A multi-layer split federated learning automatic modulation classification identification method based on a ring topology

Abstract

This invention relates to a multi-layer split federated learning automatic modulation classification and recognition method based on ring topology, belonging to the field of wireless communication signal processing and edge intelligent computing technology. It includes signal model construction and data preprocessing, deployment of a multi-layer split federated learning architecture, execution of local training based on ring topology, introduction of a dynamic regularization mechanism with adaptive signal-to-noise ratio in local training, weighted aggregation by the edge server to generate a local model, and a blockchain consensus layer executing consensus based on diversity-based node quality proofs for global model updates and recognition result output. The advantages are that a ring-based collaborative training mechanism is introduced at the terminal monitoring station layer and the edge server layer, and a quality evaluation mechanism is introduced at the blockchain consensus layer, effectively solving the communication bottleneck and model bias problem caused by heterogeneous monitoring equipment in federated learning, significantly reducing training latency while improving the accuracy and real-time performance of automatic modulation classification and recognition.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication signal processing and edge intelligent computing technology, specifically to a multi-layer split federated learning automatic modulation classification and recognition method based on ring topology, particularly a dynamic regularization mechanism with adaptive signal-to-noise ratio and a consensus mechanism based on diversity quality proof of blockchain nodes, to improve the accuracy and training efficiency of automatic modulation recognition in heterogeneous radio monitoring scenarios. Background Technology

[0002] With the exponential growth in the number of wireless devices, the electromagnetic spectrum environment faces increasingly congested and complex challenges. As a core function supporting cognitive radio operation, automatic modulation classification and identification technology can intelligently identify the modulation type of received signals under non-cooperative conditions; this technology is a key supporting means for achieving efficient spectrum utilization, interference management, and ensuring wireless communication security. Currently, deep learning-based automatic modulation classification methods, with their powerful feature extraction capabilities, have significantly surpassed traditional methods relying on instantaneous statistics or higher-order cumulants in recognition performance. However, the data-driven model faces multiple challenges in practical deployment: traditional centralized monitoring architectures require distributed sensing nodes to transmit all raw I / Q sampling data back to a central server for unified training; centralized architectures not only incur high backhaul bandwidth overhead and network congestion but also carry privacy risks such as sensitive data leakage, and the resulting processing latency is difficult to meet the real-time requirements of the task. To overcome the limitations of centralized training, Federated Learning (FL), as a privacy-preserving distributed collaborative paradigm, has been introduced into the field of wireless communication, allowing terminals to collaborate by sharing only model updates rather than raw data.

[0003] While federated learning offers significant advantages in communication efficiency and privacy protection, its direct application to wireless automatic modulation and classification tasks still faces several bottlenecks: First, multipath fading and background noise in wireless channels lead to significant differences in the signal-to-noise ratio (SNR) of signals collected by different monitoring station nodes. Traditional federated learning algorithms typically assume uniform data quality across nodes, thus ignoring the negative interference of noise gradients generated by low SNR nodes on the convergence of the global model. Second, due to the uneven spatial distribution of signal types, the data from sensing nodes in different geographical locations exhibit highly non-independent and identically distributed characteristics, which severely restricts the generalization ability of the recognition model. Third, terminal monitoring station equipment is often limited by computing power and energy consumption, and the star topology architecture used in traditional federated learning is difficult to adapt to the hardware heterogeneity of terminal devices, easily leading to the problem of "lagging" devices slowing down the overall training process. Fourth, although the introduction of blockchain technology can improve security, the excessive computational overhead of mainstream consensus mechanisms makes them difficult to deploy in resource-constrained edge environments, and existing consensus protocols focus more on transaction verification and ledger consistency, lacking evaluation and screening mechanisms for model quality. Summary of the Invention

[0004] This invention provides a multi-layer split-fed federated learning automatic modulation classification and recognition method based on ring topology to address the problems of uneven data quality among nodes, non-independent and identically distributed data, training limitations due to lagging nodes, and the lack of evaluation and screening mechanisms in blockchains in heterogeneous radio monitoring environments. By synergistically employing ring topology-based split-fedded learning, adaptive signal-to-noise ratio dynamic regularization, and blockchain node quality proof consensus technology based on diversity, efficient and robust signal recognition is achieved.

[0005] The technical solution adopted by this invention includes the following steps:

[0006] (1) Perform signal model construction and data preprocessing, extract the received continuous radio frequency signal into in-phase and quadrature components, and construct a two-dimensional tensor as the input of the deep learning model;

[0007] (2) Deploy a multi-layered split federated learning architecture and build a local ring monitoring network composed of terminal monitoring station devices under each edge server;

[0008] (3) Perform local training based on ring topology. Each terminal monitoring station device is only responsible for calculating and updating some parameters of the local model and relaying the intermediate calculation results in the communication link within the ring.

[0009] (4) Introduce a dynamic regularization mechanism with adaptive signal-to-noise ratio in local training. By sensing the signal-to-noise ratio of the data in real time, the regularization constraint is dynamically adjusted to suppress the pollution of the global model by noise interference.

[0010] (5) After local training is completed, the terminal monitoring station device will upload the complete local model parameters, and the corresponding edge server will perform weighted aggregation to generate a local model;

[0011] (6) The blockchain consensus layer executes a node quality proof consensus based on diversity, switches between quality proof score and diversity score according to the current learning rate, evaluates, filters and weights each local model, and then aggregates to generate a global model.

[0012] (7) Perform global model update and recognition result output. Each blockchain node performs performance verification and block storage on the generated global model, and securely sends the parameters to the underlying network to start the next round of iteration or output classification judgment.

[0013] The signal model construction and data preprocessing process in step (1) of this invention is as follows:

[0014] The receiver converts continuous radio frequency signals into discrete-time complex baseband signals. , is represented as:

[0015]

[0016] in, This is the sampling sequence number. The original modulated signal from the transmitting end. This is the multipath fading impulse response function. For additive white Gaussian noise, for the first The first signal sample There are several sampling points, and the following relationships exist:

[0017]

[0018] in, The imaginary unit, For the first A discrete-time complex baseband signal sample This is the in-phase component of the signal sample. These are the orthogonal components of the signal sample;

[0019] To enable a deep learning model to learn the temporal features of a signal, a segment of length is extracted from the continuous sequence. A window of observation, which will... The in-phase component sequence corresponding to each sampling point and orthogonal component sequences The top and bottom edges are concatenated to form a matrix, which is then transposed to ultimately construct a matrix with dimension [dimensional value missing]. real two-dimensional tensors :

[0020]

[0021] The process of deploying the multi-layer split federated learning architecture based on ring topology in step (2) of this invention is as follows:

[0022] The architecture consists of a terminal monitoring station layer, an edge server layer, and a blockchain consensus layer from bottom to top, wherein the terminal monitoring station layer is deployed... Each terminal monitoring station is equipped with a local ring monitoring network, which is constructed based on geographical location and actual spectrum environment characteristics. The terminal monitoring stations within the local ring monitoring network are connected by communication links, and each terminal monitoring station is maintained using a local model with the assistance of other terminal monitoring stations within the local ring monitoring network. The edge server layer is deployed... Each edge server is responsible for managing all terminal monitoring station devices within its local ring monitoring network in the terminal monitoring station layer. The edge server and the local terminal monitoring station devices it manages jointly maintain a local model. The blockchain consensus layer consists of several decentralized blockchain nodes, each blockchain node corresponding to an edge server in the edge server layer. All edge servers jointly maintain a global model through the blockchain nodes.

