Federal learning-based industrial internet 5g-a communication evaluation model training method and device

Abstract

The application provides a federal learning-based industrial internet 5G-A communication evaluation model training method and device. The method comprises the following steps: a server determines a client identifier set participating in training; the server distributes a first initial artificial intelligence model to the clients; the clients are used for training the first initial artificial intelligence model according to local simulation communication data, and upload the obtained model parameters to the server; the server aggregates the received model parameters, and updates the first initial artificial intelligence model; the server iterates until the parameters converge, obtains a second initial artificial intelligence model, and distributes the second initial artificial intelligence model to all the clients; and the clients are used for training the second initial artificial intelligence model according to local real communication data, and obtain corresponding exclusive artificial intelligence models. The server aggregates the received model parameters, and uploads the parameters instead of the original data, so that multi-client knowledge sharing is realized under the premise of protecting data privacy, and the model training efficiency and generalization ability are improved.

Description

Technical Field

[0001] This invention relates to the field of federated learning, and more specifically to a training method and apparatus for an industrial internet 5G-A communication evaluation model based on federated learning. Background Technology

[0002] With the deep integration of the Industrial Internet and next-generation communication technologies such as 5G-A, the performance requirements of communication systems in industrial scenarios are becoming increasingly diverse and sophisticated. Traditional Industrial Internet applications are complex and varied, ranging from time-sensitive industries like machinery manufacturing to energy-sensitive industries like lightweight inspection. Different industrial scenarios have significantly different specific requirements for the same parameters in communication systems. Due to the different target audiences, the specific requirements for the same parameters in communication systems vary significantly, making it difficult to directly apply general communication performance evaluation methods to meet the unique needs of various industrial scenarios.

[0003] Existing evaluation systems generally suffer from problems such as limited indicator dimensions, lack of quantitative analysis standards, and poor scenario adaptability, failing to accurately reflect the actual performance of new communication technologies like 5G-A in complex industrial environments. Furthermore, federated learning faces a conflict between communication costs and performance in industrial scenarios: while increasing the communication frequency between edge nodes and the central server can enhance the synchronization effect of the global model and improve model accuracy, this also means more communication rounds and local training rounds, resulting in greater computational resource consumption and communication overhead. In actual industrial deployments, increased local training rounds lead to higher equipment energy consumption and training time costs, while frequent model uploads and downloads significantly increase transmission energy consumption and network latency.

[0004] Therefore, there is an urgent need to construct a dynamic evaluation method with scenario adaptability to provide theoretical basis and technical support for the intelligent evaluation and optimization of industrial communication systems, thereby promoting the application of precise allocation and standardized management of communication resources in the industrial internet. Summary of the Invention

[0005] In view of the aforementioned problems, this application is proposed to provide a method and apparatus for training an industrial internet 5G-A communication evaluation model based on federated learning that overcomes or at least partially solves the aforementioned problems, comprising:

[0006] A training method for an industrial internet 5G-A communication evaluation model based on federated learning, the method involving a server and at least one client; wherein the server includes a first initial artificial intelligence model;

[0007] The method includes:

[0008] S110, The server determines the set of client identifiers participating in the training;

[0009] S120. The server distributes the first initial artificial intelligence model to the corresponding client based on the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters, and upload the model parameters to the server.

[0010] S130. When the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters.

[0011] S140. The server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is further configured to, when receiving the second initial artificial intelligence model, obtain local real communication data, and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model.

[0012] Furthermore, the model parameters include communication metrics, attention weights, training update information, training duration, training data size, and training accuracy.

[0013] Furthermore, the step of the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters includes:

[0014] The server performs weighted aggregation on the received model parameters according to the federated average algorithm to obtain the aggregated global model parameters;

[0015] The server updates the first initial artificial intelligence model based on the global model parameters.

[0016] Furthermore, the step of the server performing weighted aggregation of the received model parameters according to the federated average algorithm to obtain the aggregated global model parameters includes:

[0017] The formula for weighted aggregation of the received model parameters based on the federated average algorithm is as follows:

[0018]

[0019] in, This represents the aggregated global model parameters. This represents the total number of clients participating in this round of training. Indicates the first The amount of local training data per client This represents the total amount of data from all clients participating in the training. Indicates the first Local model parameters uploaded by each client.

[0020] Furthermore, the step of the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters includes:

[0021] Training was conducted at the first learning rate during the first objective phase of federal training.

[0022] Training is performed at a second learning rate during the second objective phase of federated training; wherein the second learning rate is lower than the first learning rate, and the first learning rate or the second learning rate is determined by scaling the base learning rate using a preset adjustment factor.

[0023] A training method for an industrial internet 5G-A communication evaluation model based on federated learning, the method involving a server and at least one client; the server includes a first initial artificial intelligence model, used to determine a set of client identifiers participating in the training, and to distribute the first initial artificial intelligence model to the corresponding client according to the client identifier set;

[0024] The method includes:

[0025] When the client receives the first initial artificial intelligence model, it acquires local simulation communication data and trains the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters.

[0026] The client uploads the model parameters to the server; wherein, when the received model parameters reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters, iterating in a loop until the parameters converge to obtain a second initial artificial intelligence model; the server is also used to distribute the second initial artificial intelligence model to all the clients;

[0027] When the client receives the second initial artificial intelligence model, it acquires local real communication data and trains the second initial artificial intelligence model based on the local real communication data to obtain the corresponding exclusive artificial intelligence model.

[0028] Furthermore, the model parameters include communication metrics, attention weights, training update information, training duration, training data size, and training accuracy; the step of training the first initial artificial intelligence model based on the local simulated communication data to obtain the model parameters includes:

[0029] The client extracts local correlation features between communication indicators of the local simulation communication data through the first initial artificial intelligence model to obtain the corresponding feature map.

[0030] The client determines the time-series feature vector of the communication index changing over time based on the feature map;

[0031] The client determines the attention weight of the communication metric based on the temporal feature vector.

[0032] Furthermore, the step of training the second initial artificial intelligence model based on the local real communication data to obtain the corresponding dedicated artificial intelligence model includes:

[0033] The client inputs the local real communication data into the dedicated artificial intelligence model to obtain the dedicated attention weights of the communication metrics corresponding to the client; wherein, the local real communication data includes the real weights of the metrics;

[0034] The client determines the KL divergence between the communication metric's specific attention weight and the metric's true weight;

[0035] If the KL divergence is less than a preset threshold, the client outputs the dedicated artificial intelligence model and the corresponding dedicated attention weights for the communication metrics.

[0036] A training device for an industrial internet 5G-A communication evaluation model based on federated learning, the device comprising a server and at least one client; wherein the server includes a first initial artificial intelligence model;

[0037] The device includes:

[0038] The client selection module is used by the server to determine the set of client identifiers participating in the training.

[0039] The model distribution module is used by the server to distribute the first initial artificial intelligence model to the corresponding client based on the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters, and upload the model parameters to the server;

[0040] The parameter aggregation module is used to aggregate the received model parameters when the model parameters received by the server reach a preset value, and update the first initial artificial intelligence model based on the aggregated model parameters.

