Personalized federal learning method and system for heterogeneous multi-source industrial internet

Abstract

The invention relates to the related technical field of digital data processing, in particular to a personalized federated learning method and system for a heterogeneous multi-source industrial internet, and the method comprises the steps: connecting a client, evaluating a load, time delay and modal similarity to generate a dynamic association table, deploying a hierarchical encryption protocol, and constructing a priority queue; a cache mechanism is set to coordinate distributed iterative optimization, so that the technical problem that network oscillation and computing resource waste are aggravated due to overhigh load of part of nodes caused by frequent access and exit of equipment and data volume difference in the industrial internet and repeated migration of clients and nodes is caused is solved, cross-equipment shared knowledge base vectors are extracted, and the computing efficiency is improved. The technical effects of reducing the influence of model isomerism on aggregation, dynamically scheduling high-frequency parameter local aggregation and low-frequency parameter cloud synchronization, optimizing the association weight of a client and a fog node in real time, realizing privacy protection and efficient personalized federated learning, and ensuring the privacy and security of user data in the training process are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field related to digital data processing, and in particular to a personalized federated learning method and system for heterogeneous multi-source industrial Internet. Background Art

[0002] With the rapid development of information technology and mobile Internet, the era of big data has arrived. The problem of data heterogeneity has become increasingly prominent in many fields such as personalized recommendations and intelligent transportation, seriously affecting the accuracy of data analysis and model training. In conventional machine learning and deep learning, data usually needs to be centrally processed and analyzed on central servers. However, with the continuous growth of data volume and the improvement of privacy protection awareness, the problem of data heterogeneity has gradually become prominent.

[0003] At present, most federated learning research focuses on federated learning under homogeneous data, which is mainly divided into converting heterogeneous data into homogeneous data through technologies such as data preprocessing and feature extraction, but some information may be lost during the preprocessing process.

[0004] To sum up, the existing technology has technical problems such as frequent connection and exit of devices in the industrial Internet, and differences in data volume leading to excessive load on some nodes, causing repeated migration of clients and nodes, and further aggravating network volatility and waste of computing resources. Summary of the Invention

[0005] This application provides a personalized federated learning system for heterogeneous multi-source industrial Internet, aiming to solve the technical problems in the existing technology of frequent connection and exit of devices in the industrial Internet, excessive load on some nodes caused by differences in data volume, repeated migration of clients and nodes, and further aggravated network volatility and waste of computing resources.

[0006] In view of the above problems, the technical solution to implement this application is:

[0007] On the one hand, the present application provides a personalized federated learning method for a heterogeneous multi-source industrial Internet, wherein the method includes: connecting to a client, evaluating the load, network topology delay and data modality similarity, and generating a dynamic association mapping table between the client and the fog node; in the federated learning initialization phase, in combination with the dynamic association mapping table, deploying a multi-input function encryption protocol to perform hierarchical verifiable encryption on the local model gradient update of the client; at the same time, constructing a priority queue based on the local model update frequency and data density, and the priority queue is used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters; in the heterogeneous multi-source industrial Internet, setting a fog node edge caching mechanism according to the priority queue, and the fog node edge caching mechanism is used to coordinate the client to perform distributed iterative optimization.

[0008] On the other hand, the present application provides a personalized federated learning system for a heterogeneous multi-source industrial Internet, wherein the system includes: a dynamic association evaluation module for connecting to a client, evaluating load, network topology delay and data modality similarity, and generating a dynamic association mapping table between the client and the fog node; a protocol encryption module for deploying a multi-input function encryption protocol in combination with the dynamic association mapping table during the federated learning initialization phase, and performing hierarchical verifiable encryption on the local model gradient update of the client; a queue construction module for simultaneously constructing a priority queue based on the local model update frequency and data density, the priority queue being used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters; an iterative optimization module for setting a fog node edge caching mechanism in the heterogeneous multi-source industrial Internet according to the priority queue, the fog node edge caching mechanism being used to coordinate the client to perform distributed iterative optimization.

[0009] In summary, the one or more technical solutions provided in this application realize the analysis of the association between the client and the fog node, extract the knowledge basis vector shared across devices, reduce the impact of model heterogeneity on aggregation, dynamically schedule the local aggregation of high-frequency parameters and the cloud synchronization of low-frequency parameters, and optimize the association weights between the client and the fog node in real time, realizing privacy protection and efficient personalized federated learning, and ensuring the privacy and security of user data during the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A flowchart of a personalized federated learning method for heterogeneous multi-source industrial Internet is provided for this application;

[0011] Figure 2 A structural diagram of a personalized federated learning system for heterogeneous multi-source industrial Internet is provided for this application.

[0012] Description of reference numerals: dynamic association evaluation module M100, protocol encryption module M200, queue construction module M300, iterative optimization module M400. DETAILED DESCRIPTION

[0013] Example 1

[0014] The present application is described in detail below with reference to the accompanying drawings. Figure 1 As shown, the present application provides a personalized federated learning method for heterogeneous multi-source industrial Internet, wherein the method includes:

[0015] S1: Connect to the client, evaluate the load, network topology delay and data modality similarity, and generate a dynamic association mapping table between the client and the fog node; S2: In the federated learning initialization phase, combine the dynamic association mapping table and deploy a multi-input function encryption protocol to perform hierarchical verifiable encryption on the client's local model gradient update.

[0016] It should be noted that in the Industrial Internet, different participants have different data distribution and computing capabilities, which leads to heterogeneity in the models and training methods used. For example, different participants may have differences in model architecture, hyperparameter settings, data preprocessing methods, model parameters, data distribution, and computing resources, which makes federated learning difficult to deal with data imbalance.

[0017] Specifically, various clients are connected. Clients refer to various intelligent devices distributed at the edge of the network, such as sensors, controllers, terminal acquisition devices, etc. These devices generate a large amount of data with heterogeneous characteristics; evaluating load refers to measuring the computing power, storage capacity and other resource usage of these clients and fog nodes in the network; network topology delay refers to the time delay experienced when data is transmitted from the client to the fog node or from one fog node to another fog node in the network, which reflects the performance and stability of the network connection; data modal similarity is used to measure the similarity of different client data in the feature space. For example, temperature, pressure, vibration and other data collected by different types of sensors may have different degrees of similarity in statistical characteristics or physical sense; by comprehensively considering these factors, a dynamic association mapping table between clients and fog nodes is generated. The dynamic association mapping table is used to record the association relationship between clients and fog nodes to facilitate subsequent data transmission and task allocation.