[0023] In this system, the terminal monitoring station devices in the terminal monitoring station layer employ a ring-topology-based split federated learning approach for local model training. Each terminal monitoring station device is only responsible for calculating and updating a portion of the parameters of its local model. Other parameters are calculated by other terminal monitoring station devices within its local ring monitoring network. Each terminal monitoring station device uses local data to calculate and update a portion of the parameters of its local model and sequentially transmits the generated forward activation values ​​or backward gradients along the communication link to other terminal monitoring station devices within its local ring monitoring network, until all terminal monitoring station devices within the local ring monitoring network have been traversed. The training task of each terminal monitoring station device is jointly determined by its position within the local ring monitoring network and the number of training layers. The edge server layer is responsible for managing the collaborative training and aggregation of local models by the terminal monitoring station devices. Each edge server manages a local ring monitoring network and predefines a "propagation step size" for all terminal monitoring station devices within that local ring monitoring network. This propagation step size determines the number of training layers for each terminal monitoring station device when collaboratively training its local model. The blockchain consensus layer is responsible for global model evaluation, consensus filtering, and weighted aggregation across edge servers. It records global model updates through a distributed ledger to ensure the generalization and security of the global model.

[0024] The local training process of split federated learning based on ring topology in step (3) of this invention is as follows:

[0025] When each terminal monitoring station device in the local ring monitoring network starts local training, it needs to obtain complete initial global model parameters from the blockchain node through the edge server. In each training cycle, each terminal monitoring station device can serve as the starting point for training the local model and initiate a collaborative training in its local ring monitoring network. Based on the division of the local model hierarchy, the complete process of a single collaborative training is dynamically divided into several parts, including forward propagation and backward propagation, which are completed sequentially by each terminal monitoring station device in the local ring monitoring network.

[0026] During training, each terminal monitoring station uses local data to train and update the parameters of its assigned local model. Simultaneously, it transmits the forward activation values ​​and backward gradients of the intermediate results sequentially along the communication link to other terminal monitoring stations within the local ring monitoring network until all terminal monitoring stations within the local ring monitoring network jointly complete one iterative optimization of the local model. A local empirical risk function is used to measure the loss of each terminal monitoring station's local model on the training set. This function consists of a local loss function term and a regularization term. The local loss function term is used to fit the local data features, while the regularization term is used to constrain the deviation of the local model parameters, preventing the local model from deviating excessively from the global model.

[0027] The forward propagation described in this invention is as follows: When a terminal monitoring station device in a local ring monitoring network initiates collaborative training using its local data as the training starting point, the terminal monitoring station device completes the training of a portion of the local model based on the received global model and training task and generates forward activation values. These forward activation values ​​are then sequentially transmitted along the communication link to other terminal monitoring station devices within the local ring monitoring network. The other terminal monitoring station devices train a portion of the local model based on the received global model, training task, and forward activation values ​​and generate their corresponding forward activation values, thus completing the forward propagation of the local model training.

[0028] The backpropagation is as follows: After all terminal monitoring station devices in the local ring monitoring network have completed forward propagation, they calculate the loss function and transmit the generated gradients back to other terminal monitoring station devices in sequence along the communication link. After receiving the backpropagated gradients, the other terminal monitoring station devices update the local model parameters they are responsible for, thus completing the backpropagation of the local model training.

[0029] The local empirical risk function described in this invention: the first in the local ring monitoring network The local model training objective for each terminal monitoring station device is to minimize its local empirical risk function. :

[0030]

[0031] in, For the first The local dataset held by each terminal monitoring station For a single training sample and its corresponding label, where The input is the original I / Q signal tensor. This is the modulation scheme label corresponding to the signal. For the trained neural network model, These are the local model parameters that the terminal monitoring station equipment is currently training or updating. Let cross-entropy be the loss function. These are the global model parameters distributed to the blockchain nodes via the edge server. Regularization coefficients are used to constrain the deviation of the local model during this batch of training.

[0032] The dynamic regularization of signal-to-noise ratio adaptation in step (4) of this invention is expressed as follows:

[0033] The local model extracts the current training batch in real time during each training iteration. Average signal-to-noise ratio :

[0034]

[0035] in, The size of the current training batch. Indicates training batch The Middle Signal-to-noise ratio of each signal sample;

[0036] The regularization coefficient for this batch was then calculated. :

[0037]

[0038] in, The basic regularization constant; For the amplitude limiting function, The attenuation factor is used to control the sensitivity to signal-to-noise ratio; and The lower and upper limits of the regularization strength are defined respectively, when the average signal-to-noise ratio When it is lower, the corresponding exponential term Larger, leading to It is also relatively large; conversely, when the average signal-to-noise ratio is relatively small... At higher levels, The local model is also correspondingly smaller, allowing it to extract fine features of the signal with greater exploratory freedom.

[0039] The process of generating a local model through local model aggregation in step (5) of this invention is as follows:

[0040] After each terminal monitoring station within the local ring monitoring network has traversed the local dataset, it signifies that the local ring monitoring network has completed one round of local training. Each terminal monitoring station will then transmit its generated local model parameters. The data is uploaded to the edge server that manages the local ring monitoring network via a communication link;

[0041] Edge server After collecting complete local models submitted by all terminal monitoring stations within the local ring monitoring network, a weighted aggregation operation is performed to merge and generate a local model of the local ring monitoring network. Its mathematical expression is:

[0042]

[0043] in, This is the set of terminal monitoring stations managed by the edge server. For the first The aggregation weight assigned to each terminal monitoring station device is determined based on the proportion of local data samples of that terminal monitoring station device, and satisfies the following conditions: Once the aggregation is complete, this local model will participate in the quality proof consensus of the subsequent blockchain consensus layer.