[0041] The iterative training module is used by the server to repeatedly trigger the client selection module, the model distribution module, and the parameter aggregation module to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when it receives the second initial artificial intelligence model, and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding dedicated artificial intelligence model.

[0042] A computer electronic device includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When executed by the processor, the computer program implements the steps of a federated learning-based training method for an industrial internet 5G-A communication evaluation model as described above.

[0043] This application has the following advantages:

[0044] In the embodiments of this application, addressing the issues of poor scenario adaptability in existing technologies and the conflict between communication costs and performance in industrial scenarios arising from federated learning, this application provides a solution for training a dynamic evaluation model of communication performance adapted to different industrial scenarios by fusing multi-scenario simulation data and real data through federated learning. Specifically, the solution involves: S110, the server determining the client identifier set participating in the training; S120, the server distributing the first initial artificial intelligence model to the corresponding client based on the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters, and then... The model parameters are uploaded to the server; S130, when the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters; S140, the server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtains a second initial artificial intelligence model, and distributes the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when receiving the second initial artificial intelligence model, and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model. By aggregating and updating the received model parameters by the server, parameter uploading replaces the original data uploading, achieving knowledge sharing among multiple clients while protecting data privacy, and improving the efficiency and generalization ability of model training. Attached Figure Description

[0045] To more clearly illustrate the technical solution of this application, the drawings used in the description of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the steps of a method for training an industrial internet 5G-A communication evaluation model based on federated learning, as provided in one embodiment of this application.

[0047] Figure 2 This is a flowchart illustrating the steps of a training method for an industrial internet 5G-A communication evaluation model based on federated learning, provided in another embodiment of this application.

[0048] Figure 3 This is a schematic diagram of a time-delay sensitive parameter provided in an embodiment of this application;

[0049] Figure 4 This is a time-delay-sensitive weight distribution diagram provided in an embodiment of this application;

[0050] Figure 5 This is a graph showing the trend of time-latency-sensitive training accuracy provided in one embodiment of this application;

[0051] Figure 6 This is a schematic diagram of energy-sensitive parameters provided in an embodiment of this application;

[0052] Figure 7 This is an energy-sensitive weight distribution diagram provided in an embodiment of this application;

[0053] Figure 8 This is a trend chart of the accuracy change of energy-sensitive training provided in one embodiment of this application;

[0054] Figure 9 This is a structural block diagram of a training device for an industrial internet 5G-A communication evaluation model based on federated learning, provided in one embodiment of this application.

[0055] Figure 10 This is a schematic diagram of the structure of a computer electronic device provided in an embodiment of the present invention;

[0056] 1. Computer electronic device; 2. External device; 3. Processing unit; 4. Bus; 5. Network adapter; 6. I / O interface; 7. Display; 8. Memory; 9. Random access memory; 10. Cache memory; 11. Storage system; 12. Program / utility; 13. Program module. Detailed Implementation

[0057] To make the objectives, features, and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0058] Analysis of existing technologies revealed that:

[0059] Existing Technology 1: Evaluation System for Performance Indicators of 5G-A Technology and Communication Systems

[0060] In industrial applications, the required functions are highly compatible with the key performance indicators of 5G-A technology. Currently, primary performance indicators are mainly implemented at two levels: the service layer and the network layer. The service layer serves terminal devices, sensors, and actuators in the industrial field, providing basic functions such as high-precision positioning, clock synchronization, and stable communication to support the core requirements of timeliness and accuracy in industrial automation scenarios. In particular, positioning accuracy and synchronization accuracy have a crucial impact on production cycle control and equipment collaborative operation.

[0061] Network layer metrics focus on the overall connectivity and scheduling capabilities of the system, encompassing dimensions such as transmission latency, bandwidth capacity, and data throughput. The aim is to ensure efficient data flow and orderly collaboration between devices within industrial systems. Latency control, service rates, and system reliability at this layer are fundamental to the stable operation of industrial networks.

[0062] To achieve the aforementioned primary indicators, the current evaluation system is further refined into a series of secondary performance indicators, such as bandwidth, latency, synchronization accuracy, and positioning accuracy. These secondary indicators can be quantitatively evaluated for different scenario requirements. For example, positioning and low latency are particularly important in high-speed mobile environments, while the stability and reliability of the communication link need to be emphasized in latency-sensitive scenarios.

[0063] In summary, 5G-A technology, with its flexible hierarchical indicator system, can cover complex and diverse industrial scenarios. Through a refined performance evaluation mechanism, it provides precise, efficient, and stable communication support for industrial systems, thereby promoting the development of intelligent manufacturing and flexible production.

[0064] However, the existing technology has the following problems:

[0065] Single-dimensional indicators: Taking the current IEEE 802.1Qcc standard and the "Industrial Internet White Paper" released by the Industrial Internet Industry Alliance in 2017 as examples, they mainly focus on network transmission layer indicators such as transmission speed, packet loss rate, and bandwidth. These indicators highly overlap with traditional communication industry system evaluation indicators and do not have sufficient characteristics of the industrial internet, namely multi-dimensional universal connectivity and high time sensitivity. Furthermore, they are not considered in the context of the new technology 5G-A to achieve the computing-communication coupling effect.

[0066] Insufficient scenario adaptation: The complexity and variability of industrial scenarios stem from the different functional priorities they ultimately achieve. These range from time-sensitive industries like machinery manufacturing to energy-sensitive industries like lightweight testing. Currently, simply applying industrial internet standards directly to the unified performance indicator system for terminal devices in the communications field fails to adequately meet the needs of these diverse scenarios. Furthermore, the lack of a targeted indicator evaluation system for industrial enterprise applications also influences the direction of system optimization.

[0067] Existing Technology 2: Evaluation System for Performance Indicators of Federated Learning and Communication Systems

[0068] In industrial environments, data typically exhibits non-independent and identically distributed characteristics, with data sources in different industrial scenarios being independent of each other. Federated learning is suitable for such heterogeneous distributed environments. Its basic mechanism is to store data or model parameters on various distributed nodes, and use a central server to coordinate data flow and computing resources for joint modeling and training, thereby improving the model's generalization ability while ensuring data privacy and security.

[0069] The iterative parameter transmission mechanism of federated learning has good coupling potential with the dynamic communication characteristics of 5G-A networks. It can help build a communication performance evaluation system with real-time response capabilities, thereby realizing dynamic evaluation and intelligent optimization of communication system performance.

[0070] However, federated learning faces a trade-off between communication costs and performance in industrial scenarios: while increasing the communication frequency between edge nodes and the central server can enhance the synchronization effect of the global model and improve model accuracy, this also means more communication rounds and local training rounds, resulting in greater computational resource consumption and communication overhead. In actual industrial deployments, increased local training rounds lead to higher equipment energy consumption and training time costs, while frequent model uploads and downloads significantly increase transmission energy consumption and network latency. Therefore, a balance must be struck between model performance and resource consumption, comprehensively considering factors such as equipment computing power, training time, bandwidth, and energy consumption, and establishing a dynamic weight allocation strategy to ensure that the federated learning system possesses both good generalization capabilities and meets the dual requirements of cost and efficiency in industrial environments.