[0018] During the initialization phase of federated learning, a multi-input function encryption protocol is deployed in combination with the dynamic association mapping table. Multi-input function encryption is an encryption technology that allows multiple data providers to encrypt their own data. The decryptor can then directly perform a specified function operation on the encrypted data and obtain the result without decrypting a single data item. Layered verifiable encryption uses a layered encryption strategy when encrypting the client's local model gradient update amount, and adds a verification mechanism to ensure that the encrypted data can maintain privacy during transmission and processing, and can be verified by the recipient for its integrity and validity, preventing malicious tampering or injection of erroneous data.

[0019] First, connect to the client and evaluate the load, network topology delay, and data modality similarity. By deploying monitoring agents on the client and fog nodes, resource usage, network status, and data feature information are collected in real time. The load is evaluated by monitoring indicators such as the CPU usage, memory usage, and network bandwidth of these sensors. At the same time, network diagnostic tools are used to measure the transmission delay from the sensor to the fog node to measure the network topology delay.

[0020] Through data preprocessing and feature extraction algorithms, the similarity between these sensor data is calculated. Specifically, cosine similarity is used to calculate the similarity between temperature sensor and pressure sensor data in time series. Based on these evaluation results, an association mapping table is dynamically generated. Each sensor is reasonably allocated to each fog node so that the number of clients connected to each fog node is balanced, and clients with low network latency and similar data modalities are connected to the same fog node as much as possible.

[0021] During the federated learning initialization phase, a multi-input function encryption protocol is deployed in combination with a dynamic association mapping table to perform hierarchical verifiable encryption on the local model gradient updates. After each client trains the model locally, it generates a gradient update vector. These gradient vectors contain key information about the model improvement, but direct transmission may leak privacy. Through the multi-input function encryption protocol, the gradient vectors of multiple clients can be encrypted, allowing the server or fog node to directly perform aggregation operations on these encrypted gradients without decrypting individual gradients.

[0022] At the same time, a layered encryption strategy is adopted. Specifically, according to the importance or sensitivity of the gradient, encryption algorithms of different strengths are applied to different parts of the gradient to improve encryption efficiency. After adding a verification mechanism, such as using digital signatures or hash value verification, the receiver can verify the integrity and authenticity of the encrypted gradient, thereby effectively aggregating the gradient while ensuring privacy and improving model training efficiency.

[0023] S3: At the same time, a priority queue is constructed based on the local model update frequency and data density. The priority queue is used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters; S4: In the heterogeneous multi-source industrial Internet, a fog node edge caching mechanism is set according to the priority queue. The fog node edge caching mechanism is used to coordinate the client to perform distributed iterative optimization.

[0024] Specifically, in the federated learning process, the local model update frequency refers to the rate of iterative optimization of the client's local model parameters, that is, the number of times the client adjusts the model parameters within a certain period of time, reflecting the speed of the model's learning progress on local data; data density refers to the compactness of the distribution of client data in the feature space, showing the concentration of data points within a certain range. Areas with high data density contain more representative patterns and features; priority queues are used to dynamically arrange the transmission and processing order of different parameters; dynamic scheduling refers to the flexible adjustment of parameter transmission and aggregation methods according to the real-time status of the system and the instructions of the priority queue to adapt to different loads and network conditions.

[0025] High-frequency parameter local aggregation refers to the aggregation of frequently updated model parameters at fog nodes close to the client, reducing the frequent transmission of these parameters between the cloud and the local; low-frequency parameter cloud synchronization refers to sending those parameters with less frequent updates to the cloud for unified synchronization and processing; the fog node edge caching mechanism is a caching strategy set in the edge computing environment of the fog node, which is used to temporarily store model parameters and intermediate calculation results in order to quickly respond to client requests, reduce data transmission delays, and improve the overall efficiency of the system; distributed iterative optimization refers to the collaborative iterative update of model parameters between multiple clients and fog nodes, and improves the speed and effect of model training through parallel computing and information sharing.

[0026] In the process of federated learning, it is first necessary to build a priority queue based on the local model update frequency and data density. Each client has its own local data and model. By monitoring the local model update frequency and data density of each client, a priority can be assigned to the parameters of each client. Preferably, for clients with a high local model update frequency (30% of clients have a frequency higher than the system average) and parameters with a high data density (40% of parameters have a data density higher than the threshold), these parameters will be given a higher priority. The priority queue will be dynamically adjusted according to the real-time monitoring data. The priority queue plays a key role in resource allocation and scheduling in the entire federated learning solution.

[0027] High-frequency parameters contain data information that allows the model to quickly adapt to local data, while parameters with high data density have a greater impact on the model's generalization ability. By prioritizing local aggregation of these parameters, these key information can be quickly integrated at the fog node, reducing frequent transmission between the cloud and local, thereby reducing communication overhead. Local aggregation reduces the number of times parameters are transmitted in the cloud. At the same time, cloud synchronization of low-frequency parameters ensures the consistency and stability of the global model. Low-frequency parameters update slowly, and synchronization in the cloud can ensure that all clients obtain these globally consistent parameters.

[0028] In a heterogeneous multi-source industrial Internet environment, a fog node edge caching mechanism is set up according to the priority queue; the fog node edge cache allocates cache resources according to the parameter priority in the priority queue. High-priority parameters will be cached first, and the number of cached copies may be more to ensure fast access and high availability; when the client needs to perform distributed iterative optimization, it first obtains the required parameters from the fog node edge cache. If the required high-priority parameters exist in the cache, the client can quickly obtain and start local iterative optimization, thereby reducing waiting time; this not only improves the efficiency of data transmission and model training, but also enhances the robustness and reliability of the system through reasonable caching strategies, ensuring that federated learning can be carried out efficiently and stably in a complex industrial Internet environment.

