j in a past time window, reflecting the network stability; bandwidth fluctuation sensitivity factor, controlling the influence of the bandwidth fluctuation degree on the migration score; adjustment coefficient, which can adjust the response strength of the urgency of the computing task, the bandwidth status, and the bandwidth fluctuation degree.

[0089] S6.5 Final selection of migration score the node with the maximum migration score as the migration target, i.e. the node with the value as the migration target, the computing task at the node is migrated to the node for processing, realizing the migration of the computing flow.

[0090] After obtaining the migration scores of all candidate target nodes , all reachable nodes are scored and sorted, and the target node with the maximum score value is selected as the optimal migration object of the current computing task. The migration score has comprehensively considered the cooperative trust score of the source node, the urgency of the current task, and the bandwidth load status and bandwidth fluctuation degree of the target node, and can reflect the optimal adaptability of task migration in the current environment.

[0091] S6.6 After the above global model is updated, the process of global model iteration in the next time will continue, and the updated global model θ(t+1) will be fed back to each edge node each time. The edge node continues local training and updating based on the new global model until the change of the global model parameter is less than the threshold preset by the expert experience method, indicating that the global model has converged, and the training is terminated.

[0092] Embodiment 2

[0093] The embodiment provides an industrial edge cooperative scheduling optimization system based on federated digital twinning, comprising:

[0094] The data acquisition module is configured to:

[0095] A computer-readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded and executed by a processor of a terminal device.

[0096] A terminal device comprises a processor and a computer readable storage medium, the processor is used to realize instructions; the computer readable storage medium is used to store a plurality of instructions, the instructions are suitable for being loaded by the processor and executing the based on federal digital twin industrial edge collaborative scheduling optimization method.

[0097] The above are preferred embodiments of the present application, not limited by the protection scope of the present application, therefore: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. An industrial edge collaborative scheduling optimization method based on federal digital twinning, characterized in that, The method comprises: obtaining industrial time series data of an edge node; extracting a compressed feature vector by using local convolution and principal component analysis; mapping the feature vector into a digital twin model of the edge node; The method comprises preprocessing real-time sensor data and equipment state information data collected by the edge node, including removing noise, data normalization and feature extraction, combining the preprocessed data with a compressed feature vector and a resource capability vector of the edge node, and inputting them into the digital twin model respectively, wherein the real-time sensor data and the equipment state information are directly mapped by using an API interface to write corresponding parameters of the existing digital twin model; the compressed feature vector and the resource capability vector of the edge node are mapped by using a rule engine, and are converted into executable instructions of the digital twin model by using a preset rule, thereby forming a digital twin model of each edge node; the digital twin model of each edge node represents the current state, resource capability and task execution of the edge node; calculating a scheduling weight score of each edge node based on the digital twin model; The global model based on the neural network is constructed, wherein the global model parameters are initialized based on the scheduling weight score and the learning rate; the updated parameter gradient of each edge node is aggregated by combining the hyperbolic tangent function and the collaborative trust score, and a task migration score is calculated; the updated parameter gradient of each edge node is aggregated by combining the hyperbolic tangent function and the collaborative trust score, which comprises: first, taking the edge node The scheduling weight score in the current round , as the edge node The overall reflection of resources and data quality, then calculate the current global model parameters in the edge node The gradient of the compressed feature vector in the local digital twin model , indicates the current global model in the edge node The fitting direction of the compressed feature vector in the local digital twin model, and the parameter update direction of the global model between the last round and the current round Cosine similarity calculation is performed, which measures the degree of fit of the edge node The change trend of the current global model, finally, the scheduling weight score is multiplied by the cosine value to obtain the edge node The collaborative trust score in the current round of training, reflecting the collaborative effectiveness and trust degree of the edge node In the current round; ， wherein, represents the direction and magnitude of the global parameter update between the previous round and the current round, i.e., the change trend of the global model between two rounds of training, represents the edge node In the first collaborative trust score in the round of training; Migration scheduling is performed based on task migration scores; the aggregation of updated parameter gradients for each edge node using the hyperbolic tangent function and collaborative reliability scores also includes assessing whether the computational task needs to be migrated to other nodes based on node status and resource capability information in the global digital twin model; firstly, edge nodes... Collaborative Trust Scoring in this Round of Training As a fundamental factor, a task urgency function is introduced. The task urgency function is combined with the resource intensity required for the task calculated on edge node i. The task has been waiting time. By taking the logarithm of the product of adding one to each of the two, we ensure that urgent tasks are prioritized during the migration process; finally, we consider the target node. Current bandwidth status Network volatility over the past time window ,in The volatility term represents the variance or standard deviation of bandwidth over a fixed time window in the past, expressed by the bandwidth volatility sensitivity factor. Adjust the weights, and finally use the power exponent. The overall calculation above is enhanced nonlinearly to generate a transfer score. The higher the transfer score, the more suitable the target section is as a transfer target for the current computation flow. The formula is as follows: ， wherein, represents a computation task migration score, at the round computation task is assigned to the node overall fitness of the node ; represents the resource intensity requirement of task i.