[0071] To address the aforementioned issues, this paper proposes integrating existing performance evaluation systems for various communication systems, using mathematical modeling of key parameters, and employing federated learning as the primary tool. This approach applies standardized performance indicator construction methods to different real-world scenarios within the Industrial Internet, forming a three-dimensional joint indicator system encompassing "network-resource-model." This aims to further standardize the construction of communication system performance indicators within the Industrial Internet industry.

[0072] To better illustrate the principles of this invention, the following explanation will cover four aspects: the communication performance index system framework, the platform modular design, the training process, and the application scenarios.

[0073] I. Framework of Communication Performance Index System

[0074] 1.1 Constructing a hierarchical structure for the indicator system

[0075] The hierarchical structure of the indicator system of this invention includes a basic layer, a resource layer, and a model layer.

[0076] The foundational layer is primarily responsible for building communication performance. First, a dynamic latency grading mechanism is introduced, requiring real-time constraints to be met during critical phases (such as model delivery), while allowing for flexible latency during non-critical phases. Second, for throughput control, bandwidth allocation is dynamically adjusted based on the number of clients to avoid traffic congestion affecting simulation results and ensure communication quality matches the 5G-A standard. Finally, a fault-tolerant threshold is set for packet loss rate; minor packet loss is considered normal, while severe packet loss triggers retransmission or model interpolation compensation mechanisms. Client liveness detection and breakpoint recovery functions are also introduced to ensure training stability.

[0077] The resource layer is primarily responsible for the coordinated scheduling of computation and energy consumption. First, a device capability profiling system is established to monitor computing power and energy consumption indicators in real time. Tasks are allocated based on resource availability, with strong nodes performing complex feature extraction and weaker nodes participating in lightweight computation. Energy consumption optimization algorithms incorporate energy consumption into model adjustment conditions, automatically reducing computational load and shortening training steps to balance performance and power consumption. Simultaneously, an edge caching mechanism is deployed to preload model parameters, improving training efficiency and reducing network transmission burden.

[0078] The model layer is primarily responsible for the federated performance enhancement mechanism. First, a CNN-LSTM hybrid architecture is employed to fuse spatiotemporal features, and an attention mechanism is introduced to improve anomaly detection capabilities and model interpretability. Through structural optimization, a closed-loop processing from perception to optimization is achieved, enhancing the model's robustness and adaptability in resource-heterogeneous environments.

[0079] 1.2 Design of Network Transmission Indicators and Resource Efficiency Indicators

[0080] In one embodiment of the present invention, commonly used core indicators in industrial communication are first extracted to construct an indicator system mainly based on bandwidth, latency, packet loss rate, jitter, throughput, bit error rate, and signal strength. These indicators comprehensively cover various communication environments from wired to wireless and from single nodes to system clusters, possessing good representativeness and adaptability, and can effectively characterize the communication performance characteristics under different industrial scenarios.

[0081] Secondly, regarding energy consumption costs, the energy consumption during local training and communication is the primary reference. Priority is given to solutions that reduce overall energy consumption while ensuring accuracy, with particular emphasis on training efficiency in 5G-A transmission scenarios. In terms of computational efficiency, the FLOPs / Byte metric is introduced to evaluate computational output per unit of communication load. Combined with a real-time parameter tuning mechanism, resource allocation is optimized to avoid wasting computing power and improve the overall system operating efficiency.

[0082] 1.3 Model Performance Index Design

[0083] The model should possess strong data fitting and inference capabilities, a simple structure, and be easy to iteratively optimize. It should support intelligent functions such as bias weighting, convergence monitoring, and autonomous parameter tuning to achieve continuous optimization within the federated learning framework, ensuring reasonable weight distribution and clear heatmap representation, and effectively improving the accuracy and reliability of communication performance evaluation.

[0084] II. Platform Modular Design

[0085] In any embodiment of the present invention, the platform modular design is divided into two parts: a server and a client. Specifically, it involves a server and at least one client; wherein the server includes a first initial artificial intelligence model.

[0086] The server, acting as the central node in the federated learning framework, is responsible for the initialization, distribution, aggregation, and updating of the global model, as well as communication and interaction with various clients. As an example, the server can be a physical server deployed in an industrial internet center, a virtual cluster of multiple servers, or a computing node deployed at the edge of a 5G network. The physical form of the server can be flexibly chosen according to the actual needs of the industrial scenario, as long as centralized management and scheduling of the model can be achieved.

[0087] The first initial AI model refers to the global model initialized on the server side, which is distributed to the clients participating in the training at the start of federated learning training. This model can adopt a hybrid architecture of CNN-LSTM-attention mechanism to establish the correspondence between communication performance metric values ​​and attention weights.

[0088] The client is a local device or node deployed in various industrial scenarios of the Industrial Internet. It is used to receive models from the server, train models using local data, and upload the trained model parameters to the server. The federated learning framework of this invention can support multiple clients to participate in training simultaneously, and the number of clients can be flexibly configured according to the actual needs of the industrial scenario. As an example, the client can be a communication terminal on an AGV, a data acquisition node in a sensor network, a control unit of an industrial robot, a monitoring device on a production line, etc. The type and number of clients can be flexibly configured according to the actual scale of the industrial scenario.

[0089] This method achieves distributed model training and knowledge sharing through collaborative interaction between the server and the client, and trains a communication performance evaluation model that can be adapted to different industrial scenarios while protecting local data privacy.

[0090] III. Training Process

[0091] In this embodiment, the model training process is described from the perspective of the server.

[0092] Reference Figure 1 This application illustrates a training method for an industrial internet 5G-A communication evaluation model based on federated learning, according to an embodiment of this application.

[0093] S110. The server determines the set of client identifiers participating in the training;

[0094] S120. The server distributes the first initial artificial intelligence model to the corresponding client based on the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters, and upload the model parameters to the server.

[0095] S130. When the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters.

[0096] S140. The server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is further configured to, when receiving the second initial artificial intelligence model, obtain local real communication data, and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model.

[0097] In the embodiments of this application, addressing the issues of poor scenario adaptability in existing technologies and the conflict between communication costs and performance in industrial scenarios arising from federated learning, this application provides a solution for training a dynamic evaluation model of communication performance adapted to different industrial scenarios by fusing simulation data from multiple scenarios with real data through federated learning. By aggregating and updating the received model parameters on the server side, parameter uploads replace raw data uploads, enabling knowledge sharing among multiple clients while protecting data privacy, thereby improving the efficiency and generalization ability of model training.

[0098] The following will further explain a training method for an industrial internet 5G-A communication evaluation model based on federated learning in this exemplary embodiment.

[0099] In one embodiment of the present invention, the specific process of "the server determining the set of client identifiers participating in training" in step S110 can be further explained in conjunction with the following description.

[0100] It's important to note that the server is the core of the system, responsible for the global management and scheduling of the model. It probes clients via UDP broadcast, assigns unique IDs, and tracks client status in real time. During each training round, the server randomly selects clients to participate. Specifically, in the system initialization and preparation phase, the server first loads the preset initial AI model structure and related configuration parameters (e.g., batch size = 600, Dropout ratio = 0.5, total training rounds = 500, and 10 local training iterations per round of communication). It then sequentially opens the UDP port for control command transmission and the TCP port for model data exchange, while activating the visualization module to support dynamic display of the subsequent training process. After the client starts, it sends a registration request to the server via UDP communication. The server responds, assigns a unique identifier to each client, and records its network address and port information. During the federated training cycle, the server selects a subset of nodes from the registered clients to participate in each training round.

