Federal learning-based cross-regional edge node congestion collaborative early warning method

Abstract

The invention discloses a cross-regional edge node congestion collaborative early warning method based on federated learning. The method comprises the steps of data set construction, federated learning framework initialization, local model training, global model aggregation, congestion detection and congestion early warning. The invention relates to the technical field of cross-regional edge network intelligent monitoring, in particular to a cross-regional edge node congestion collaborative early warning method based on federated learning, which comprises the following steps of: initializing a sequence model at an edge node, constructing local loss combined with binary cross entropy and time sequence smooth punishment, and stably training by Adam and early stop; through weighted average aggregation disturbance parameters based on data volume and loss, performance degradation after fusion is prevented by adopting momentum fusion and a rollback threshold value; and meanwhile, an attention mechanism is embedded to realize joint prediction of probability, severity and duration, cross-regional propagation parameters of the graph neural network are utilized, loss is verified, and invalid sharing is rejected, so that prediction accuracy and timeliness are improved, and node congestion is effectively relieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of cross-regional edge networks, and particularly relates to a cross-regional edge node congestion collaborative early warning method based on federated learning. BACKGROUND

[0002] With the rapid development of 5G and edge computing technology, cross-regional network systems are facing the challenge of real-time processing of massive data; federated learning, as a distributed machine learning paradigm, has been widely applied in edge networks to realize model training under data privacy protection, such as congestion prediction in the fields of intelligent transportation and the Internet of Things; this technology avoids the risk of centralized data transmission through local computing and global aggregation, and improves the processing capability of time series data by combining sequence models.

[0003] However, the current technology still has significant defects, traditional centralized learning is prone to privacy leakage and single point failure, and the existing federated learning framework has insufficient generalization performance when processing heterogeneous data, ignoring the cross-regional correlation, resulting in low congestion detection accuracy; in addition, the local model training lacks differential privacy mechanism and is easily disturbed by noise, and the global aggregation does not consider the contribution difference of nodes, causing unstable model fusion; these problems are particularly prominent in high dynamic network environment, limiting the timeliness and reliability of real-time early warning. SUMMARY

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present application provides a cross-regional edge node congestion collaborative early warning method based on federated learning, which aims at the problem that the time series dependency of time series data is ignored during the initialization of the federated learning framework in the prior art, resulting in a slow response of the local prediction model to dynamic network changes; this scheme initializes a sequence model based on at each edge node, and defines a local loss function combining binary cross-entropy and time series smoothing penalty, achieving accurate time series modeling of network performance indicators; for the problem that the data heterogeneity and contribution difference between nodes are not fully considered during global model aggregation in the prior art, resulting in insufficient generalization performance of the global model, this scheme collects disturbance parameters for weighted average aggregation based on data volume and loss value to generate a high-quality global model, and adopts a momentum optimization mechanism to smooth the fusion of global and local parameters, while setting a loss threshold rollback mechanism; for the problem that congestion detection is limited to local data processing in the prior art, ignoring cross-regional correlation leads to prediction isolation and lack of global optimization, this scheme integrates an model with attention mechanism to realize comprehensive prediction of real-time probability, severity and duration, and uses graph neural network to propagate adjacent region parameters for cross-regional collaborative adjustment, while applying validation loss constraint to reject invalid sharing.

[0005] The technical scheme adopted by the present application is as follows: a cross-regional edge node congestion cooperative early warning method based on federated learning, which comprises the following steps:

[0006] Step S1: data set construction, real-time collection of network performance index data on each cross-regional edge node, construction of a data vector in the form of a time series, subsequent normalization processing and introduction of an adaptive threshold, identification and deletion of noise samples;

[0007] Step S2: federated learning framework initialization, initialization of a sequence model based on on each edge node for local congestion prediction, the model comprising an input layer, a hidden layer and an output layer, and setting initial parameters, defining a local training loss function of each edge node, the function combining binary cross-entropy and time series smoothing penalty regularization term, and using an optimizer for iterative optimization;

[0008] Step S3: local model training, gradient descent training using local normalized data on each edge node, updating model parameters, while introducing an early stopping mechanism and L2 norm clipping strategy, then, after local training, adding Gaussian noise to the model parameters, and adaptively adjusting the noise level according to the regional data sensitivity;