[0029] Furthermore, the load, network topology delay, and data modality similarity are evaluated to generate a dynamic association mapping table between the client and the fog node. The method of this application also includes:

[0030] A reinforcement learning agent is embedded in the data-aware grouping, which is used to optimize the association weight between the client and the fog node in real time. At the same time, the sharding storage strategy of the encryption key is dynamically adjusted according to the channel quality indicator. When it is detected that the fog node load exceeds the load threshold, a load migration decision tree based on topology reconstruction is triggered.

[0031] Specifically, in the data perception and grouping stage of the federated learning system, a reinforcement learning agent is embedded. The reinforcement learning agent can learn the optimal strategy in continuous interaction with the environment, and is used to optimize the association weight between the client and the fog node in real time. The association weight refers to a quantitative indicator that measures the closeness and importance of the connection between the client and the fog node, which can affect the priority of data transmission and resource allocation.

[0032] At the same time, based on the channel quality indicators, which are used to measure the stability, reliability and transmission efficiency of network communication links, such as parameters such as signal strength, bit error rate and bandwidth, the encryption key fragmentation storage strategy is dynamically adjusted, that is, the encryption key is divided into multiple fragments, and these fragments are reasonably allocated to different storage locations according to changes in channel quality to ensure the security and availability of the key.

[0033] When it is detected that the fog node load exceeds the load threshold, the load threshold is a pre-set standard for measuring the maximum load that the fog node can withstand. It is used to determine whether load migration is needed and trigger the load migration decision tree based on topology reconstruction. The load migration decision tree is used to determine whether part of the load will be migrated to other nodes to achieve load balancing based on the network topology and node load when the fog node load is too high.

[0034] In the federated learning process, data perception and grouping are key links. Each client generates a large amount of data, which needs to be reasonably distributed to different fog nodes for processing. After embedding the reinforcement learning agent, the interaction between the client and the fog node is monitored and learned in real time. The reinforcement learning agent continuously adjusts the association weight through trial and error. The reinforcement learning agent gradually converges the association weight to the optimal value based on historical data and reward mechanisms (such as transmission success rate, response time, etc. as reward signals). For those clients with large data volumes, frequent transmissions and good network conditions, the association weight will be increased, and vice versa, it will be reduced, making data transmission more efficient and reducing invalid transmission and resource waste.

[0035] At the same time, the shard storage strategy of the encryption key is dynamically adjusted according to the channel quality indicators. Common channel quality indicators include signal strength, bit error rate and bandwidth. When the channel quality is good (for example, signal strength is higher than -70dBm, bit error rate is lower than 1%, and bandwidth is higher than 10Mbps), the encryption key shards can be stored in a dispersed manner on multiple fog nodes, with each fog node storing 2-3 shards; when the channel quality deteriorates (for example, signal strength is lower than -80dBm, bit error rate is higher than 2%, and bandwidth is lower than 5Mbps), the key shards will be concentrated on a few stable connected fog nodes, with each fog node storing 4-5 shards to ensure the availability and security of the key. This dynamic adjustment strategy improves the adaptability and reliability of the system while ensuring data security.

[0036] When the fog node load exceeds the load threshold, a load migration decision tree based on topology reconstruction is triggered. The load migration decision tree migrates the load based on the network topology and node load. Furthermore, based on the load of adjacent nodes, some of the load is migrated to adjacent fog nodes with lower loads. This process is performed dynamically within the network, for example, the load is checked every 5 minutes, and migration is initiated once the threshold is exceeded. This mechanism effectively avoids single-point overload, improves overall performance and stability, and ensures the smooth progress of the federated learning process. These steps not only enhance the security and efficiency of data transmission, but also improve the robustness and scalability of the system through dynamic adjustment and load balancing, enabling it to adapt to the complex and changing industrial Internet environment.

[0037] Furthermore, the present application method also includes:

[0038] An attribute-based access control strategy is introduced to define fine-grained model parameter access rights. After verifying the legitimacy of the client identity based on the attribute-based access control strategy, a dynamic threshold decryption unit is activated to perform homomorphic aggregation on the encrypted gradient.

[0039] Specifically, the attribute-based access control strategy is an attribute-based access control mechanism that determines the access rights to model parameters according to the attributes of the client or user (such as device type, user role, data category, etc.). It can achieve fine-grained control of access to model parameters, ensuring that only clients with appropriate attributes can access and use specific model parameters; the dynamic threshold decryption unit can dynamically adjust the conditions required for decryption (such as how many key shards are required for decryption) according to the system status and security requirements. After verifying the legitimacy of the client identity, the unit is activated to decrypt the encrypted gradient to ensure that only legitimate clients can participate in the subsequent homomorphic aggregation process; homomorphic aggregation processing refers to performing aggregation operations directly on encrypted data, so that the aggregation results can be obtained without decrypting the data, thereby protecting the privacy of the data while achieving effective model updates and optimization.

[0040] In the federated learning system, an attribute-based access control strategy is introduced to ensure the security and privacy of model parameters. Each fog node is responsible for managing a certain number of clients, each with different attributes, such as sensor type (temperature, pressure, vibration, etc.), region (workshop A, workshop B, etc.), and data sensitivity level (high, medium, low). An attribute-based access control strategy is defined to set access rules for different model parameters. For example, it is stipulated that only clients with the "temperature sensor" attribute and belonging to the "workshop A" region can access and use temperature-related model parameters, while clients with the "pressure sensor" attribute and belonging to the "workshop B" region can only access pressure-related parameters.

[0041] Based on the attribute-based access control strategy, when a client requests to participate in federated learning, its identity legitimacy is first verified; the verification process includes checking the client's certificate, key, or interacting with the identity authentication protocol; after the verification is passed, the dynamic threshold decryption unit is activated, which dynamically determines the conditions required for decryption based on the current security requirements and system status; once the dynamic threshold decryption unit is activated and the encrypted gradients are decrypted, these gradients will be homomorphically aggregated; through homomorphic aggregation, the aggregation result can be directly calculated without decrypting each individual gradient to obtain the aggregated gradient vector, thereby not only protecting the privacy of each client's data but also improving the efficiency of model updates.

[0042] The attribute-based access control strategy ensures that only clients with appropriate attributes can participate in model training, preventing unauthorized access and data leakage; the dynamic threshold decryption unit and homomorphic aggregation processing achieve efficient data aggregation and model updates while ensuring data security, allowing federated learning to proceed smoothly while protecting privacy; this not only improves the overall security of the system, but also enhances the reliability and accuracy of model training, providing a strong guarantee for building a secure and efficient federated learning environment.