[0101] As an example, the server may randomly select a preset number of clients from all registered online clients to form a client identifier set for training; or, the server may select eligible clients to form a client identifier set for training based on factors such as the client's online status, historical participation frequency, and device computing power; or, the server may select clients for training based on the size of the local data on the client.

[0102] In one specific implementation, the server maintains a client status table, recording information such as the online status, the time of the last participation, and the amount of local data for each client. At the beginning of each training round, the server selects three online clients that have not recently participated from the status table, and randomly selects one of them as the client identifier set for this round of training.

[0103] In one embodiment of the present invention, the specific process of step S120, "the server distributes the first initial artificial intelligence model to the corresponding client according to the client identifier set; wherein the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model according to the local simulation communication data, obtain model parameters, and upload the model parameters to the server," can be further described in conjunction with the following description.

[0104] Understandably, after the server distributes the model, it doesn't directly control the client's training process. Instead, it triggers the client to perform local training autonomously by distributing the model and hyperparameters. The client only receives the model; the local data remains within the server's system. After distributing the model, the server needs to wait for the client to complete training and upload the model parameters. The server can receive the client's uploads by listening on a TCP port.

[0105] During the model and data distribution process, the server distributes the first initial artificial intelligence model to the corresponding client based on the client identifier set determined in step S110. Specifically, the server uses the TCP protocol to encrypt and compress the current global model parameters and distribute the model architecture and hyperparameters (such as learning rate, batch size, Dropout ratio, or training epochs) to the selected client. At the same time, it allocates data indexes to ensure that the client only processes local data. Based on this, the client generates local simulation communication data that conforms to the preset parameter range and completes the division of the training set and test set. After all clients have completed data preparation and confirmed their readiness status, federated training officially starts and uploads the trained model parameters back to the server.

[0106] It should be noted that the local simulated communication data refers to data generated by the client within the corresponding parameter range according to a preset industrial application scenario type, used to simulate the communication performance indicators in an actual industrial environment. The local simulated communication data includes values ​​for indicators such as bandwidth, latency, packet loss rate, jitter, throughput, bit error rate, and signal strength.

[0107] The model parameters include communication metrics, attention weights, training update information, training duration, training data size, and training accuracy.

[0108] The communication metric attention weights refer to the degree of contribution of each communication metric to the current communication performance calculated by the client model through the attention mechanism. They are usually represented by a set of normalized values, with the sum of each weight being 1.

[0109] The training update information refers to the gradient change of model parameters during local training on the client side, which indicates the direction and magnitude of the adjustment of model parameters.

[0110] The training duration refers to the time spent by the client to complete this round of local training.

[0111] The training data scale refers to the amount of local data used by the client in this training, such as the number of sample records.

[0112] The training accuracy refers to the prediction accuracy of the client model on the local validation set.

[0113] As an example, the server maintains a "client identifier-TCP connection" mapping table. When a model needs to be distributed, the server first compresses the initial AI model, then looks up the corresponding TCP connection for each identifier in the client identifier set, and sends the compressed model data to the corresponding client through that connection. After sending, the server waits for a confirmation message from the client; if no confirmation is received, it retransmits the data. After receiving the global model, the client trains the model based on locally generated data. After training, the client packages and compresses metadata containing gradient update information, training duration, data size, local accuracy, and the weight distribution of each parameter, and uploads it to the server via the TCP protocol.

[0114] In one embodiment of the present invention, the specific process of step S130, "when the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters," can be further explained in conjunction with the following description.

[0115] It's important to note that during each training round, the server randomly selects clients to participate, uses a weighted average algorithm to aggregate the locally updated parameters, generates and distributes a new global model. The server only aggregates the model parameters from each client, without accessing the original data, thus meeting data privacy requirements. The server integrates the knowledge learned locally by multiple clients into the global model, achieving cross-scenario knowledge sharing. Through weighted aggregation, clients with larger datasets contribute more knowledge, adapting to heterogeneous data distributions in industrial scenarios.

[0116] In one embodiment of the present invention, the specific process of "the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters" can be further described in conjunction with the following description.

[0117] The server performs weighted aggregation on the received model parameters according to the federated average algorithm to obtain the aggregated global model parameters;

[0118] The server updates the first initial artificial intelligence model based on the global model parameters.

[0119] As an example, when the number of model parameters received by the server reaches a preset number, or the waiting time exceeds a preset duration, the server triggers an aggregation operation. The server merges the local model parameters uploaded by multiple clients into new global model parameters. A federated averaging algorithm can be used for weighted aggregation, with weight coefficients determined based on the data scale of each client. The server assigns the aggregated global model parameters to the first initial AI model, completing the model update for this round. Specifically, the aggregated global model parameters can directly replace the original parameters of the first initial AI model. The updated model will serve as the base model for the next round of training.

[0120] The formula for weighted aggregation of the received model parameters based on the federated average algorithm is as follows:

[0121]

[0122] in, This represents the aggregated global model parameters. This represents the total number of clients participating in this round of training. Indicates the first The amount of local training data per client This represents the total amount of data from all clients participating in the training. Indicates the first Local model parameters uploaded by each client.

[0123] In one embodiment of the present invention, the specific process of "the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters" can be further described in conjunction with the following description.

[0124] Training was conducted at the first learning rate during the first objective phase of federal training.

[0125] Training is performed at a second learning rate during the second objective phase of federated training; wherein the second learning rate is lower than the first learning rate, and the first learning rate or the second learning rate is determined by scaling the base learning rate using a preset adjustment factor.

[0126] It's important to note that the learning rate is a hyperparameter in machine learning training, controlling the step size of model parameter updates. An excessively large learning rate may cause the model to oscillate and fail to converge, while an excessively small learning rate may lead to slow training speed and getting stuck in local optima. The server dynamically adjusts the learning rate based on the training progress, accelerating convergence in the early stages and preventing overfitting in the later stages. To ensure training continuity, the server periodically checks the client status and automatically retrys failed transmissions.

[0127] As an example, a dynamic learning rate adaptive mechanism is introduced to control the training process. A relatively high learning rate (adjustment factor M=1.5) is set in the early stages of federated training to accelerate model convergence and facilitate rapid exploration of the parameter space. In the later stages of training, the system automatically reduces the learning rate to achieve fine-tuning of parameters, avoid overfitting or performance oscillations, and improve the overall model convergence quality. The base learning rate can be 0.01.

[0128] In one embodiment of the present invention, the specific process of step S140, which states that "the server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtains a second initial artificial intelligence model, and distributes the second initial artificial intelligence model to all clients; wherein, the client is further configured to, when receiving the second initial artificial intelligence model, obtain local real communication data and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model", can be further described in conjunction with the following description.