[0009] Step S4: global model aggregation, collecting perturbed parameters of all edge nodes, setting weights based on the local data volume and loss value of each node, generating a global model through weighted averaging, then sending the global model back to each edge node and fusing with the local parameters, and using a momentum optimization mechanism to smooth the fusion process;

[0010] Step S5: congestion detection, using the fused model to predict congestion in real-time data on each edge node, integrating an attention mechanism into the model, outputting congestion probability, severity and predicted duration, then using a graph neural network to represent regional relationships, realizing cross-regional cooperation by propagating model parameters of adjacent regions, adjusting the local model and integrating a global view;

[0011] Step S6: congestion early warning, based on the prediction result to determine whether to trigger an early warning, if the probability and severity exceed the threshold, generating an early warning information, sending the early warning information to high-risk areas according to priority, and considering local probability and neighborhood influence in priority.

[0012] Further, in step S1: the data set construction specifically comprises the following steps:

[0013] Step S11: Data collection, real-time collection of network performance index data on each cross-regional edge node, including network traffic, bandwidth usage, delay, packet loss rate, CPU and memory usage, data collection is in time series form, and a data vector is constructed;

[0014] Step S12: Data normalization, introducing adaptive threshold detection based on sliding window to identify noise, and converting data to standard distribution form with mean of 0 and variance of 1 through normalization processing.

[0015] Further, in step S2, the federated learning framework is initialized, specifically including the following steps:

[0016] Step S21: Distributed deep learning model initialization, initializing a deep learning model on each edge node, using Sequence model processes time series data for local congestion prediction, the model includes an input layer, a hidden layer and an output layer;

[0017] Step S22: Definition of local loss function, adding time series smoothing penalty, defining the local training loss function of each edge node, combining binary cross-entropy and regularization term.

[0018] Further, in step S3, the local model training, specifically including the following steps:

[0019] Step S31: Local gradient descent training, using local normalized data for model training on each edge node, while introducing early stopping mechanism, updating model parameters through gradient descent algorithm;

[0020] Step S32: Differential privacy parameter perturbation, after local training, adding Gaussian noise to model parameters to achieve differential privacy protection, and adjusting noise level adaptively according to regional data sensitivity.

[0021] Further, in step S4, the global model aggregation, specifically including the following steps:

[0022] Step S41: Parameter aggregation, collecting perturbed parameters of all edge nodes, performing weighted average to generate global model, introducing contribution weight, adjusting weight according to local data volume of each node;

[0023] Step S42: Local fusion, sending the global model back to each edge node and fusing with local parameters, using momentum optimization mechanism to smooth the fusion process.

[0024] Further, in step S5, the congestion detection, specifically including the following steps:

[0025] Step S51: Real-time congestion probability prediction, using the fused model on each edge node to predict congestion for real-time data, integrating attention mechanism into The output probability, severity and predicted duration are outputted.

[0026] Step S52: Cross-regional parameter sharing, using a graph neural network to represent regional relationships and propagate parameters, achieving cross-regional collaboration by sharing parameters of adjacent regions, and adjusting local models to incorporate a global view.

[0027] Further, in step S6, the congestion warning specifically includes the following steps:

[0028] Step S61: Warning threshold judgment, based on the prediction result to determine whether to trigger a warning, if the probability exceeds the threshold, a warning information containing time, location and duration is generated;

[0029] Step S62: Warning distribution, sending the generated warning information to high-risk areas according to priority.

[0030] The beneficial effects achieved by the above-mentioned scheme are as follows:

[0031] (1) In view of the problem that the existing technology ignores the time series dependence of time series data during the initialization of the federated learning framework, resulting in a slow response of the local prediction model to dynamic network changes, the scheme initializes a sequence model based on at each edge node, and defines a local loss function combining binary cross-entropy and time series smoothing penalty, achieving accurate time series modeling of network performance indicators, improving the accuracy and robustness of local congestion prediction, and ensuring the efficiency and stability of the training process through Adam optimizer and iteration stopping condition.