[0043] Furthermore, the present application method also includes:

[0044] A collaborative cache network across fog nodes is constructed, and distributed redundant storage is performed based on the importance scores of model parameters. Based on the collaborative cache network, a cache replacement chain is set up, and the future request distribution is predicted by combining parameter freshness and access frequency. When regional network congestion is detected, a dynamic cache resource allocation mechanism is activated.

[0045] Specifically, a collaborative cache network across fog nodes refers to a collaborative cache architecture built between multiple fog nodes, allowing different fog nodes to share cache resources, improving cache utilization and data availability; model parameter importance score refers to an indicator that evaluates the impact of model parameters on overall model performance, and is used to determine which parameters need priority storage and protection; distributed redundant storage refers to the storage of data copies on multiple different storage nodes to improve data reliability and fault tolerance.

[0046] The cache replacement chain is a data structure used to manage cache resources. It defines which cache contents will be replaced to make room for new data when cache resources are insufficient. Parameter freshness refers to the update time or version information of the model parameters, reflecting the timeliness and effectiveness of the parameters. Access frequency refers to how frequently a certain data or resource is accessed, which can be used to predict possible future access needs. Future request distribution refers to the number and pattern of access requests that each data or resource may receive in the future based on historical access data. The dynamic allocation mechanism of cache resources refers to the flexible adjustment of the allocation method of cache resources according to the current network conditions, load conditions and data access needs to improve the performance and efficiency of the system.

[0047] When building a collaborative caching network across fog nodes, the model parameters must first be scored for importance. The importance score of each parameter can be determined by calculating its sensitivity to the model output. Based on these scores, the parameters are distributed on different fog nodes for redundant storage to ensure high availability and reliability of key parameters.

[0048] According to the collaborative cache network, a cache replacement chain is set up. The cache replacement chain defines how to select the data to be replaced when the cache space is insufficient. For example, a replacement chain based on algorithms such as first-in-first-out (FIFO), least recently used (LRU) or adaptive replacement cache (ARC) can be used. At the same time, the parameter freshness and access frequency are combined to predict the future request distribution. Specifically, there is an access data record with a time window of 1 hour, and the access frequency and update time of each parameter are counted. For example, parameter A has been accessed 100 times in the past 1 hour, and the last update time was 5 minutes ago; parameter B has been accessed 20 times, and the last update time was 30 minutes ago. Furthermore, it can be predicted that the access request for parameter A may still be high in the next 15 minutes, while the request for parameter B may decrease. Therefore, parameter A is retained in the cache, and parameter B is marked as replaceable.

[0049] When regional network congestion is detected, for example, when the network bandwidth utilization in a certain area exceeds 80% and the data transmission delay exceeds a preset threshold (such as 100 milliseconds), the dynamic allocation mechanism of cache resources is activated. At this time, based on the predicted future request distribution, parameters with high access frequency and high freshness are preferentially cached to local fog nodes in the network congested area, reducing the need to obtain data from other areas, thereby alleviating network congestion and improving data access efficiency.

[0050] By building a collaborative cache network and dynamic cache management mechanism across fog nodes, the storage efficiency and access speed of model parameters are improved, ensuring that data can be accessed and used quickly and reliably in complex industrial Internet environments. This not only improves the speed and effectiveness of model training, but also enhances the overall performance and stability of the system, enabling it to better adapt to changing network conditions and load requirements.

[0051] Furthermore, the present application method also includes:

[0052] During the local training phase on the client, a lightweight feature distiller is deployed to extract cross-device shared knowledge base vectors. By contrasting the learning loss function, the semantic alignment between the personalized model and the global model is constrained, and adversarial samples are generated using the cross-device shared knowledge base vectors.

[0053] Specifically, during the local training phase on the client, a lightweight feature distiller is deployed. Feature distillation is a knowledge extraction process that extracts the knowledge of a complex model or multiple models into a smaller model to achieve more efficient model performance. The lightweight feature distiller is a feature distillation tool optimized for resource-constrained devices. It can reduce computing resource consumption while ensuring feature extraction efficiency. Its function is to extract cross-device shared knowledge base vectors, which are condensed representations of the common features of data from multiple devices, reflecting the common knowledge between data from different devices, and can be used to enhance the generalization ability and cross-device adaptability of the model.

[0054] The semantic alignment between the personalized model and the global model is constrained by the contrastive learning loss function. The contrastive learning loss function refers to the function used to train the model to distinguish the similarities and differences of different data features. It can effectively improve the model's understanding of the deep semantics of the data and ensure that the personalized model does not deviate from the common semantic representation of the global model while learning local data features, so that the personalized model can meet local needs while being consistent with the global model; adversarial samples are generated using cross-device shared knowledge basis vectors. Adversarial samples refer to data samples generated by adding small perturbations to the original data, which can cause the model to make incorrect judgments and are used to test and improve the robustness of the model. Using cross-device shared knowledge basis vectors to generate adversarial samples can enhance the model's adaptability to cross-device data changes and improve the model's generalization and stability.

[0055] During local training on the client, a lightweight feature distiller is first deployed. The lightweight feature distiller runs on each client to extract key features from local data. The feature distiller compresses the data into a three-dimensional cross-device shared knowledge base vector through a self-supervised learning mechanism. This process retains the most important features of the data. The cross-device shared knowledge base vector is easy to transmit and store, and it also condenses the common knowledge of data from different devices.

[0056] Then, the contrastive learning loss function is used to constrain the semantic alignment between the personalized model and the global model. Through the contrastive learning loss function, the similarity between the local sample features and the global sample features is determined, and the parameters of the personalized model are adjusted to make its semantic representation close to that of the global model, avoiding overfitting of local data.

[0057] At the same time, adversarial samples are generated using cross-device shared knowledge base vectors. Specifically, the average value of the shared knowledge base vector is μ; by adding a Gaussian distribution (mean 0, standard deviation ε) to the vector σ) noise to generate adversarial samples; these adversarial samples are fed into the model for training, improving the model's adaptability and robustness to data changes. In these steps, lightweight feature distillation and adversarial sample generation improve the efficiency and generalization capabilities of the client's local model. It also enhances the model's understanding and adaptability to global semantics, ensuring that personalized models meet local needs while remaining consistent with the global model. This helps build an efficient, stable, and highly adaptable federated learning system to meet the complex and ever-changing application requirements of the Industrial Internet.