[0044] The consensus mechanism for proving the quality of blockchain nodes based on diversity in step (6) of this invention is as follows:

[0045] The diversity-based node quality proof consensus relies on a public verification set pre-deployed on each blockchain node. The public verification set serves as a unified performance metric and is divided into multiple non-overlapping verification subsets, which are then distributed to each blockchain node. Each blockchain node calculates the local score of its local model on its respective verification subset and exchanges and summarizes these local scores through a consensus algorithm to obtain the aggregate weight of each local model.

[0046] After blockchain nodes collect all local models, these models may exhibit significant biases due to the non-independent and identically distributed nature of the data. To improve the generalization ability of the global model, the quality of each generated local model needs to be evaluated. This is achieved using a diversity-based node quality proof consensus mechanism. This mechanism evaluates the models from two dimensions: local model quality and diversity, defining corresponding quality proof scores and diversity scores. Each blockchain node selects between the quality proof score and the diversity score, using its current learning rate (LR) as the metric.

[0047] Quality proof score: Let the switching threshold be... ,when At this point, the blockchain node determines that the loss function of the global model is in a period of rapid decline, and the blockchain node collects the local models submitted by each edge server. Then, based on the loss function values ​​on the common validation set... Calculate the quality proof score :

[0048]

[0049] Quality proof score This enables blockchain nodes to achieve initial and rapid convergence of the global model by assigning higher aggregation weights to low-loss local models.

[0050] Diversity score: when the conditions are met At this point, each blockchain node determines that the global model training has entered a plateau period and selects a local model diversity score. As an evaluation metric, the diversity score during the stable fine-tuning phase for:

[0051]

[0052] in, The arithmetic mean accuracy of the local model across all modulation categories. The standard deviation of the accuracy for each category is given. For the parameters of the local model, This is a preset dispersion penalty coefficient;

[0053] Each blockchain node assigns a corresponding aggregation weight to each local model based on the quality proof score or diversity score, and generates a new round of global model, thereby completing the global model update in a decentralized environment.

[0054] The global model update and recognition result output process in step (7) of this invention is as follows:

[0055] For the newly generated global model, each blockchain node first performs performance verification and stores the evidence on the blockchain. Each blockchain node will quickly evaluate the newly generated model on the public verification set. The resulting key performance indicators, along with the hash values ​​of the local model parameters, will be used as metadata for this aggregation and recorded in the new block. The public verification set is logically a unified performance metric. During physical deployment, it is divided into multiple non-overlapping verification subsets and distributed to each blockchain node. Each node exchanges the evaluation results based on its own subset, and finally establishes a globally consistent aggregation weight through a consensus algorithm. The immutability of the blockchain ensures the auditability of the entire decision-making and execution process from weight calculation to global model generation.

[0056] Each blockchain node performs a secure distribution operation of global model parameters to achieve a logical closed loop in the training process. After the complete block of the latest aggregation result is confirmed by each blockchain node, the blockchain consensus layer securely distributes the latest global model parameters to each edge server. Subsequently, each edge server synchronizes the received parameters to the underlying terminal monitoring station equipment under its jurisdiction and uses it as the starting point of the next training cycle. If the performance of the obtained global model is verified to meet the preset requirements, the blockchain node uses the current global model as the final output to complete the modulation recognition task.

[0057] The beneficial effects of this invention are:

[0058] 1. Enhanced model convergence stability and robustness in heterogeneous signal quality environments. Addressing the issues of significant signal-to-noise ratio (SNR) differences between nodes due to multipath fading and background noise in wireless channels, and the tendency of low-quality noise gradients to contaminate the global model, this invention innovatively introduces a dynamic regularization mechanism with adaptive SNR. This mechanism dynamically adjusts the regularization coefficient by sensing the average SNR of the training batch in real time. Under low SNR conditions, it strengthens the constraints on the local model to suppress noise interference, while under high SNR conditions, it relaxes the constraints to extract fine features, thereby ensuring that the model maintains a stable convergence trajectory even in harsh and time-varying electromagnetic environments.

[0059] 2. Significantly improves model generalization ability in scenarios with highly non-independent and identically distributed (Non-IID) data. Addressing the problem that uneven spatial distribution of signal types leads to highly non-independent and identically distributed data from sensing nodes, severely limiting model generalization ability, this invention designs a diversity-based node quality proof consensus technology at the blockchain consensus layer. This technology no longer relies solely on a loss function but introduces a diversity score as a penalty term. This causes blockchain nodes to prioritize local models that perform evenly across all modulation types when aggregating local models to generate a global model, effectively correcting the model bias caused by skewed data distribution and greatly improving the recognition accuracy for minority or easily confused signals. Even in extremely heterogeneous scenarios, the performance of the architecture only shows a slight decrease; under high signal-to-noise ratio conditions, the average accuracy of the global model remains above 93%, verifying the stability and usability of the method in extreme communication environments.

[0060] 3. Effectively mitigates the "fallback" effect caused by heterogeneous terminal device hardware, significantly reducing training latency. Addressing the issue that traditional star topologies, limited by computing power and energy consumption, easily lead to "fallback" nodes slowing down global training, this invention breaks with traditional communication topologies and innovatively introduces a ring-based split-fed learning framework. By dynamically allocating "propagation step size" within the local ring monitoring network, the computational load is distributed among heterogeneous devices. Devices with weaker computing power only undertake a small number of model layer calculations matching their processing capabilities, avoiding overall progress being constrained by the weakest node. This significantly improves concurrent training efficiency while maintaining high accuracy. Compared to standard parallel multi-layer federated learning schemes, this invention can reduce total training latency by approximately 70%.

[0061] 4. A decentralized edge security system with low overhead and a focus on model quality evaluation has been constructed. Addressing the shortcomings of mainstream blockchain consensus mechanisms, such as high computational overhead and a lack of model quality screening mechanisms, this invention deeply integrates multi-layered federated learning with lightweight blockchain technology. Relying on a pre-deployed public validation set, "model quality and diversity scores" replace energy-intensive computing power competition (such as traditional Proof-of-Work) as the consensus standard. This mechanism not only keeps computational overhead within the tolerance range of edge devices but also effectively identifies and isolates inferior model updates caused by sensor failures or poor channel conditions, ensuring the secure and stable operation of large-scale radio monitoring networks. Attached Figure Description

[0062] Figure 1 This is a diagram of the three-layer architecture described in this invention;

[0063] Figure 2 This is a flowchart of the automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning described in this invention.