[0032] (2) In view of the problem that the existing technology does not fully consider data heterogeneity and contribution difference between nodes during global model aggregation, resulting in insufficient generalization performance of the global model, the scheme collects disturbance parameters for weighted average aggregation based on data volume and loss value, generates a high-quality global model, and uses a momentum optimization mechanism to smoothly integrate global and local parameters, while setting a loss threshold rollback mechanism to avoid performance degradation after fusion, improving the convergence speed and model adaptability of the overall federated learning.

[0033] (3) In view of the problem that congestion detection in the prior art is limited to local data processing, ignoring cross-regional association leading to prediction isolation and lack of global optimization, the scheme integrates an attention mechanism into The model realizes real-time probability, severity and duration comprehensive prediction, and utilizes a graph neural network to propagate adjacent area parameters for cross-area collaborative adjustment, while applying a verification loss constraint to reject invalid sharing, thereby improving the accuracy, timeliness and overall collaborative efficiency of detection. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 A schematic diagram of the cross-area edge node congestion collaborative early warning method based on federated learning provided by the present application is shown in the figure.

[0035] Figure 2 A schematic diagram of step S1 is shown in the figure.

[0036] Figure 3 A schematic diagram of step S2 is shown in the figure.

[0037] Figure 4 A schematic diagram of step S3 is shown in the figure.

[0038] Figure 5 A schematic diagram of step S4 is shown in the figure.

[0039] Figure 6 A schematic diagram of step S5 is shown in the figure.

[0040] Figure 7 A schematic diagram of step S6 is shown in the figure.

[0041] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] In the description of the present application, it should be understood that the terms “up”, “down”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inner”, “outer” and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the present application and simplify the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0044] Embodiment one, refer to Figure 1 The cross-area edge node congestion collaborative early warning method based on federated learning provided by the present application comprises the following steps:

[0045] Step S1: Dataset construction. Real-time network performance index data is collected at each cross-regional edge node, and a data vector is constructed in the form of time series. Then, the data is normalized and an adaptive threshold is introduced to identify and remove noisy samples.

[0046] Step S2: Federated learning framework initialization, initializing a framework based on... A sequence model for local congestion prediction, comprising an input layer, The layer and output layer are defined, and initial parameters are set. A local training loss function is defined for each edge node, combining binary cross-entropy with a time-series smoothing penalty regularization term. The optimizer performs iterative optimization;

[0047] Step S3: Local model training. Gradient descent training is performed on local normalized data at each edge node to update the model parameters. At the same time, an early stopping mechanism and L2 norm pruning strategy are introduced. Subsequently, after local training, Gaussian noise is added to the model parameters, and the noise level is adaptively adjusted according to the sensitivity of regional data.

[0048] Step S4: Global model aggregation. Collect the perturbation parameters of all edge nodes, set weights based on the local data volume and loss value of each node, generate a global model by weighted averaging, and then send the global model back to each edge node and fuse it with the local parameters. Use momentum optimization mechanism to smooth the fusion process.

[0049] Step S5: Congestion Detection. At each edge node, the fused model is used to predict congestion in real-time data, incorporating an attention mechanism. The model outputs congestion probability, severity, and expected duration. Subsequently, a graph neural network is used to represent regional relationships. Cross-regional collaboration is achieved by propagating model parameters from neighboring regions, adjusting the local model, and integrating it into the global view.

[0050] Step S6: Congestion warning. Based on the prediction results, determine whether to trigger a warning. If the probability and severity exceed the threshold, generate a warning message and send the warning message to high-risk areas according to priority, taking into account local probability and neighborhood impact.

[0051] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the construction of the dataset specifically includes the following steps:

[0052] Step S11: Data Acquisition. Real-time network performance metrics data are collected at each cross-regional edge node, including network traffic, bandwidth utilization, latency, packet loss rate, CPU and memory usage. The data is collected in time series format and a data vector is constructed, represented as follows:

[0053] ;

[0054] in, Indicates the node region index. Represents the node region In time The collected data vector, This represents network traffic, measured in megabits per second (Mbps). ; Indicates bandwidth utilization. Indicates latency, in milliseconds. ; Indicates packet loss rate; express Utilization rate; Indicates memory usage; This indicates the transpose symbol; the data collection frequency is once every 5 seconds, and the data is stored in a local buffer. The buffer size is the data from the most recent hour. A preset congestion threshold is used; if any indicator exceeds the preset threshold, it is initially marked as a potential congestion label. ,otherwise ;

[0055] Step S12: Data normalization. An adaptive threshold detection based on a sliding window is introduced to identify noise. The data is transformed into a standard distribution with a mean of 0 and a variance of 1 through normalization, as shown below:

[0056] ;

[0057] in, This represents the normalized data vector. Indicates the area In the time window The mean vector within, Indicates the area Standard deviation vector, time window Corresponding to 1 hour Indicates the division by zero factor; noise monitoring rule: for each component, if If it is, then it is considered noise and deleted.