[0058] Furthermore, the present application method also includes:

[0059] Obtain the entropy contribution of each client parameter update to the global model; in low-bandwidth scenarios, dynamically adjust the federation aggregation frequency based on the entropy contribution, enable real-time streaming aggregation for high-entropy clients, activate the difference compression parameter, and reconstruct the gradient transmission tensor dimension.

[0060] Specifically, the contribution entropy value is a quantitative indicator used to measure the contribution of each client parameter update to the global model. Based on the concept of entropy in information theory, it reflects the amount of information brought by the client parameter update; the federated aggregation frequency refers to the frequency of the global model aggregating client parameter updates during the federated learning process, usually in units of iterations or time, which determines the intensity of communication between the client and the server; real-time streaming aggregation can continuously aggregate the parameter updates uploaded by the client without waiting for all clients to complete the update before performing batch aggregation, thereby speeding up the model update speed; the difference compression parameter is an algorithm parameter used to optimize data transmission. It reduces the amount of transmitted data by comparing the differences between new and old data, and only transmits the changed part instead of the entire data set; reconstructing the gradient transmission tensor dimension refers to reorganizing and adjusting the transmission format of the gradient data to improve data transmission efficiency and the parallelism of model updates.

[0061] During the federated learning process, it is necessary to obtain the contribution entropy value of each client parameter update to the global model; after each model update, the contribution entropy value is determined by calculating the difference between each client parameter update and the global model parameters. The contribution entropy value reflects the relative contribution of each client to the global model optimization.

[0062] In low-bandwidth scenarios, all clients uploading parameters at the same time may cause network congestion; the federation aggregation frequency is dynamically adjusted according to the contribution entropy value; for high-entropy clients (such as clients with a contribution entropy value greater than 0.5), real-time streaming aggregation is enabled, allowing these clients to upload parameter updates at a higher frequency (such as every 30 seconds) and perform aggregation processing immediately; at the same time, the difference compression parameter is activated to reduce the amount of gradient transmission data. In addition, the gradient transmission tensor dimension is reconstructed, and the original high-dimensional sparse gradient vector (such as a size of 10000×1) is reorganized into multiple low-dimensional dense vectors (such as 10 1000×1 vectors) to improve transmission efficiency and parallel processing capabilities.

[0063] In the above steps, through the quantitative evaluation of the contribution entropy value and targeted transmission optimization strategies, it is ensured that critical client updates can be processed in a timely manner under low bandwidth constraints. This not only improves the speed and efficiency of model training, but also enhances the system's adaptability to changes in network conditions, ensuring that the federated learning process can proceed smoothly even in a resource-constrained environment, and ultimately improving the performance and accuracy of the global model.

[0064] Furthermore, the present application method also includes:

[0065] A hash fingerprint chain that records the parameter update path triggers a historical version rollback when malicious parameter injection is detected; at the same time, a security verification protocol based on the contributed entropy value is triggered to force the encryption key to be updated and isolate suspicious client nodes.

[0066] Specifically, the hash fingerprint chain is a data structure generated based on the hash algorithm, which is used to record the parameter update path. By linking the data summary (hash value) of each parameter update, an unalterable record chain is formed, which can effectively track the historical trajectory of parameter updates and ensure the integrity and traceability of the data; the security verification protocol is used to verify the rules and processes for the security of various operations and data in the system, and define the measures to be taken when potential security threats are detected to ensure the normal operation of the system and data security; historical version rollback means that when the system detects an abnormal or error state, it can restore the system state to a previous stable time point to ensure that the system can quickly recover from the error and continue to operate normally.

[0067] Forced encryption key update means that when the system detects potential security risks, it will be forced to use new encryption keys to replace old keys to prevent potential key leakage or malicious use; isolating suspicious client nodes means temporarily or permanently removing client nodes that are detected to have possible malicious behavior from the system to prevent malicious behavior from affecting the entire system and ensure the security and reliability of the system.

[0068] During the federated learning process, the parameter update path of each client is recorded through the hash fingerprint chain. Each time each client uploads a parameter update, the hash value of the update is calculated and added to the hash fingerprint chain. This forms a chain structure, allowing the system to clearly track the parameter update history of each client and ensure the transparency and integrity of the update path.

[0069] When the system detects malicious parameter injection, for example, when it finds that the parameters uploaded by a client are significantly different from the normal distribution (such as parameter values outside the normal range or abnormal update frequency), it triggers a rollback of the historical version; finds the most recent stable version based on the hash fingerprint chain to ensure that the global model is not contaminated. At the same time, it triggers a security verification protocol based on the contribution entropy value, which is used here as an indicator to measure the credibility of the client; for those clients with low contribution entropy values (for example, less than 0.3), the system forces the encryption key to be updated to prevent potential key leakage. In addition, the system isolates suspicious client nodes and temporarily removes them from the federated learning process, restricting their participation in model training until the security check and repair are completed.

[0070] In the above steps, the system's security and stability are enhanced through the hash fingerprint chain and security verification protocol. This not only effectively prevents malicious parameter injection attacks, but also promptly restores the system to a safe state, ensuring the reliability of the federated learning process and the integrity of the data. This helps to build a secure and trustworthy federated learning environment, enabling the system to respond quickly and resume normal operation when faced with potential security threats.

[0071] Furthermore, the present application method also includes:

[0072] Traverse each client and build a device resource profile, which is used to map storage capacity indicators and energy consumption constraint indicators; predict federation participation based on the device resource profile and assign training task priorities; dynamically bind heterogeneous task flows in the priority queue based on the training task priorities, and activate a progressive model pruning strategy.

[0073] Specifically, the device resource profile is a data model that comprehensively describes the characteristics of device resources. By collecting and analyzing various resource indicators of the device, such as storage capacity, energy consumption, computing power, etc., a portrait that comprehensively reflects the status of device resources is generated; the storage capacity indicator refers to the size of the space available for data storage on the device, usually in bytes, which limits the amount of data and model size that the device can store; the energy consumption constraint indicator refers to the limit on energy consumption of the device during operation, usually in watts or joules, which affects the device's operating time and task scheduling; federation participation refers to the willingness and ability of the device to participate in the federated learning process, reflecting the device's activity and contribution potential in the federated learning system; the training task priority is a classification of the importance and urgency of the training task based on the device resource situation and federated learning needs, which is used to rationally allocate computing resources and task scheduling.