[0064] Figure 3 This is an accuracy curve of the global model under different signal-to-noise ratios when parameter α=0.5 in this invention;

[0065] Figure 4 This is the confusion matrix under partial signal-to-noise ratio when parameter α=0.5 in this invention;

[0066] Figure 5 This is an accuracy curve of the global model under different α values ​​and different signal-to-noise ratios in this invention;

[0067] Figure 6 The graph shows the variation of the test accuracy of the three schemes with the number of communication rounds.

[0068] Figure 7 The graph shows the change in accuracy of the three test schemes over the time of the simulated wall clock. Detailed Implementation

[0069] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, which integrate a ring topology-based split federated learning framework, a signal-to-noise ratio adaptive dynamic regularization mechanism, and a diversity-based blockchain node quality proof technology.

[0070] Includes the following steps:

[0071] (1). Perform signal model construction and data preprocessing, extract the received continuous radio frequency signal into in-phase and quadrature components, and construct a two-dimensional tensor as the input of the deep learning model;

[0072] The receiver converts continuous radio frequency signals into discrete-time complex baseband signals. , is represented as:

[0073]

[0074] in, This is the sampling sequence number. The original modulated signal from the transmitting end. This is the multipath fading impulse response function. As additive white Gaussian noise, in communication principles, a complex baseband signal naturally consists of an in-phase component (I, real part) and a quadrature component (Q, imaginary part), that is, for the first... The first signal sample There are several sampling points, and the following relationships exist:

[0075]

[0076] in, The imaginary unit, For the first A discrete-time complex baseband signal sample This is the in-phase component of the signal sample. These are the orthogonal components of the signal sample;

[0077] To enable a deep learning model to learn the temporal features of a signal, a segment of length is extracted from the continuous sequence. A window of observation, which will... The in-phase component sequence corresponding to each sampling point and orthogonal component sequences The top and bottom edges are concatenated to form a matrix, which is then transposed to ultimately construct a matrix with dimension [dimensional value missing]. real two-dimensional tensors :

[0078]

[0079] (2) Deploy a multi-layered split federated learning architecture, and build a local ring monitoring network composed of terminal monitoring station devices under each edge server; refer to Figure 1 This embodiment constructs a multi-layer split federated learning automatic modulation classification and recognition architecture based on ring topology:

[0080] The architecture, from bottom to top, consists of a terminal monitoring station layer, an edge server layer, and a blockchain consensus layer. The terminal monitoring station layer is deployed... Each terminal monitoring station is equipped with a local ring monitoring network, which is constructed based on geographical location and actual spectrum environment characteristics. The terminal monitoring stations within the local ring monitoring network are connected by communication links. Each terminal monitoring station is maintained using a local model with the assistance of other terminal monitoring stations within the local ring monitoring network. An edge server layer is deployed. Each edge server is responsible for managing all terminal monitoring station devices within its local ring monitoring network in the terminal monitoring station layer. The edge server and the local terminal monitoring station devices it manages jointly maintain a local model. The blockchain consensus layer consists of several decentralized blockchain nodes, each corresponding to an edge server in the edge server layer. All edge servers jointly maintain a global model through the blockchain nodes.

[0081] In the terminal monitoring station layer, terminal monitoring station devices employ a ring-topology-based split-fed learning approach for local model training. Each terminal monitoring station device is only responsible for calculating and updating a portion of its local model's parameters; other parameters are calculated by other terminal monitoring station devices within its local ring monitoring network. Each terminal monitoring station device uses local data to calculate and update a portion of its local model's parameters and sequentially transmits the generated forward activation values ​​or backward gradients along the communication link to other terminal monitoring station devices within its local ring monitoring network, until all terminal monitoring station devices within the local ring monitoring network have been traversed. The training task of each terminal monitoring station device is jointly determined by its position within the local ring monitoring network and the number of training layers. The edge server layer manages the collaborative training and aggregation of local models by the terminal monitoring station devices. Each edge server manages a local ring monitoring network and predefines a "propagation step size" for all terminal monitoring station devices within that local ring monitoring network. This propagation step size determines the number of training layers each terminal monitoring station device uses when collaboratively training its local model. The blockchain consensus layer is responsible for global model evaluation, consensus filtering, and weighted aggregation across edge servers. It records global model updates through a distributed ledger, ensuring the generalization and security of the global model.

[0082] (3) Perform local training based on ring topology. Each terminal monitoring station device is only responsible for calculating and updating some parameters of the local model and relaying the intermediate calculation results in the communication link within the ring.

[0083] To adapt to the heterogeneity of terminal monitoring station equipment in terms of computing power and communication bandwidth, and to alleviate the problem of "lagging" equipment slowing down the overall training process in traditional federated learning, this invention introduces a split federated learning method based on ring topology within the local ring monitoring network at the monitoring station level.

[0084] In a multi-layered split-fed federated learning architecture based on ring topology, terminal monitoring station devices construct several local ring monitoring networks according to their geographical location and spectrum environment characteristics. The training of a local model for a given terminal monitoring station device is not independent, but rather involves the joint participation of all terminal monitoring station devices within its local ring monitoring network. Each terminal monitoring station device is only responsible for training different layers of its local model. During local model training, the specific local model layer that a terminal monitoring station device should be responsible for is determined by its position within its local ring monitoring network and the predefined "propagation step size" of the edge server. Specifically, the position of the terminal monitoring station device within the local ring monitoring network defines its starting position during collaborative local model training, and the propagation step size quantifies the number of local model layers it needs to continuously train, thus achieving distributed distribution of computational load among the terminal devices.

[0085] When starting local training, each terminal monitoring station in the local ring monitoring network first needs to obtain complete initial global model parameters from the blockchain node via the edge server. Within each training cycle, each terminal monitoring station can serve as the starting point for training its local model, initiating a collaborative training session within its local ring monitoring network. Based on the hierarchical division of the local model, the complete process of a single collaborative training session (including forward and backward propagation) is dynamically divided into several parts, which are sequentially completed by each terminal monitoring station within the local ring monitoring network.

[0086] During training, each terminal monitoring station uses local data to train and update the parameters of its assigned local model. Simultaneously, it transmits intermediate results such as forward activation values ​​and backward gradients sequentially along the communication link to other terminal monitoring stations within the local ring network, until all terminal monitoring stations in the local ring network jointly complete one iterative optimization of the local model. This patent uses a local empirical risk function to measure the loss of each terminal monitoring station's local model on the training set. This function consists of a local loss function term and a regularization term. The local loss function term is used to fit local data features, while the regularization term constrains the shift of local model parameters, preventing the local model from deviating excessively from the global model.