[0058] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the federated learning framework is initialized, specifically including the following steps:

[0059] Step S21: Distributed deep learning model initialization. A deep learning model is initialized on each edge node, using... Sequence models process time-series data for local congestion prediction. The model includes an input layer, The initialization of the layer and output layer is as follows:

[0060] ;

[0061] in, express The hidden state vector needs to be passed to the next time step. This represents the hidden state at the previous time step, initially set as a zero vector; Indicates the area initial The parameter set includes the weight matrix and biases; This represents the predicted congestion probability, with values ​​ranging from... between, express Activation function and These represent the output layer weight vector and bias, respectively. The layer dimension is 64, and the output layer is a single neuron;

[0062] Step S22: Define the local loss function, add a time series smoothing penalty, define the local training loss function for each edge node, and combine the binary cross-entropy and regularization term, as follows:

[0063] ;

[0064] in, Indicates the area The local loss needs to be minimized; Indicates the actual congestion label. Indicates the predicted probability. This represents the smoothing penalty coefficient. Represents the natural logarithm; Set constraints: All Optimize usage Optimizer, Learning Rate Iterate until the loss decreases to less than Or a maximum of 50 iterations.

[0065] By performing the above operations, this solution addresses the problem in existing federated learning frameworks where the initialization of time-series data ignores the temporal dependencies of the data, leading to a sluggish response of the local prediction model to dynamic network changes. Instead, this solution initializes a time-series-based model at each edge node. The sequence model is proposed, and a local loss function combining binary cross-entropy and time series smoothing penalty is defined to achieve accurate time-series modeling of network performance indicators, thereby improving the accuracy and robustness of local congestion prediction. At the same time, the efficiency and stability of the training process are ensured by using the Adam optimizer and iterative stopping conditions.

[0066] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the local model training specifically includes the following steps:

[0067] Step S31: Local gradient descent training. The model is trained using local normalized data at each edge node, and an early stopping mechanism is introduced to update the model parameters using the gradient descent algorithm, as shown below:

[0068] ;

[0069] in, Indicates the area In the The parameter set of the round iteration, Indicates the current parameter. The gradient of the loss with respect to the parameters is represented using backpropagation; the training dataset is the nearest... Data, validation set is remaining Early stopping condition: If the verification loss increases for three consecutive rounds, then stop; Set constraints: After parameter update, proceed. Norm clipping ensures ;

[0070] Step S32: Differential privacy parameter perturbation. After local training, Gaussian noise is added to the model parameters to achieve differential privacy protection. The noise level is adaptively adjusted according to the sensitivity of regional data, as shown below:

[0071] ;

[0072] in, This represents the parameters after the perturbation. This indicates that the mean is 0 and the variance is 0. Gaussian noise vector, It is the identity matrix; Indicates the standard deviation of noise. Indicate privacy parameters and set adaptive rules: if the regional data variance ,but Multiply by 0.5 to reduce noise.

[0073] Example 5, see Figure 1 and Figure 5This embodiment is based on the above embodiment. In step S4, the global model aggregation specifically includes the following steps:

[0074] Step S41: Parameter aggregation. Collect the perturbation parameters of all edge nodes, perform weighted averaging, generate a global model, introduce contribution weights, and adjust the weights according to the amount of local data for each node, as shown below:

[0075] ;

[0076] in, This represents a global parameter that needs to be sent back to the node. Indicates the total number of regions; Indicates originating from region The disturbance parameters; This represents the contribution weight; after normalization, the sum of the weights is 1. Indicates the area Number of data samples;

[0077] Step S42: Local fusion. The global model is sent back to each edge node and fused with the local parameters. The fusion process is smoothed using a momentum optimization mechanism, as shown below:

[0078] ;

[0079] in, This represents the initial parameters after fusion, which will be used in the next round of local training. This indicates the latest local parameters. This represents the fusion coefficient; at the same time, a constraint is set: after fusion, the local loss is recalculated, and if it is higher than the threshold of 0.5, it is rolled back to the local parameters.