[0074] Dynamic binding refers to the process of flexibly associating tasks with heterogeneous task streams in the priority queue based on the training task priority, ensuring that high-priority tasks can be processed in a timely manner; heterogeneous task streams refer to a set of tasks of various types and requirements generated in the federated learning system due to differences in device type, data characteristics and computing power; progressive model pruning strategy is a model optimization technology that reduces the size and computational complexity of the model by gradually removing unimportant parameters or connections in the model, while trying to maintain the performance of the model to adapt to the resource constraints of different devices.

[0075] In the federated learning system, it is first necessary to traverse each client and build a device resource profile. For example, device types include sensors, controllers, and small servers. The system collects the storage capacity indicators of each device (such as the sensor storage capacity is 1GB, the controller is 2GB, and the server is 10GB) and energy consumption constraint indicators (such as the sensor energy consumption is 5W, the controller is 10W, and the server is 50W), as well as other resource information. By analyzing this data, a device resource profile is generated for each device. For example, the sensor device resource profile shows that it has limited storage and low energy consumption, which is suitable for lightweight tasks; the server device profile shows that it has rich resources and is suitable for complex computing tasks.

[0076] Predict federated participation based on device resource profiles, analyze various indicators in the device resource profiles, and predict the device's willingness and ability to participate in federated learning. Assign training task priorities to each device based on participation and other factors (such as network connection stability, historical contributions, etc.). Devices with high participation and high resource availability are assigned high-priority tasks, while devices with low participation and limited resources are assigned low-priority tasks.

[0077] According to the training task priority, the heterogeneous task flows in the priority queue are dynamically bound. The system maintains a priority queue, which contains various heterogeneous task flows, such as model training, data preprocessing, parameter update, etc.; when tasks need to be assigned, the appropriate task flow is assigned to the corresponding device according to the training task priority of the device; at the same time, the progressive model pruning strategy is activated. For resource-constrained devices, progressive model pruning is applied according to their device resource profile and training task priority to adapt to storage limitations while trying to maintain model performance. The progressive model pruning strategy is achieved by removing unimportant weights or neurons in the model, involving multiple iterations to evaluate model performance and pruning effects.

[0078] In the above steps, through the construction and application of device resource portraits, accurate assessment of device resources and reasonable allocation of tasks are achieved, which improves the overall efficiency of the system and ensures that high-priority tasks are given priority. Through the progressive model pruning strategy, the model can adapt to the resource limitations of different devices, enhancing the compatibility and scalability of the system, and helping to build an efficient, flexible and adaptable federated learning environment to meet the diverse equipment and task requirements in the Industrial Internet.

[0079] Furthermore, in the federated learning initialization phase, in combination with the dynamic association mapping table, a multi-input function encryption protocol is deployed, and the method of the present application further includes:

[0080] Deploy a federated learning performance monitoring plug-in, determine the hyperparameter combination through multi-objective optimization, and generate an adaptive parameter adjustment strategy matrix, wherein the federated learning performance monitoring plug-in is used to monitor the model convergence speed and privacy leakage risk; based on the adaptive parameter adjustment strategy matrix, dynamically adjust the learning rate and batch size of the client's local training, and inject the intensity parameter of the differential privacy noise; at the same time, according to the real-time feedback of the data modality similarity and network topology delay, allocate the privacy protection budget of each fog node proportionally; if the consumption rate of the privacy protection budget does not meet the expected threshold, trigger the compensation mechanism of elastically scaling the training batch size, and reconstruct the noise injection distribution of the differential privacy noise based on the gradient sparsity.

[0081] Specifically, the federated learning performance monitoring plug-in is used to monitor key performance indicators in the federated learning process in real time, such as model convergence speed and privacy leakage risk; multi-objective optimization is a method of finding the optimal solution between multiple conflicting objectives, and is used to determine the hyperparameter combination to balance multiple requirements such as model performance and privacy protection; the adaptive parameter adjustment strategy matrix is a matrix structure that stores multiple parameter adjustment strategies, and dynamically selects the appropriate parameter adjustment strategy combination according to different system states and performance indicators.

[0082] The intensity parameter of differential privacy noise is used to control the amount of noise added to the data to protect data privacy. The greater the noise, the stronger the privacy protection, but it may affect data accuracy and model performance; data modal similarity reflects the degree of similarity between different data sets in the feature space and is used to guide the allocation of privacy protection budget; network topology delay refers to the communication delay between nodes in the network, which affects the efficiency of data transmission; privacy protection budget is a resource allocation concept used to measure the system resources consumed by privacy protection measures, such as computing resources and time; the compensation mechanism for elastically scaling training batch size is a dynamic adjustment mechanism used to compensate for performance loss by adjusting the training batch size when the privacy protection budget is insufficient; gradient sparsity reflects the proportion of non-zero elements in the gradient vector and is used to optimize the injection distribution of differential privacy noise.

[0083] Deploy a federated learning performance monitoring plug-in. Specifically, the performance monitoring plug-in will collect indicators such as model convergence speed and privacy leakage risk in real time. The model convergence speed is measured by calculating the rate of change of the loss function, and the privacy leakage risk can be determined by evaluating the strength of the encryption key and the data access log. Furthermore, the hyperparameter combination is determined through multi-objective optimization, and an adaptive parameter adjustment strategy matrix is generated. Preferably, the three hyperparameters of learning rate, batch size, and differential privacy noise intensity need to be optimized. A multi-objective optimization algorithm (such as NSGA-II) is used to find a balance between model convergence speed and privacy protection strength. After multiple iterative optimizations, an adaptive parameter adjustment strategy matrix is generated, which contains the optimal hyperparameter combination under different system states.

[0084] Based on an adaptive parameter adjustment strategy matrix, the learning rate and batch size of local client training are dynamically adjusted, and the intensity parameters of differential privacy noise are injected. For example, when the performance monitoring plug-in detects slow model convergence, the strategy matrix is used to increase the client's learning rate, reduce the batch size, and lower the intensity of differential privacy noise to accelerate model convergence. At the same time, the privacy protection budget of each fog node is proportionally allocated based on real-time feedback on data modality similarity and network topology latency.