[0087] The forward propagation is as follows: When a terminal monitoring station in the local ring monitoring network initiates collaborative training using its local data as the starting point, the terminal monitoring station completes the training of a portion of the local model and generates forward activation values ​​based on the received global model and training task. These forward activation values ​​are then sequentially transmitted along the communication link to other terminal monitoring station devices within the local ring monitoring network. The other terminal monitoring station devices train a portion of the local model based on the received global model, training task, and forward activation values ​​and generate their corresponding forward activation values, thus completing the forward propagation of the local model training.

[0088] The backpropagation is as follows: After all terminal monitoring station devices in the local ring monitoring network have completed forward propagation, they calculate the loss function and transmit the generated gradients back to other terminal monitoring station devices in sequence along the communication link. After receiving the backpropagated gradients, the other terminal monitoring station devices update the local model parameters they are responsible for, thus completing the backpropagation of the local model training.

[0089] Local empirical risk function: the first in the local ring monitoring network The local model training objective for each terminal monitoring station device is to minimize its local empirical risk function. :

[0090]

[0091] in, For the first The local dataset held by each terminal monitoring station For a single training sample and its corresponding label, where The input is the original I / Q signal tensor. This is the modulation scheme label corresponding to the signal. For the trained neural network model, These are the local model parameters that the terminal monitoring station equipment is currently training or updating. Let cross-entropy be the loss function. These are the global model parameters distributed to the blockchain nodes via the edge server. Regularization coefficients are used to constrain the deviation of the local model during this batch of training.

[0092] By minimizing the local empirical risk function, the trained local model can better learn the signal characteristics of the local data while effectively suppressing model drift caused by non-independent and identically distributed data, thereby improving the accuracy of classification and recognition of modulated signals.

[0093] At the start of each round of local training, the edge server dynamically adjusts its corresponding propagation step size by monitoring the hardware status of each terminal monitoring station device to achieve load balancing within the local ring monitoring network. This load balancing ensures that devices with weaker computing power only undertake a small number of local model layer calculations commensurate with their processing capabilities, thereby preventing the overall training progress from being limited by locally lagging nodes. This ring topology-based training mode effectively solves the heterogeneity problem of terminal monitoring station devices, providing a more efficient local training foundation for the entire architecture.

[0094] (4) Introduce a dynamic regularization mechanism with adaptive signal-to-noise ratio in local training. By sensing the signal-to-noise ratio of the data in real time, the regularization constraint is dynamically adjusted to suppress the pollution of the global model by noise interference.

[0095] This invention introduces a signal-to-noise ratio (SNR) adaptive dynamic regularization mechanism during the local training process of terminal monitoring station equipment. This mechanism aims to address the SNR heterogeneity problem caused by differences in channel conditions among different monitoring nodes and suppress the interference of noise gradients generated by low-quality data on the global model, thereby effectively optimizing the traditional federated learning algorithm.

[0096] Traditional federated proximal optimization algorithms use a fixed regularization coefficient μ to constrain local model parameters. With global model This deviation mitigates model drift caused by non-independent and identically distributed data. However, in automatic modulation classification and recognition scenarios, the reliability of different training batches of data varies significantly due to their different average signal-to-noise ratios (SNR). Low SNR samples contain a large amount of noise, and their calculated gradients have high variance and uncertainty, resulting in low reliability. Direct fitting without constraints can easily lead to the local model overfitting to noise or deviating from the optimization direction. Conversely, high SNR samples have clear features, and their gradients more realistically reflect the essential characteristics of the signal, resulting in high reliability. The local model should be allowed to learn more fully.

[0097] The adaptive signal-to-noise ratio (SNR) dynamic regularization mechanism described herein is an optimization strategy designed to address the significant differences in SNR between local data from different terminal monitoring station devices in a wireless communication environment. This mechanism, by introducing SNR sensing capabilities, enables dynamic intervention in the local empirical risk function.

[0098] The local model extracts the current training batch in real time during each training iteration. Average signal-to-noise ratio :

[0099]

[0100] in, The size of the current training batch. Indicates training batch The Middle The signal-to-noise ratio of each signal sample.

[0101] The regularization coefficient for this batch was then calculated. :

[0102]

[0103] in, The basic regularization constant; This is a limiting function, designed to avoid overly or underly constraining the local model. The attenuation factor is used to control the sensitivity to signal-to-noise ratio; and The lower and upper bounds of the regularization strength are defined respectively. When the average signal-to-noise ratio... When it is lower, the corresponding exponential term Larger, leading to The signal-to-noise ratio is also relatively high, making the local model more dependent on the global model, thus suppressing overfitting of the local model on noisy data. Conversely, when the average signal-to-noise ratio is lower... At higher levels, The local model is also relatively small, allowing it to extract fine features of the signal with greater exploratory freedom.

[0104] (5) After local training is completed, the terminal monitoring station device will upload the complete local model parameters, and the corresponding edge server will perform weighted aggregation to generate a local model;

[0105] Once all terminal monitoring stations within the local ring monitoring network have traversed the local dataset, it signifies that the local ring monitoring network has completed one round of local training. Each terminal monitoring station will then submit its generated local model parameters. The data is uploaded to the edge server that manages the local ring monitoring network via a communication link;

[0106] Edge server After collecting complete local models submitted by all terminal monitoring stations within the local ring monitoring network, a weighted aggregation operation is performed to merge and generate a local model of the local ring monitoring network. Its mathematical expression is:

[0107]

[0108] in, This is the set of terminal monitoring stations managed by the edge server. For the first The aggregation weight assigned to each terminal monitoring station device is determined based on the proportion of local data samples of that terminal monitoring station device, and satisfies the following conditions: Once the aggregation is complete, this local model will participate in the quality proof consensus of the subsequent blockchain consensus layer.

[0109] (6) The blockchain consensus layer executes a node quality proof consensus based on diversity, switches between quality proof score and diversity score according to the current learning rate, evaluates, filters and weights each local model, and then aggregates to generate a global model.

[0110] This step is executed at the blockchain consensus layer. The core objective of diversity-based node quality proof consensus technology is to generate a global model with high accuracy and strong generalization ability. Blockchain nodes, through diversity-based quality proof consensus technology, select local models with greater generalization ability, thereby generating a global model that can effectively cope with uneven data distribution and noise interference bias.

[0111] The diversity-based node quality proof consensus technology relies on a public verification set pre-deployed on each blockchain node. This public verification set, serving as a unified performance metric, is divided into multiple non-overlapping verification subsets and distributed to each blockchain node. Each blockchain node calculates a local score for its local model on its assigned verification subset, and then exchanges and aggregates these local scores through a consensus algorithm to obtain the aggregate weight of each local model.