[0129] It should be noted that this invention relates to two training phases. The first phase utilizes simulation data to learn common patterns across multiple scenarios, obtaining a general model with good generalization ability. The second phase uses real data for fine-tuning, obtaining a specific model accurately adapted to the specific scenario. In the second phase, the server uses a federated averaging algorithm to weight and aggregate the received model parameters, generating a global model for the next round. This process iterates until the training termination condition is met, resulting in a second initial AI model. The training termination condition is that the loss function value tends to stabilize or reaches a preset threshold. The second initial AI model refers to the global model that has converged after multiple rounds of federated learning training, having learned common communication patterns across multiple scenarios. After obtaining the converged general model, the server needs to distribute it to all clients, including those that may not have participated in the training of the general model. Upon receiving the general model, the client uses local real communication data to fine-tune and train the general model, thus obtaining a specific model adapted to the specific scenario.

[0130] As an example, during the global evaluation and report generation phase, the server broadcasts the final trained global model parameters to all clients. Each client validates the model based on its local test set and calculates the KL divergence to measure the difference between the model's predictions and the true distribution, while also generating a local parameter importance ranking. After the evaluation results are uploaded, the server summarizes and analyzes the weight matrices uploaded by each client, generating a CSV report containing statistical indicators such as average weights and standard deviations. It also uses a visualization module to output a multi-client weight distribution heatmap and a weight evolution trend chart during the training process.

[0131] During system maintenance and optimization, the server broadcasts a heartbeat packet via UDP every 5 minutes to monitor the online status of clients. If no response is received within a set 1-second response time, the client is marked as offline and removed from the training task. This invention features breakpoint resumption capability, automatically resuming the training process from the last successful round after an unexpected interruption on the server or client side. Furthermore, the server dynamically adjusts the number of participating clients and TCP transmission parameters based on real-time network load conditions to optimize communication overhead and resource allocation efficiency while ensuring system stability.

[0132] In one embodiment of the present invention, the simulation platform includes the following in terms of architecture design and functional mechanisms: hybrid model architecture design, temperature scaling mechanism optimization, federated communication mechanism, dynamic adaptation mechanism optimization, and industrial scenario adaptation design.

[0133] First, the platform introduces a CNN-LSTM-attention mechanism fusion model, comprehensively utilizing Convolutional Neural Networks (CNN) for spatial feature extraction, Long Short-Term Memory Networks (LSTM) for temporal modeling, and an attention mechanism for feature selection in key dimensions, thereby achieving the collaborative expression of multimodal communication metrics. This ternary fusion architecture significantly improves the model's generalization ability, feature interpretability, and robustness when handling multi-source heterogeneous performance metrics, breaking through the performance bottleneck of traditional single models in complex industrial communication environments.

[0134] Secondly, the platform incorporates a temperature scaling control mechanism to adjust the probability distribution entropy value of the model output. When significant differences in client data distribution are detected (i.e., KL divergence exceeds 0.2), the system automatically reduces the temperature parameter value, thereby highlighting the dominant role of key features in the model output. This temperature parameter is uniformly learned and set by the server and distributed to the client before each training round, reducing model fluctuations caused by node heterogeneity at the source and improving the stability and consistency of the global model.

[0135] In terms of communication mechanisms, the platform adopts a separate architecture for control signaling and data transmission: the UDP protocol is used to transmit lightweight messages such as training instructions and status information, ensuring that communication latency is controlled within 5ms; the TCP protocol is used to transmit model parameters and training data, ensuring communication stability and data reliability, with a packet loss rate controlled below 0.1%. Furthermore, by combining the Huffman entropy coding algorithm to compress sparse gradients, the gradient volume can be compressed to 15% of its original size, significantly improving communication efficiency and meeting the high-concurrency simulation requirements of the 5G-A environment.

[0136] Regarding the regulation of the training process, it has already been explained in step S130, and will not be repeated here.

[0137] In terms of industrial adaptability design, the platform analyzes the statistical correlation between key communication indicators through a parameter coupling mechanism. For example, bandwidth and throughput are positively correlated (correlation coefficient is approximately 0.7), while packet loss rate and signal strength are negatively correlated, accurately simulating the physical coupling characteristics in industrial communication systems. Simultaneously, it employs a Gaussian mixture distribution generation strategy, supporting dual-track operation in both preset parameter mode and random mode. This allows for detailed modeling of specific industrial scenarios as well as high-coverage simulation of typical industrial environments (such as factory workshops and outdoor base stations).

[0138] Through the aforementioned structural and mechanistic innovations, a closed-loop federated learning system of "perception-decision-optimization" for industrial communication scenarios has been constructed. This effectively overcomes the structural constraints of traditional communication systems in performance evaluation and intelligent optimization, realizing a quantifiable, transferable, and generalizable intelligent indicator analysis and model training framework, possessing significant engineering practicality and innovative value.

[0139] In this embodiment, the model training process is described from the perspective of the server.

[0140] Reference Figure 2 This paper illustrates another embodiment of a training method for an industrial internet 5G-A communication evaluation model based on federated learning, which involves a server and at least one client. The server includes a first initial artificial intelligence model for determining a set of client identifiers participating in the training and distributing the first initial artificial intelligence model to the corresponding client based on the client identifier set.

[0141] The method includes:

[0142] S210. When the client receives the first initial artificial intelligence model, it acquires local simulation communication data and trains the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters.

[0143] S220. The client uploads the model parameters to the server; wherein, when the received model parameters reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters, iterating in a loop until the parameters converge to obtain a second initial artificial intelligence model; the server is also used to distribute the second initial artificial intelligence model to all the clients;

[0144] S230. When the client receives the second initial artificial intelligence model, it acquires local real communication data and trains the second initial artificial intelligence model based on the local real communication data to obtain the corresponding exclusive artificial intelligence model.

[0145] In the embodiments of this application, the first initial artificial intelligence model is trained by the client using local simulated communication data, which solves the problems of single indicator dimensions and lack of targeted quantitative standards, enabling the model to learn general communication rules in multiple scenarios and laying the foundation for subsequent scenario adaptation. The second initial artificial intelligence model is fine-tuned and trained by the client using local real communication data to obtain a dedicated artificial intelligence model, which solves the problem of insufficient scenario adaptation and enables the final model to accurately adapt to the personalized communication performance evaluation needs of different industrial scenarios.

[0146] The following will further explain a training method for an industrial internet 5G-A communication evaluation model based on federated learning in this exemplary embodiment.

[0147] In one embodiment of the present invention, the specific process of step S210, "when the client receives the first initial artificial intelligence model, it acquires local simulation communication data and trains the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters," can be further explained in conjunction with the following description.

[0148] The client obtains the communication indicators and the industrial scenario type corresponding to the client, and generates the local simulation communication data within the preset parameter range of the corresponding communication indicators based on the industrial scenario type.

[0149] As an example, the client simulates equipment in an industrial scenario, generating simulation data and performing local training. The client generates a training set (50,000 samples) and a test set (10,000 samples) based on preset parameter ranges (bandwidth 1-100Mbps, packet loss rate 0-0.1%), simulating various indicators in real-world communication scenarios. After loading the global model, the client performs 10 rounds of training on local data, using mini-batch gradient descent (batch_size=600) and Dropout (0.5 ratio) to prevent overfitting. For communication, the client transmits control commands via UDP and model parameters and evaluation results via TCP, while compressing the data to improve transmission efficiency. To handle connection interruptions or client downtime, the client supports resume functionality to ensure that training progress is not lost due to failures.