[0085] If the consumption rate of the privacy protection budget does not meet the expected threshold, a compensation mechanism for elastically scaling the training batch size is triggered. Preferably, the training batch size is elastically scaled to increase the data processing capacity and improve the model training efficiency. At the same time, the noise injection distribution of the differential privacy noise is reconstructed based on the gradient sparsity, and the noise is injected into the non-zero element part to improve the effectiveness of the noise, thereby maintaining the strength of privacy protection while reducing the impact on model performance.

[0086] In the above steps, the efficiency of model training and the reliability of privacy protection are ensured through performance monitoring and dynamic parameter adjustment. The training parameters are automatically adjusted according to the system status and performance indicators. Through reasonable privacy protection budget allocation and noise injection optimization, the overall performance and security of the system are improved, and an efficient, secure and adaptable federated learning environment is built to meet the complex and changing application needs of the Industrial Internet.

[0087] In summary, the beneficial effects of the embodiments of the present application are:

[0088] Due to the adoption of connecting clients, evaluating load, network topology delay and data modality similarity, a dynamic association mapping table between clients and fog nodes is generated; in the federated learning initialization phase, combined with the dynamic association mapping table, a multi-input function encryption protocol is deployed to perform hierarchical verifiable encryption on the client's local model gradient update amount; at the same time, a priority queue is constructed based on the local model update frequency and data density, and the priority queue is used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters; in the heterogeneous multi-source industrial Internet, a fog node edge caching mechanism is set according to the priority queue, and the fog node edge caching mechanism is used to coordinate the client to perform distributed iterative optimization. This application provides a personalized federated learning method and system for heterogeneous multi-source industrial Internet, analyzes the association between clients and fog nodes, extracts cross-device shared knowledge basis vectors, reduces the impact of model heterogeneity on aggregation, dynamically schedules local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters, and optimizes the association weights between clients and fog nodes in real time, thereby achieving privacy protection and efficient personalized federated learning, and ensuring the privacy and security of user data during training.

[0089] Example 2

[0090] Based on the same inventive concept as the personalized federated learning method for heterogeneous multi-source industrial Internet in the above embodiment, Figure 2 As shown, an embodiment of the present application provides a personalized federated learning system for a heterogeneous multi-source industrial Internet, wherein the system includes:

[0091] The dynamic association evaluation module M100 is used to connect to the client, evaluate the load, network topology delay and data modality similarity, and generate a dynamic association mapping table between the client and the fog node.

[0092] The protocol encryption module M200 is used to deploy a multi-input function encryption protocol in combination with the dynamic association mapping table during the federated learning initialization phase to perform hierarchical verifiable encryption on the local model gradient update amount of the client.

[0093] The queue construction module M300 is used to, at the same time, construct a priority queue based on the local model update frequency and data density, and the priority queue is used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters.

[0094] The iterative optimization module M400 is used to set a fog node edge caching mechanism in a heterogeneous multi-source industrial Internet according to the priority queue, and the fog node edge caching mechanism is used to coordinate the client to perform distributed iterative optimization.

[0095] Furthermore, the dynamic association evaluation module M100 is further configured to execute the following method:

[0096] A reinforcement learning agent is embedded in the data-aware grouping, which is used to optimize the association weight between the client and the fog node in real time. At the same time, the sharding storage strategy of the encryption key is dynamically adjusted according to the channel quality indicator. When it is detected that the fog node load exceeds the load threshold, a load migration decision tree based on topology reconstruction is triggered.

[0097] Furthermore, the dynamic association evaluation module M100 is further configured to execute the following method:

[0098] An attribute-based access control strategy is introduced to define fine-grained model parameter access rights. After verifying the legitimacy of the client identity based on the attribute-based access control strategy, a dynamic threshold decryption unit is activated to perform homomorphic aggregation on the encrypted gradient.

[0099] Furthermore, the personalized federated learning system for heterogeneous multi-source industrial Internet is also used to perform the following method:

[0100] A collaborative cache network across fog nodes is constructed, and distributed redundant storage is performed based on the importance scores of model parameters. Based on the collaborative cache network, a cache replacement chain is set up, and the future request distribution is predicted by combining parameter freshness and access frequency. When regional network congestion is detected, a dynamic cache resource allocation mechanism is activated.

[0101] Furthermore, the personalized federated learning system for heterogeneous multi-source industrial Internet is also used to perform the following method:

[0102] During the local training phase on the client, a lightweight feature distiller is deployed to extract cross-device shared knowledge base vectors. By contrasting the learning loss function, the semantic alignment between the personalized model and the global model is constrained, and adversarial samples are generated using the cross-device shared knowledge base vectors.

[0103] Furthermore, the personalized federated learning system for heterogeneous multi-source industrial Internet is also used to perform the following method:

[0104] Obtain the entropy contribution of each client parameter update to the global model; in low-bandwidth scenarios, dynamically adjust the federation aggregation frequency based on the entropy contribution, enable real-time streaming aggregation for high-entropy clients, activate the difference compression parameter, and reconstruct the gradient transmission tensor dimension.

[0105] Furthermore, the personalized federated learning system for heterogeneous multi-source industrial Internet is also used to perform the following method:

[0106] A hash fingerprint chain that records the parameter update path triggers a historical version rollback when malicious parameter injection is detected; at the same time, a security verification protocol based on the contributed entropy value is triggered to force the encryption key to be updated and isolate suspicious client nodes.

[0107] Furthermore, the personalized federated learning system for heterogeneous multi-source industrial Internet is also used to perform the following method:

[0108] Traverse each client and build a device resource profile, which is used to map storage capacity indicators and energy consumption constraint indicators; predict federation participation based on the device resource profile and assign training task priorities; dynamically bind heterogeneous task flows in the priority queue based on the training task priorities, and activate a progressive model pruning strategy.