[0112] After blockchain nodes collect all local models, these models may exhibit significant biases due to the non-independent and identically distributed nature of the data. To improve the generalization ability of the global model, the quality of each generated local model needs to be evaluated. This invention employs a node quality proof consensus mechanism based on diversity. This mechanism evaluates local models from two dimensions: quality and diversity, defining corresponding quality proof scores and diversity scores. Each blockchain node selects between the quality proof score and the diversity score, using its current learning rate (LR) as the indicator.

[0113] Quality proof score: Let the switching threshold be... ,when At this point, the blockchain node determines that the loss function of the global model is in a period of rapid decline, and the blockchain node collects the local models submitted by each edge server. Then, based on the loss function values ​​on the common validation set... Calculate the quality proof score :

[0114]

[0115] Quality proof score This enables blockchain nodes to achieve initial and rapid convergence of the global model by assigning higher aggregation weights to low-loss local models.

[0116] Diversity score: when the conditions are met At this point, each blockchain node determines that the global model training has entered a plateau period and selects a local model diversity score. As an evaluation metric, the diversity score in the stable fine-tuning stage aims to break the local optima caused by non-independent and identically distributed data, and improve the recognition accuracy of modulation schemes with a small sample size. The expression is as follows:

[0117]

[0118] in, The arithmetic mean accuracy of the local model across all modulation categories. The standard deviation of the accuracy for each category is given. For the parameters of the local model, This is the preset dispersion penalty coefficient.

[0119] Diversity score This includes both rewards for average accuracy and penalties for class imbalance. Diversity score. This enables each blockchain node to prioritize and aggregate local models that perform evenly across all modulation categories, thereby effectively alleviating the model convergence problem caused by non-independent and identically distributed data in the global model.

[0120] Each blockchain node assigns a corresponding aggregation weight to each local model based on the quality proof score or diversity score, and generates a new round of global model, thereby completing the global model update in a decentralized environment.

[0121] (7) Perform global model update and recognition result output. Each blockchain node performs performance verification and block notarization on the generated global model, and securely sends the parameters to the underlying network to start the next round of iteration or output classification judgment.

[0122] As the final step in a single training cycle of the aforementioned method, this step is responsible for receiving the consensus results and performing global model generation, performance verification, secure storage, and parameter distribution. Through the implementation of this step, the entire architecture can initiate the next iteration cycle or complete the delivery of the final usable global model; at the same time, this step ensures the logical closed loop of the overall training process and achieves reliable synchronization of global model parameters among heterogeneous nodes.

[0123] For the newly generated global model, each blockchain node first performs performance verification and stores the results on the blockchain. Each blockchain node then performs a rapid evaluation of the newly generated model on a public validation set. The resulting key performance metrics (such as accuracy and loss) are combined with the hash values ​​of the local model parameters and recorded as metadata for this aggregation in the new block. The public validation set, logically serving as a unified performance metric, is physically divided into multiple non-overlapping validation subsets and distributed to each blockchain node. Each node exchanges evaluation results based on its respective subset, and finally, a globally consistent aggregation weight is established through a consensus algorithm. The immutability of the blockchain ensures the auditability of the entire decision-making and execution process, from weight calculation to global model generation.

[0124] Each blockchain node performs a secure distribution operation of global model parameters to achieve a logical closed loop in the training process. After the complete block of the latest aggregation result is confirmed by each blockchain node, the blockchain consensus layer securely distributes the latest global model parameters to each edge server. Subsequently, each edge server synchronizes the received parameters to the underlying terminal monitoring station equipment under its jurisdiction and uses it as the starting point for the next training cycle. If the performance of the obtained global model is verified to meet the preset requirements, the blockchain node uses the current global model as the final output, thereby completing the modulation recognition task.

[0125] The following experimental examples will further illustrate the effects of the present invention.

[0126] Experimental setup: The RadioML2018.01a dataset was used. This dataset contains 24 modulation types with a signal-to-noise ratio ranging from -20dB to +30dB. The raw I / Q signals without any processing were used as input in the experiment. To simulate a real heterogeneous environment, a Dirichlet distribution was employed. The data is divided into non-independent and identically distributed partitions.

[0127] This method controls the skewness of the data distribution by adjusting the concentration parameter α of the Dirichlet distribution: the smaller the α value, the greater the difference in tag distribution among different terminal monitoring stations, meaning that some terminal monitoring stations mainly collect one type of data, while other types of data are relatively scarce; the larger the α value, the more uniform the data distribution. By flexibly adjusting the α parameter, different degrees of non-independent and identically distributed data partitioning can be generated, thereby more realistically simulating the data heterogeneity of monitoring nodes in a radio environment and providing a more realistic experimental basis for subsequent model training and evaluation.

[0128] The architecture described in the experiment includes A local ring-shaped monitoring network is established, with each edge server managing three terminal monitoring station devices. Global communication rounds are set to 600 rounds, and an early stopping strategy is enabled: training stops when the loss value no longer decreases after a certain number of consecutive rounds. The initial learning rate is set to... In addition, a dynamic learning rate decay strategy is employed: if the loss does not decrease for several consecutive rounds, the learning rate is automatically decayed.

[0129] In real-world scenarios, terminal monitoring station equipment is typically deployed in harsh electromagnetic environments, and sensing terminals may malfunction due to hardware aging or battery depletion. To simulate this instability, starting from the 50th communication round, high-intensity Gaussian noise interference was introduced into the model update of one of the local ring monitoring networks to simulate severe data distortion caused by channel fading or sensor failure. Based on this, to verify the effectiveness and robustness of the proposed scheme, this experimental example uses three comparative schemes for evaluation:

[0130] Option 1: This option adopts a standard federated learning star topology architecture, with parallel communication between the terminal monitoring station equipment and the server. However, due to its lack of awareness of physical layer signal quality and robust design against abnormal updates, it performs extremely poorly when faced with malicious noise injection or sensor failure, and is prone to catastrophic performance drops.

[0131] Option 2 introduces a blockchain-based quality proof consensus mechanism at the top level. Although it still uses a star-shaped parallel architecture in terms of topology, compared to Option 1, Option 2 significantly enhances security. It can identify and isolate abnormal updates generated by faulty nodes through validation set loss evaluation, thereby maintaining a stable convergence trajectory under noise interference. However, Option 2 has a significant efficiency bottleneck. Because its communication mode requires the central server to wait for all heterogeneous terminal monitoring station devices to complete transmission, the overall model training speed is limited by the slowest "laggard," restricting the training efficiency under large-scale deployment.