[0080] By performing the above operations, this solution addresses the problem in existing technologies where the global model aggregation does not fully consider the heterogeneity of data and differences in contributions between nodes, resulting in insufficient generalization performance of the global model. This solution generates a high-quality global model by collecting perturbation parameters and performing weighted average aggregation based on data volume and loss value. It also employs a momentum optimization mechanism to smoothly fuse global and local parameters and sets a loss threshold rollback mechanism to avoid performance degradation after fusion, thereby improving the convergence speed and model adaptability of the overall federated learning.

[0081] Example 6, see Figure 1 and Figure 6 This embodiment is based on the above embodiment. In step S5, the congestion detection specifically includes the following steps:

[0082] Step S51: Real-time congestion probability prediction. The fused model is used on each edge node to predict congestion based on real-time data, integrating the attention mechanism into... In the output, the probability, severity, and expected duration are represented as follows:

[0083] ;

[0084] in, This indicates the probability of congestion. This represents the attention matrix, and the elements in the matrix are... q and k represent the query and key vectors, respectively, from projection; The score represents the severity level, and its value ranges from [value missing]. ; Represents an exponential function with the natural constant as its base; , Indicates the fractional adjustment factor; expected duration Through forward simulation, the maximum simulation time is 5 minutes, with each step lasting 5 seconds. Used to find the independent variable that makes the function take the maximum value;

[0085] Step S52: Cross-regional parameter sharing. A graph neural network is used to represent regional relationships and propagate parameters. Cross-regional collaboration is achieved by sharing parameters from adjacent regions. The local model is adjusted to integrate with the global view, as shown below:

[0086] ;

[0087] in, Indicates the adjusted parameters; Indicates the area The set of adjacent regions, specifically those with a geographical distance of less than [a certain value]. The area within the range, Indicates propagation weight, Indicates the area and region Geographical distance between them Set constraints: the adjusted verification loss shall not exceed 1.1 times the original loss; otherwise, refuse to share.

[0088] By performing the above operations, this solution addresses the problems in existing technologies where congestion detection is limited to local data processing and ignores cross-regional correlations, leading to isolated predictions and a lack of global optimization. It integrates an attention mechanism... The model achieves comprehensive prediction of real-time probability, severity, and duration, and uses graph neural networks to propagate parameters from adjacent regions for cross-regional collaborative adjustment. At the same time, it applies validation loss constraints to reject invalid sharing, thereby improving the accuracy, timeliness, and overall collaborative efficiency of the detection system.

[0089] Example 7, see Figure 1 and Figure 7This embodiment is based on the above embodiment. In step S6, the congestion warning specifically includes the following steps:

[0090] Step S61: Warning threshold determination. Based on the prediction results, determine whether to trigger a warning. If the probability exceeds the threshold, generate warning information including time, location, and duration, as shown below:

[0091] ;

[0092] in, This indicates a warning signal; 1 indicates triggering. Indicates the probability threshold. Indicates the severity threshold. Indicates the confidence interval. express Confidence level, The number of samples used for estimation is determined by truncating the upper and lower bounds when selecting the interval. The warning information includes: time. Node region index Duration Confidence interval ;

[0093] Step S62: Warning distribution. The generated warning information is sent to high-risk areas according to priority, as shown below:

[0094] ;

[0095] in, Indicates the distribution priority; messages are sent according to priority. To indicate the influence of the neighborhood, set constraints: prioritize sending. The area is checked once per minute.