[0150] In one embodiment of the present invention, the specific process of "training the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters" can be further explained in conjunction with the following description.

[0151] The client extracts local correlation features between communication indicators of the local simulation communication data through the first initial artificial intelligence model to obtain the corresponding feature map;

[0152] The client determines the time-series feature vector of the communication index changing over time based on the feature map;

[0153] The client determines the attention weight of the communication metric based on the time-series feature vector.

[0154] Specifically, after receiving the global model, each client trains the model based on the locally generated simulation communication data. The training process includes normalizing the original input data and sequentially passing it through CNN, pooling layers, BiLSTM, and attention mechanism modules to extract the potential temporal correlations and feature weights between multidimensional communication parameters.

[0155] The model design focuses on feature extraction and weight regression. The model adopts a CNN-LSTM-attention hybrid architecture to extract features from different perspectives.

[0156] The first convolutional layer has 64 3x1 convolutional kernels, which capture the local relationship between adjacent parameters (such as the relationship between bandwidth and throughput). The operation steps are to first perform convolution operation, then batch normalization, ReLU activation, and finally max pooling. The second convolutional layer is equipped with 128 3x1 convolutional kernels to expand the receptive field and extract higher-order features.

[0157] The bidirectional LSTM layers, with 128 units each for the forward and backward directions, capture temporal dependencies. Attention scores are calculated using two fully connected layers, and then weighted summed after Softmax normalization. The calculation formula is as follows:

[0158]

[0159] in, It is the output of a scoring function with i indicators. These are the corresponding attention weights.

[0160] Finally, the data enters the feature fusion and decision layer, which includes a 256-node fully connected layer, and the final weight output is formed after temperature scaling.

[0161] Each client generates massive amounts of data during a complete training cycle, including performance metrics for the current training epoch and the overall model accuracy. To enable real-time monitoring and continuous recording, a visualization module (Visualizer.py) was designed. This module generates real-time monitoring charts after each local training epoch. The most important parameter importance bar chart is sorted by weight for the current epoch, showcasing the influence of each parameter and aiding in rapid decision-making. Secondly, after the overall training is complete, a weight change curve over epochs is plotted, revealing the model's convergence dynamics. For data logs, CSV files are exported, including metrics such as mean and standard deviation, supporting offline in-depth analysis. Data retention is also ensured; all charts are archived with timestamps for easy tracking of historical training results and model evolution.

[0162] In one embodiment of the present invention, the specific process of step S220, "the client uploads the model parameters to the server; wherein, when the received model parameters reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model according to the aggregated model parameters, iterating cyclically until the parameters converge to obtain a second initial artificial intelligence model; the server is also used to distribute the second initial artificial intelligence model to all the clients," can be further explained in conjunction with the following description.

[0163] It should be noted that after training is completed, the model parameters on the client are updated to include local knowledge. The client uploads the updated model parameters to the server, but does not upload the original data to protect data privacy.

[0164] As an example, after training is complete, a TCP connection is established with the server. The client packages and compresses metadata containing gradient update information, training duration, data size, local accuracy, and the weight distribution of each parameter, and uploads it to the server via the TCP protocol. After sending, the client waits for an acknowledgment message from the server. If no acknowledgment is received within a timeout period, it automatically retransmits. The server then performs weighted aggregation of the received model parameters according to the federated averaging algorithm and generates the global model for the next round. This process iterates until the training termination condition is met. The specific processing steps of the server are not detailed here.

[0165] In one embodiment of the present invention, the specific process of step S230, "when the client receives the second initial artificial intelligence model, it acquires local real communication data and trains the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model," can be further explained in conjunction with the following description.

[0166] Understandably, this step constitutes the second phase of the client-side training process. After receiving the second initial AI model (general model) from the server, the client uses local real-world communication data to fine-tune and train the model, obtaining a custom AI model adapted to this scenario.

[0167] As an example, the client can identify and receive the second initial AI model in several ways: the server includes an identifier when distributing the model, indicating whether it is the final general-purpose model; the client then parses the message header to identify it upon receiving the model; or, the server sends a training completion notification before distributing the final model. When the client receives the second initial AI model, it acquires local real-world communication data. This local real-world communication data refers to real data collected by the client in an actual industrial environment, which differs from the simulation data used in the first stage. The client can read historically collected real-world communication data from local storage and train the second initial AI model based on this data, making the model more suitable for the communication characteristics of the current scenario, ultimately obtaining a client-specific AI model.

[0168] In one embodiment of the present invention, the specific process of "training the second initial artificial intelligence model based on the local real communication data to obtain the corresponding exclusive artificial intelligence model" can be further described in conjunction with the following description.

[0169] The client inputs the local real communication data into the dedicated artificial intelligence model to obtain the dedicated attention weights of the communication metrics corresponding to the client; wherein, the local real communication data includes the real weights of the metrics;

[0170] The client determines the KL divergence between the communication metric's specific attention weight and the metric's true weight;

[0171] If the KL divergence is less than a preset threshold, the client outputs the dedicated artificial intelligence model and the corresponding dedicated attention weights for the communication metrics.

[0172] It should be noted that, to improve the accuracy and efficiency of the model, the system also introduces KL divergence and attention mechanisms to perform a global impact assessment on key communication metrics (such as bandwidth and latency). During the model validation phase, model validation is conducted based on a local test set, and KL divergence is calculated to measure the difference between the model's predictions and the true distribution, while also generating a local parameter importance ranking.

[0173] Real-world communication data contains the true weight labels of the metrics. These weights can be obtained through expert annotation, historical experience, or back-calculation based on actual results. When the client determines that the KL divergence between the communication metric's specific attention weight and the metric's true weight is less than a preset threshold, the model passes validation. The client then outputs the model's specific attention weight as the final result for subsequent communication performance evaluation and resource optimization.

[0174] IV. Application Scenarios

[0175] In industrial communication systems, transmission rate and latency directly impact user experience, while signal strength and bandwidth are more limited by terminal performance. For latency-sensitive scenarios such as telemedicine and precision machining, latency, jitter, and bit error rate must be strictly controlled; while in energy-sensitive scenarios such as sensor monitoring and image recognition, signal frequency and strength must be limited to reduce equipment energy consumption and ensure stable system operation.

[0176] To verify the system's adaptability to different communication indicator weights in industrial applications, we selected two typical scenarios for simulation testing: AGV (Automated Guided Vehicle) scenario (latency-sensitive scenario) and large-scale terminal logistics network architecture (energy-sensitive scenario). The specific parameter settings are shown in Table 1. By setting the respective parameter ranges, we evaluated the system's performance prediction capabilities under various constraints.

[0177]

[0178] Table 1. Specific parameter setting range for AGV cart and large-scale terminal logistics network architecture.

[0179] Simulation results and adaptability analysis:

[0180] Based on practical experience, considering the characteristics of high bandwidth, low latency, and low packet loss in latency-sensitive scenarios, the predicted experimental results indicate that when latency and bandwidth are controlled within a relatively high and stable range, the overall throughput and bit error rate are expected to rank relatively high in the overall weight ranking. Adjusting the randomly generated data range in the program, such as... Figure 3 As shown.