[0109] Furthermore, the protocol encryption module M200 is also used to perform the following method:

[0110] Deploy a federated learning performance monitoring plug-in, determine the hyperparameter combination through multi-objective optimization, and generate an adaptive parameter adjustment strategy matrix, wherein the federated learning performance monitoring plug-in is used to monitor the model convergence speed and privacy leakage risk; based on the adaptive parameter adjustment strategy matrix, dynamically adjust the learning rate and batch size of the client's local training, and inject the intensity parameter of the differential privacy noise; at the same time, according to the real-time feedback of the data modality similarity and network topology delay, allocate the privacy protection budget of each fog node proportionally; if the consumption rate of the privacy protection budget does not meet the expected threshold, trigger the compensation mechanism of elastically scaling the training batch size, and reconstruct the noise injection distribution of the differential privacy noise based on the gradient sparsity.

[0111] In summary, any step can be stored as a computer instruction or program in an unlimited computer memory and can be called and recognized by an unlimited computer processor, without any unnecessary restrictions.

[0112] Furthermore, the above technical solution only reflects the preferred technical solution of the technical solution of the embodiment of the present application. Some changes that may be made to certain parts thereof by technical personnel in this technical field all reflect the novel principles of the embodiment of the present application. Obviously, technical personnel in this field can make various changes and modifications to the present application without departing from the scope of the present application.

Claims

1. A personalized federated learning method for heterogeneous multi-source industrial Internet, characterized by: The method comprises: Connect to the client, evaluate the load, network topology latency, and data modality similarity, and generate a dynamic association mapping table between the client and the fog node; During the federated learning initialization phase, a multi-input function encryption protocol is deployed in combination with the dynamic association mapping table to perform hierarchical verifiable encryption on the client's local model gradient update amount; At the same time, a priority queue is built based on the local model update frequency and data density. The priority queue is used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters. In the heterogeneous multi-source industrial Internet, a fog node edge caching mechanism is set according to the priority queue, and the fog node edge caching mechanism is used to coordinate the client to perform distributed iterative optimization.

2. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 1, characterized in that: Evaluating load, network topology delay, and data modality similarity to generate a dynamic association mapping table between clients and fog nodes, the method further comprising: In the data-aware group, a reinforcement learning agent is embedded, wherein the reinforcement learning agent is used to optimize the association weight between the client and the fog node in real time; At the same time, the sharding storage strategy of the encryption key is dynamically adjusted according to the channel quality index. When it is detected that the fog node load exceeds the load threshold, the load migration decision tree based on topology reconstruction is triggered.

3. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 2, characterized in that: The invention also includes: introducing an attribute-based access control strategy, wherein the attribute-based access control strategy is used to define fine-grained model parameter access rights; Based on the attribute-based access control strategy, after verifying the legitimacy of the client identity, the dynamic threshold decryption unit is activated to perform homomorphic aggregation processing on the encrypted gradient.

4. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 1, characterized in that: Also includes: Build a collaborative cache network across fog nodes and perform distributed redundant storage based on model parameter importance scores; According to the collaborative cache network, a cache replacement chain is set up, and the future request distribution is predicted by combining parameter freshness and access frequency. When regional network congestion is detected, a cache resource dynamic allocation mechanism is started.

5. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 4, characterized in that: Also includes: During the local client training phase, a lightweight feature distiller is deployed to extract cross-device shared knowledge base vectors. By contrasting the learning loss function, the semantic alignment between the personalized model and the global model is constrained, and adversarial samples are generated using the cross-device shared knowledge base vector.

6. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 5, characterized in that: Also includes: Obtain the entropy value of each client parameter update's contribution to the global model; In low-bandwidth scenarios, the federation aggregation frequency is dynamically adjusted based on the contribution entropy value, real-time streaming aggregation is enabled for high-entropy clients, and the difference compression parameters are activated to reconstruct the gradient transmission tensor dimension.

7. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 6, characterized in that: Also includes: Record the hash fingerprint chain of the parameter update path, and trigger the historical version rollback when malicious parameter injection is detected; At the same time, a security verification protocol based on the contributed entropy value is triggered to force the encryption key to be updated and isolate suspicious client nodes.

8. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 7, characterized in that: Also includes: Traverse each client and build a device resource profile, which is used to map storage capacity indicators and energy consumption constraint indicators; Predicting federation participation based on the device resource profile and assigning training task priorities; According to the training task priority, the heterogeneous task flows in the priority queue are dynamically bound, and a progressive model pruning strategy is activated.

9. The personalized federated learning method for heterogeneous multi-source industrial Internet according to claim 1, characterized in that: In the federated learning initialization phase, a multi-input function encryption protocol is deployed in combination with the dynamic association mapping table. The method further includes: Deploy a federated learning performance monitoring plug-in, determine the hyperparameter combination through multi-objective optimization, and generate an adaptive parameter adjustment strategy matrix. The federated learning performance monitoring plug-in is used to monitor the model convergence speed and privacy leakage risk. Based on the adaptive parameter adjustment strategy matrix, the learning rate and batch size of the client's local training are dynamically adjusted, and the intensity parameter of the differential privacy noise is injected; At the same time, based on the real-time feedback of data modality similarity and network topology latency, the privacy protection budget of each fog node is allocated proportionally; If the consumption rate of the privacy protection budget does not meet the expected threshold, a compensation mechanism for elastically scaling the training batch size is triggered, and the noise injection distribution of the differential privacy noise is reconstructed based on the gradient sparsity.

10. A personalized federated learning system for heterogeneous multi-source industrial Internet, characterized by: A system for implementing the personalized federated learning method for heterogeneous multi-source industrial Internet according to any one of claims 1 to 9, comprising: Dynamic association evaluation module, used to connect to the client, evaluate the load, network topology delay and data modality similarity, and generate a dynamic association mapping table between the client and the fog node; A protocol encryption module is used to deploy a multi-input function encryption protocol in combination with the dynamic association mapping table during the federated learning initialization phase to perform hierarchical verifiable encryption on the client's local model gradient update amount; A queue construction module is used to, simultaneously, construct a priority queue based on the local model update frequency and data density, wherein the priority queue is used to dynamically schedule a hybrid transmission mode of local aggregation of high-frequency parameters and cloud synchronization of low-frequency parameters; The iterative optimization module is used to set a fog node edge caching mechanism in a heterogeneous multi-source industrial Internet according to the priority queue, and the fog node edge caching mechanism is used to coordinate the client to perform distributed iterative optimization.

Images