2. The industrial edge collaborative scheduling optimization method based on federal digital twinning according to claim 1, characterized in that, The method of extracting the compressed feature vector by using local convolution and principal component analysis comprises: firstly applying a local convolution operation to the original data to extract local spatial features; then applying principal component analysis (PCA) to the extracted local spatial features, and only retaining part of the principal components to finally generate a compressed feature vector corresponding to the edge node: ， in, This represents the compressed feature vector of edge node i at time t; The original data represents edge node i at time [time]. Industrial time series data; It refers to the size of the window during local convolution operations; Indicates the number of principal components retained; This indicates that for edge node i at time... raw data Perform convolution processing. Indicates the size of the convolution kernel; This indicates that principal component analysis is performed on the local spatial features extracted by convolution, and the most representative features are retained. One principal component.

3. The industrial edge collaborative scheduling optimization method based on federal digital twinning according to claim 2, characterized in that, The scheduling weight score of each edge node is calculated based on the digital twin model, including calculating the scheduling weight score of each edge node at the current time based on the compressed feature vector and the resource capability vector in the digital twin model of each edge node, wherein the two-norm of the compressed feature vector in the edge node digital twin model is calculated first to measure the importance of the compressed feature vector of the edge node in its digital twin model; then the two-norm of the resource capability vector is calculated, representing the comprehensive computing power level of the edge node; finally, the feature intensity is divided by the two-norm of the resource capability vector in the digital twin model of each edge node plus a positive number epsilon, thereby obtaining the scheduling weight score of the edge node i in the node feature model at time t : ， wherein, represents a positive number, ensures that the denominator is not zero in the calculation, is a constant; represents the two-norm of the compressed feature vector ; represents the two-norm of the resource capability vector.

4. The industrial edge collaborative scheduling optimization method based on federal digital twinning according to claim 3, characterized in that, The global model parameter initialization based on the scheduling weight score and the learning rate, comprising defining a global model based on a neural network, initializing the current global model parameters, including a weight matrix, a bias vector and a learning rate, using a hyperbolic tangent function and the collaborative trust score aggregates the updated parameter gradients uploaded by each edge node, and then performs a federated optimization iteration: ， in, Indicates the current number Global model parameters of the wheel; This indicates the global model parameters after the next update. Indicates the learning rate. This represents the total number of edge nodes participating in federated learning; Indicates the first The index of each edge node, Represents edge nodes The updated parameter gradients uploaded to the global model; Represents the hyperbolic tangent function. Represents edge nodes In the Collaborative reliability scoring in round training.

5. An industrial edge collaborative scheduling optimization system based on federated digital twinning, performing an industrial edge collaborative scheduling optimization method based on federated digital twinning as claimed in claim 1, characterized in that, The method comprises: a data acquisition module configured to obtain industrial time series data of an edge node; a feature module configured to extract a compressed feature vector by using local convolution and principal component analysis; a mapping module configured to map the feature vector into a digital twin model of the edge node; a scheduling module configured to calculate a scheduling weight score of each edge node based on the digital twin model; a global module configured to construct a global model based on a neural network, wherein the global model parameters and the learning rate are initialized based on the scheduling weight score; the updated parameter gradient of each edge node is aggregated by using a hyperbolic tangent function and a collaborative trust score, and a task migration score is calculated; a migration module configured to perform migration scheduling according to the task migration score.

6. A computer-readable storage medium having stored therein a plurality of instructions, wherein the instructions, when executed by a processor, cause the processor to perform operations comprising: The instructions are adapted to be loaded and executed by a processor of a terminal device to implement the industrial edge collaborative scheduling optimization method based on federated digital twin. 7.A terminal device, comprising a processor and a computer readable storage medium, the processor is configured to implement instructions; the computer readable storage medium is configured to store a plurality of instructions, characterized in that, The instructions are adapted to be loaded and executed by a processor to implement the industrial edge collaborative scheduling optimization method based on federated digital twin.

Images