[0181] After setting the model parameters, set up 3 client nodes, randomly selecting 1 node for local training and communication with the server in each round; set the total number of training rounds to 500 rounds, with 50 local training iterations in each communication round. The heatmap of the seven weights after federated learning is shown below. Figure 4 As shown.

[0182] Observing the accuracy trend throughout the training process, as the number of training rounds increases, the model's accuracy in fitting the weight ratio gradually increases, reaching approximately 63% after 500 training rounds. Slight overfitting occurs in the last few training rounds. The overall trend is as follows: Figure 5 As shown.

[0183] The final weight distribution is shown in Table 2:

[0184]

[0185] Table 2 Weights of Delay-Sensitive Parameters

[0186] However, considering the characteristics of energy-sensitive scenarios with high tolerance for bit error rates and control over throughput limits, the prediction experiment results show that the signal strength index is estimated to be relatively high in the overall weight ranking. This is because stronger signal strength means higher communication costs and energy consumption. Adjusting the randomly generated data range in the program, such as... Figure 6 As shown.

[0187] After setting the model parameters, three client nodes were configured. One node was randomly selected each round for local training and communication with the server. The total number of training rounds was set to 500, with 10 local training iterations per communication round. Increasing the number of communication rounds forced the energy-intensive model to make decisions. The overall accuracy trend after increasing the number of communication rounds is shown in Figure 7.

[0188] Stability is improved, but overall training accuracy remains almost unchanged. Therefore, when the total number of training epochs is fixed, the main function of increasing the number of communication epochs is to unify the global model in a timely manner and reduce the generation of bias nodes. Under the same final prediction accuracy, the heatmap of the seven weights of the communication model parameters after federated learning is as follows: Figure 8 As shown.

[0189] As can be seen, in energy-sensitive industrial environments, signal strength directly affects the overall energy consumption of the model. Furthermore, the model has a high tolerance for other parameters and its throughput is deliberately limited to meet low-power consumption requirements. Therefore, the weighting differences are quite significant, as shown in Table 3.

[0190]

[0191] Table 3 Weights of Energy-Sensitive Parameters

[0192] When parameters are constrained under different conditions, they are divided into time-sensitive industrial scenarios, which meet the characteristics of high bandwidth, low latency, and low packet loss. The higher the throughput and the lower the bit error rate, the better the performance. In energy-sensitive scenarios, signal strength has a greater impact on energy consumption and is ranked higher in weight, which basically meets the experimental expectations. This proves that the present invention can provide quantitative and adaptive weight analysis in different typical industrial environments, which has certain reference value.

[0193] In summary, this invention proposes a dynamic performance evaluation framework for 5G-A communication systems. By introducing performance indicators across three dimensions—network, resources, and model—including bandwidth, latency, packet loss rate, energy consumption, computational efficiency, and model convergence, a multi-parameter coupled evaluation system is constructed. This system supports the automatic generation of personalized evaluation weights based on pre-input parameter ranges and can automatically adjust the evaluation focus according to the characteristics of different industrial scenarios, achieving dynamic and accurate evaluation of communication performance in current and future emerging industrial scenarios.

[0194] This invention designs a lightweight federated learning architecture that combines attention mechanisms, gradient compression algorithms, and hybrid communication protocols to significantly reduce resource consumption for model transmission and local computation. Furthermore, it introduces a dynamic learning rate adjustment strategy and a breakpoint resume mechanism to further improve system training efficiency and stability, making it suitable for large-scale distributed industrial node collaborative modeling and real-time communication performance monitoring.

[0195] This invention develops a federated learning simulation verification platform for industrial communication performance evaluation, featuring multimodal data generation, dynamic parameter adjustment, and visualization analysis. The platform supports communication fluctuation simulation in typical industrial scenarios and visually displays the importance ranking and variation patterns of various communication parameters through heatmaps and trend curves. This platform not only provides tools to support communication resource allocation and optimization but also enhances the transparency and decision-making support capabilities of the evaluation process.

[0196] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0197] Reference Figure 9This application illustrates a training device for an industrial internet 5G-A communication evaluation model based on federated learning, according to an embodiment of this application. The device involves a server and at least one client; wherein the server includes a first initial artificial intelligence model; specifically, it includes the following modules:

[0198] Specifically, it includes:

[0199] Select client module 910, which is used by the server to determine the set of client identifiers participating in the training;

[0200] The model distribution module 920 is used by the server to distribute the first initial artificial intelligence model to the corresponding client according to the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model according to the local simulation communication data, obtain model parameters, and upload the model parameters to the server;

[0201] The parameter aggregation module 930 is used to aggregate the received model parameters when the model parameters received by the server reach a preset value, and update the first initial artificial intelligence model based on the aggregated model parameters.

[0202] The iterative training module 940 is used by the server to repeatedly trigger the client selection module, the model distribution module, and the parameter aggregation module to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when it receives the second initial artificial intelligence model, and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model.

[0203] In one embodiment of the present invention, the parameter aggregation module 930 includes:

[0204] The weighted aggregation submodule is used by the server to perform weighted aggregation on the received model parameters according to the federated average algorithm to obtain the aggregated global model parameters;

[0205] The model update submodule is used by the server to update the first initial artificial intelligence model based on the global model parameters.

[0206] In one embodiment of the present invention, the weighted aggregation submodule includes:

[0207] The aggregation calculation unit is used to perform weighted aggregation of the received model parameters according to the federated average algorithm. The calculation formula is as follows:

[0208]

[0209] in, This represents the aggregated global model parameters. This represents the total number of clients participating in this round of training. Indicates the first The amount of local training data per client. This represents the total amount of data from all clients participating in the training. Indicates the first Local model parameters uploaded by each client.

[0210] Reference Figure 10 The illustration shows a computer electronic device for implementing a federated learning-based industrial internet 5G-A communication evaluation model training method of the present invention, which may specifically include the following:

[0211] The aforementioned computer electronic device 1 is manifested in the form of a general-purpose computing device. The components of the computer electronic device 1 may include, but are not limited to: one or more processors or processing units 3, memory 8, and a bus 4 connecting different system components (including memory 8 and processing unit 3).

[0212] Bus 4 represents one or more of several bus architectures, including memory buses or memory controllers, peripheral buses, graphics acceleration ports, processors, or local buses using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Audio / Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0213] Computer electronic device 1 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer electronic device 1, including volatile and non-volatile media, removable and non-removable media.

[0214] Memory 8 may include computer system readable media in the form of volatile memory, such as random access memory 9 and / or cache memory 10. Computer electronic device 1 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 11 may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). Although Figure 10As not shown, a disk drive for reading and writing to a removable non-volatile disk (such as a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (such as a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 4 via one or more data media interfaces. The memory may include at least one program product having a set (e.g., at least one) of program modules 13 configured to perform the functions of the embodiments of this application.

[0215] A program / utility 12 having a set (at least one) of program modules 13 may be stored, for example, in memory. Such program modules 13 include—but are not limited to—an operating system, one or more application programs, other program modules 13, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 13 typically perform the functions and / or methods described in the embodiments of this application.