[0096] 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 apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0098] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A cross-regional edge node congestion collaborative early warning method based on federated learning, characterized in that: The method includes the following steps: Step S1: Dataset construction. Real-time network performance index data is collected at each cross-regional edge node, and a data vector is constructed in the form of time series. Then, the data is normalized and an adaptive threshold is introduced to identify and remove noisy samples. Step S2: Federated learning framework initialization, initializing a framework based on... The sequence model is constructed, and initial parameters are set, defining the local training loss function for each edge node; Step S3: Local model training. Gradient descent training is performed on local normalized data at each edge node to update the model parameters. At the same time, an early stopping mechanism and L2 norm pruning strategy are introduced. Subsequently, after local training, Gaussian noise is added to the model parameters, and the noise level is adaptively adjusted according to the sensitivity of regional data. Step S4: Global model aggregation. Collect all edge node parameters with added noise, set weights based on the local data volume and loss value of each node, generate a global model by weighted averaging, and then send the global model back to each edge node and fuse it with the local parameters. Use momentum optimization mechanism to smooth the fusion process. Step S5: Congestion Detection. At each edge node, the fused model is used to predict congestion in real-time data. The model incorporates an attention mechanism to output congestion probability, severity, and expected duration. Subsequently, a graph neural network is used to represent regional relationships, and cross-regional collaboration is achieved by propagating model parameters from adjacent regions, adjusting the local model, and integrating it into the global view. Step S6: Congestion warning. Based on the prediction results, determine whether to trigger a warning. If the probability or severity exceeds the threshold, generate a warning message and send the warning message to high-risk areas according to priority, taking into account local probability and neighborhood impact.

2. The cross-regional edge node congestion collaborative early warning method based on federated learning according to claim 1, characterized in that: In step S2, the federated learning framework is initialized, specifically including the following steps: Step S21: Distributed deep learning model initialization. A deep learning model is initialized on each edge node, using... Sequence models process time-series data for local congestion prediction. The model includes an input layer, Layers and output layers; Step S22: Define the local loss function, add a time series smoothing penalty, define the local training loss function for each edge node, and combine binary cross-entropy and regularization terms.

3. The cross-regional edge node congestion collaborative early warning method based on federated learning according to claim 1, characterized in that: In step S3, the local model training specifically includes the following steps: Step S31: Local gradient descent training. The model is trained using local normalized data at each edge node, and an early stopping mechanism is introduced to update the model parameters through the gradient descent algorithm. Step S32: Differential privacy parameter perturbation. After local training, Gaussian noise is added to the model parameters to achieve differential privacy protection, and the noise level is adaptively adjusted according to the sensitivity of regional data.

4. The cross-regional edge node congestion collaborative early warning method based on federated learning according to claim 1, characterized in that: In step S4, the global model aggregation specifically includes the following steps: Step S41: Parameter aggregation: Collect the perturbation parameters of all edge nodes, perform weighted averaging, generate a global model, introduce contribution weights, and adjust the weights according to the amount of local data of each node; Step S42: Local fusion. The global model is sent back to each edge node and fused with the local parameters. The fusion process is smoothed using a momentum optimization mechanism.

5. The cross-regional edge node congestion collaborative early warning method based on federated learning according to claim 1, characterized in that: In step S5, the congestion detection specifically includes the following steps: Step S51: Real-time congestion probability prediction. The fused model is used on each edge node to predict congestion based on real-time data, integrating the attention mechanism into... The output includes the probability, severity, and expected duration. Step S52: Cross-regional parameter sharing. Use graph neural networks to represent regional relationships, propagate parameters, achieve cross-regional collaboration by sharing parameters of adjacent regions, and adjust the local model to integrate into the global view.

6. The cross-regional edge node congestion collaborative early warning method based on federated learning according to claim 1, characterized in that: In step S6, the congestion warning specifically includes the following steps: Step S61: Warning threshold determination. Based on the prediction results, determine whether to trigger a warning. If the probability exceeds the threshold, generate warning information including time, location and duration. Step S62: Warning distribution, sending the generated warning information to high-risk areas according to priority.

7. The cross-regional edge node congestion collaborative early warning method based on federated learning according to claim 1, characterized in that: In step S1: The dataset construction specifically includes the following steps: Step S11: Data collection. Real-time network performance metrics data are collected at each cross-regional edge node, including network traffic, bandwidth utilization, latency, packet loss rate, CPU and memory usage. The data collection is in time series form, and a data vector is constructed. Step S12: Data normalization. An adaptive threshold detection based on a sliding window is introduced to identify noise. The data is transformed into a standard distribution with a mean of 0 and a variance of 1 through normalization processing.

Images