[0216] The computer electronic device 1 can also communicate with one or more external devices 2 (e.g., keyboard, pointing device, display 7, camera, etc.), and with one or more devices that enable an operator to interact with the computer electronic device 1, and / or with any device that enables the computer electronic device 1 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed through the I / O interface 6. Furthermore, the computer electronic device 1 can also communicate with one or more networks (e.g., local area network (LAN)), wide area network (WAN), and / or public networks (e.g., the Internet) through the network adapter 5. Figure 10 As shown, network adapter 5 communicates with other modules of computer electronic device 1 via bus 4. It should be understood that, although... Figure 10 Not shown, it may be combined with other hardware and / or software modules, including but not limited to: microcode, device drivers, redundant processing unit 3, external disk drive array, RAID system, tape drive and data backup storage system 11, etc.

[0217] The processing unit 3 executes various functional applications and data processing by running programs stored in memory 8, such as implementing a federated learning-based industrial internet 5G-A communication evaluation model training method provided in the embodiments of this application.

[0218] That is, when the above-mentioned processing unit 3 executes the above program, it implements the following: S110, the server determines the client identifier set participating in the training; S120, the server distributes the first initial artificial intelligence model to the corresponding client according to the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model according to the local simulation communication data, obtain model parameters, and upload the model parameters to the server; S130, when the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model according to the aggregated model parameters; S140, the server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when receiving the second initial artificial intelligence model, and train the second initial artificial intelligence model according to the local real communication data to obtain a corresponding dedicated artificial intelligence model.

[0219] In this application embodiment, the application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a federated learning-based training method for an industrial internet 5G-A communication evaluation model as provided in all embodiments of the application.

[0220] That is, when the program is executed by the processor, the following steps are implemented: S110, the server determines the client identifier set participating in the training; S120, the server distributes the first initial artificial intelligence model to the corresponding client according to the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model according to the local simulation communication data, obtain model parameters, and upload the model parameters to the server; S130, when the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model according to the aggregated model parameters; S140, the server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when receiving the second initial artificial intelligence model, and train the second initial artificial intelligence model according to the local real communication data to obtain a corresponding dedicated artificial intelligence model.

[0221] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0222] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0223] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the operator's computer, partially on the operator's computer, as a standalone software package, partially on the operator's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the operator's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider). The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

[0224] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0225] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0226] The above provides a detailed description of the training method and apparatus for the industrial internet 5G-A communication evaluation model based on federated learning provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A training method for an industrial internet 5G-A communication evaluation model based on federated learning, characterized in that, The method involves a server and at least one client; wherein the server includes a first initial artificial intelligence model; The method includes: S110, The server determines the set of client identifiers participating in the training; S120. The server distributes the first initial artificial intelligence model to the corresponding client based on the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model based on the local simulation communication data, obtain model parameters, and upload the model parameters to the server; the model parameters include communication metrics attention weights, training update information, training duration, training data scale, and training accuracy; S130. When the model parameters received by the server reach a preset value, the server aggregates the received model parameters and updates the first initial artificial intelligence model based on the aggregated model parameters. S140. The server repeats steps S110-S130 to converge the parameters of the first initial artificial intelligence model, obtain the second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when it receives the second initial artificial intelligence model, and train the second initial artificial intelligence model based on the local real communication data to obtain the corresponding exclusive artificial intelligence model. In step S120, the attention weight of the communication indicator is calculated according to the following steps: the client is used to extract the local correlation features between the communication indicators of the local simulation communication data through the first initial artificial intelligence model to obtain the corresponding feature map; the client is used to determine the temporal feature vector of the communication indicator changing over time based on the feature map; the client is used to determine the attention weight of the communication indicator based on the temporal feature vector. In step S140, the step of the client training the second initial artificial intelligence model based on the local real communication data to obtain the corresponding exclusive artificial intelligence model includes: the client inputting the local real communication data into the exclusive artificial intelligence model to obtain the communication indicator exclusive attention weights corresponding to the client; wherein, the local real communication data includes indicator real weights; the client determining the KL divergence between the communication indicator exclusive attention weights and the indicator real weights; if the KL divergence is less than a preset threshold, the client outputs the exclusive artificial intelligence model and the corresponding communication indicator exclusive attention weights.

2. The method according to claim 1, characterized in that, The step of the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters includes: The server performs weighted aggregation on the received model parameters according to the federated average algorithm to obtain the aggregated global model parameters; The server updates the first initial artificial intelligence model based on the global model parameters.

3. The method according to claim 2, characterized in that, The step of the server performing weighted aggregation of the received model parameters according to the federated average algorithm to obtain the aggregated global model parameters includes: The formula for weighted aggregation of the received model parameters based on the federated average algorithm is as follows: in, This represents the aggregated global model parameters. This represents the total number of clients participating in this round of training. Indicates the first The amount of local training data per client. This represents the total amount of data from all clients participating in the training. Indicates the first Local model parameters uploaded by each client.

4. The method according to claim 1, characterized in that, The step of the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters includes: Training was conducted at the first learning rate during the first objective phase of federal training. Training is performed at a second learning rate during the second objective phase of federated training; wherein the second learning rate is lower than the first learning rate, and the first learning rate or the second learning rate is determined by scaling the base learning rate using a preset adjustment factor.

5. A training device for an industrial internet 5G-A communication evaluation model based on federated learning, characterized in that, The device involves a server and at least one client; wherein the server includes a first initial artificial intelligence model; The device includes: The client selection module is used by the server to determine the set of client identifiers participating in the training. The model distribution module is used by the server to distribute the first initial artificial intelligence model to the corresponding client based on the client identifier set; wherein, the client is used to receive the first initial artificial intelligence model, obtain local simulation communication data, train the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters, and upload the model parameters to the server; the model parameters include communication metrics attention weights, training update information, training duration, training data scale, and training accuracy; The parameter aggregation module is used to aggregate the received model parameters when the model parameters received by the server reach a preset value, and update the first initial artificial intelligence model based on the aggregated model parameters. An iterative training module is used by the server to repeatedly trigger the client selection module, the model distribution module, and the parameter aggregation module to converge the parameters of the first initial artificial intelligence model, obtain a second initial artificial intelligence model, and distribute the second initial artificial intelligence model to all clients; wherein, the client is also used to obtain local real communication data when it receives the second initial artificial intelligence model, and train the second initial artificial intelligence model based on the local real communication data to obtain a corresponding exclusive artificial intelligence model; The attention weights of the communication metrics are obtained by the client performing the following functions: the client is used to extract local correlation features between communication metrics of the local simulation communication data through the first initial artificial intelligence model to obtain a corresponding feature map; the client is used to determine the temporal feature vector of the communication metrics changing over time based on the feature map; the client is used to determine the attention weights of the communication metrics based on the temporal feature vector. The dedicated AI model is obtained by the client performing the following functions: the client inputs the local real communication data into the dedicated AI model to obtain the communication indicator dedicated attention weights corresponding to the client; wherein, the local real communication data includes the indicator real weights; the client determines the KL divergence between the communication indicator dedicated attention weights and the indicator real weights; if the KL divergence is less than a preset threshold, the client outputs the dedicated AI model and the corresponding communication indicator dedicated attention weights.

6. A computer electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the steps of a training method for an industrial internet 5G-A communication evaluation model based on federated learning as described in any one of claims 1 to 4.

Images