[0132] Option 3: This invention incorporates the ring topology and split learning technology at the monitoring station layer. Option 3 achieves asynchronous parameter transmission, effectively alleviating communication congestion and latency issues. This option combines a deep collaborative signal-to-noise ratio adaptive regularization mechanism with a diversity-based blockchain node quality proof consensus technology, significantly improving operational efficiency while maintaining high reliability. This alleviates the conflict between real-time performance and reliability in radio monitoring scenarios, providing a feasible solution for collaborative training of large-scale heterogeneous terminals.

[0133] Simulated clock time setting: Since the experiment was conducted serially in a single-machine environment, to fairly evaluate the communication efficiency in real heterogeneous networks, "simulated clock time" was defined as a performance metric. Total model training latency. Latency calculated locally and transmission delay composition.

[0134] Computation latency: Based on a model computation load of 0.5 GFLOPs, and according to the set heterogeneous computing power. GFLOPS calculates the GFLOPS of each node. :

[0135]

[0136] in, The computational cost of the representative model is expressed in GFLOPs. The heterogeneous computing power (in GFLOPS) represents the heterogeneous computing power of each terminal monitoring station device.

[0137] Communication latency: based on model size and set link bandwidth (Average 2Mbps, including 10% of lagging nodes) Calculate transmission time :

[0138]

[0139] in, This represents the number of model parameters that need to be transmitted during a single communication.

[0140] The following analysis, based on the attached figures, experimental content, and results, is presented:

[0141] Figure 2 A schematic diagram illustrating the workflow of the method described herein demonstrates the operational mechanism of the method of this invention;

[0142] Figure 3 The classification accuracy curves of the global model under different signal-to-noise ratios are given when the parameter α=0.5;

[0143] Figure 4 The confusion matrix of the model under different signal-to-noise ratios is shown under the same parameter settings;

[0144] Figure 5 The accuracy curves of the model under different α values ​​and different signal-to-noise ratios were compared. Experimental results show that when the α parameter is appropriately selected, the curves exhibit a significant convergence trend. This indicates that as long as the data of each node has a certain degree of inter-class diversity, the designed SNR adaptive dynamic regularization mechanism can effectively suppress the offset of local model updates, making the global model performance approach the ideal state. Even in extreme heterogeneous scenarios, the performance of the architecture only shows a slight decrease; under high SNR (SNR=10dB) conditions, the average accuracy of the global model remains above 93%, verifying the stability and usability of the method of this invention in extreme communication environments.

[0145] Figure 6 The accuracy of three schemes on the test set was compared with the number of communication rounds. Experimental results show that the performance of the traditional scheme 1 immediately drops after being disturbed, with the accuracy rapidly decreasing to 4.46%, close to the level of random guessing. This indicates that the traditional federated aggregation mechanism cannot perceive link quality and is unable to filter out contaminated parameter updates, leading to the destruction of the global model. In contrast, scheme 2 and the scheme 3 proposed in this invention both maintain stable convergence trajectories and do not fluctuate significantly due to environmental noise. This result verifies the effectiveness of the diversity-based quality proof consensus mechanism designed in the blockchain consensus layer. By combining the dual evaluation of loss function and diversity, the architecture can identify and isolate abnormal updates from faulty local ring monitoring networks, thereby maintaining the integrity of the global model in unreliable communication environments.

[0146] Figure 7 The test accuracy of the three schemes was compared with the analog clock time. Compared with Scheme 1 and Scheme 2, Scheme 3 proposed in this invention shows superior communication efficiency. Thanks to the ring topology, this scheme effectively alleviates the "fallback" effect caused by the heterogeneity of terminal monitoring station equipment. Experimental results show that Scheme 3 achieves improved training speed while maintaining a high accuracy comparable to Scheme 2.

Claims

1. A multi-layer split federated learning automatic modulation classification and recognition method based on ring topology, characterized in that: Includes the following steps: (1) Perform signal model construction and data preprocessing, extract the received continuous radio frequency signal into in-phase and quadrature components, and construct a two-dimensional tensor as the input of the deep learning model; (2) Deploy a multi-layered split federated learning architecture and build a local ring monitoring network composed of terminal monitoring station devices under each edge server; (3) Perform local training based on ring topology. Each terminal monitoring station device is only responsible for calculating and updating some parameters of the local model and relaying the intermediate calculation results in the communication link within the ring. (4) Introduce a dynamic regularization mechanism with adaptive signal-to-noise ratio in local training. By sensing the signal-to-noise ratio of the data in real time, the regularization constraint is dynamically adjusted to suppress the pollution of the global model by noise interference. (5) After local training is completed, the terminal monitoring station device will upload the complete local model parameters, and the corresponding edge server will perform weighted aggregation to generate a local model; (6) The blockchain consensus layer executes a node quality proof consensus based on diversity, switches between quality proof score and diversity score according to the current learning rate, evaluates, filters and weights each local model, and then aggregates to generate a global model. (7) Perform global model update and recognition result output. Each blockchain node performs performance verification and block storage on the generated global model, and securely sends the parameters to the underlying network to start the next round of iteration or output classification judgment.

2. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The signal model construction and data preprocessing process in step (1) is as follows: The receiver converts continuous radio frequency signals into discrete-time complex baseband signals. , is represented as: ； in, This is the sampling sequence number. The original modulated signal from the transmitting end. This is the multipath fading impulse response function. For additive white Gaussian noise, for the first The first signal sample There are several sampling points, and the following relationships exist: ； in, The imaginary unit, For the first A discrete-time complex baseband signal sample This is the in-phase component of the signal sample. These are the orthogonal components of the signal sample; To enable a deep learning model to learn the temporal features of a signal, a segment of length is extracted from the continuous sequence. A window of observation, which will... The in-phase component sequence corresponding to each sampling point and orthogonal component sequences The top and bottom edges are concatenated to form a matrix, which is then transposed to ultimately construct a matrix with dimension [dimensional value missing]. real two-dimensional tensors : 。 3. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The process for deploying the multi-layer split federated learning architecture based on ring topology in step (2) is as follows: The architecture consists of a terminal monitoring station layer, an edge server layer, and a blockchain consensus layer from bottom to top, wherein the terminal monitoring station layer is deployed... Each terminal monitoring station is equipped with a local ring monitoring network, which is constructed based on geographical location and actual spectrum environment characteristics. The terminal monitoring stations within the local ring monitoring network are connected by communication links, and each terminal monitoring station is maintained using a local model with the assistance of other terminal monitoring stations within the local ring monitoring network. The edge server layer is deployed... Each edge server is responsible for managing all terminal monitoring station devices within its local ring monitoring network in the terminal monitoring station layer. The edge server and the local terminal monitoring station devices it manages jointly maintain a local model. The blockchain consensus layer consists of several decentralized blockchain nodes, each blockchain node corresponding to an edge server in the edge server layer. All edge servers jointly maintain a global model through the blockchain nodes. In this system, the terminal monitoring station devices in the terminal monitoring station layer employ a ring-topology-based split federated learning approach for local model training. Each terminal monitoring station device is only responsible for calculating and updating a portion of the parameters of its local model. Other parameters are calculated by other terminal monitoring station devices within its local ring monitoring network. Each terminal monitoring station device uses local data to calculate and update a portion of the parameters of its local model and sequentially transmits the generated forward activation values ​​or backward gradients along the communication link to other terminal monitoring station devices within the local ring monitoring network until all terminal monitoring station devices within the local ring monitoring network have been traversed. The training task of each terminal monitoring station device is jointly determined by its position within the local ring monitoring network and the number of training layers. The edge server layer is responsible for managing the collaborative training and aggregation of local models by the terminal monitoring station devices. Each edge server manages a local ring monitoring network and predefines a "propagation step size" for all terminal monitoring station devices within that local ring monitoring network. This propagation step size determines the number of training layers for each terminal monitoring station device when collaboratively training its local model. The blockchain consensus layer is responsible for global model evaluation, consensus filtering, and weighted aggregation across edge servers. It records global model updates through a distributed ledger to ensure the generalization and security of the global model.

4. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The local training process for the split federated learning based on ring topology in step (3) is as follows: When each terminal monitoring station device in the local ring monitoring network starts local training, it needs to obtain complete initial global model parameters from the blockchain node through the edge server. In each training cycle, each terminal monitoring station device can serve as the starting point for training the local model and initiate a collaborative training in its local ring monitoring network. Based on the division of the local model hierarchy, the complete process of a single collaborative training is dynamically divided into several parts, including forward propagation and backward propagation, which are completed sequentially by each terminal monitoring station device in the local ring monitoring network. During training, each terminal monitoring station uses local data to train and update the parameters of its assigned local model. Simultaneously, it transmits the forward activation values ​​and backward gradients of the intermediate results sequentially along the communication link to other terminal monitoring stations within the local ring monitoring network until all terminal monitoring stations within the local ring monitoring network jointly complete one iterative optimization of the local model. A local empirical risk function is used to measure the loss of each terminal monitoring station's local model on the training set. This function consists of a local loss function term and a regularization term. The local loss function term is used to fit the local data features, while the regularization term is used to constrain the deviation of the local model parameters, preventing the local model from deviating excessively from the global model.

5. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 4, characterized in that, The forward propagation is as follows: When a terminal monitoring station device in the local ring monitoring network initiates collaborative training using its local data as the training starting point, the terminal monitoring station device completes the training of part of the local model and generates forward activation values ​​based on the received global model and training task. The forward activation values ​​are then transmitted sequentially to other terminal monitoring station devices in the local ring monitoring network along the communication link. The other terminal monitoring station devices train part of the local model based on the received global model, training task and forward activation values ​​and generate their corresponding forward activation values, thus completing the forward propagation of the local model training. The backpropagation is as follows: After all terminal monitoring station devices in the local ring monitoring network have completed forward propagation, they calculate the loss function and transmit the generated gradients back to other terminal monitoring station devices in sequence along the communication link. After receiving the backpropagated gradients, the other terminal monitoring station devices update the local model parameters they are responsible for, thus completing the backpropagation of the local model training.

6. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 4, characterized in that, The local empirical risk function: the first in the local ring monitoring network The local model training objective for each terminal monitoring station device is to minimize its local empirical risk function. : ； in, For the first The local dataset held by each terminal monitoring station For a single training sample and its corresponding label, where The input is the original I / Q signal tensor. This is the modulation scheme label corresponding to the signal. For the trained neural network model, These are the local model parameters that the terminal monitoring station equipment is currently training or updating. Let cross-entropy be the loss function. These are the global model parameters distributed to the blockchain nodes via the edge server. Regularization coefficients are used to constrain the deviation of the local model during this batch of training.

7. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The dynamic regularization of signal-to-noise ratio adaptation in step (4) is expressed as follows: The local model extracts the current training batch in real time during each training iteration. Average signal-to-noise ratio : ； in, The size of the current training batch. Indicates training batch The Middle Signal-to-noise ratio of each signal sample; The regularization coefficient for this batch was then calculated. : ； in, The basic regularization constant; For the amplitude limiting function, The attenuation factor is used to control the sensitivity to signal-to-noise ratio; and The lower and upper bounds of the regularization strength are defined respectively, when the average signal-to-noise ratio When it is lower, the corresponding exponential term Larger, leading to It is also relatively large; conversely, when the average signal-to-noise ratio is relatively small... At higher levels, The local model is also correspondingly smaller, allowing it to extract fine features of the signal with greater exploratory freedom.

8. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The process of generating local models through local model aggregation in step (5) is as follows: After each terminal monitoring station within the local ring monitoring network has traversed the local dataset, it signifies that the local ring monitoring network has completed one round of local training. Each terminal monitoring station will then transmit its generated local model parameters. The data is uploaded to the edge server that manages the local ring monitoring network via a communication link; Edge server After collecting complete local models submitted by all terminal monitoring stations within the local ring monitoring network, a weighted aggregation operation is performed to merge and generate a local model of the local ring monitoring network. Its mathematical expression is: ； in, This is the set of terminal monitoring stations managed by the edge server. For the first The aggregation weight assigned to each terminal monitoring station device is determined based on the proportion of local data samples of that terminal monitoring station device, and satisfies the following conditions: Once the aggregation is complete, this local model will participate in the quality proof consensus of the subsequent blockchain consensus layer.

9. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The consensus mechanism for proving the quality of blockchain nodes based on diversity in step (6) is as follows: The diversity-based node quality proof consensus relies on a public verification set pre-deployed on each blockchain node. The public verification set serves as a unified performance metric and is divided into multiple non-overlapping verification subsets, which are then distributed to each blockchain node. Each blockchain node calculates the local score of its local model on its respective verification subset and exchanges and summarizes these local scores through a consensus algorithm to obtain the aggregate weight of each local model. After the blockchain nodes collect all local models, each local model may have a large deviation due to the non-independent and identically distributed data. In order to improve the generalization of the global model, it is necessary to evaluate the quality of each generated local model. This is achieved by adopting a node quality proof consensus mechanism based on diversity. This mechanism evaluates from two dimensions: local model quality and diversity. It defines corresponding quality proof scores and diversity scores. Each blockchain node selects between the quality proof score and the diversity score based on the current learning rate (LR). Quality proof score: Let the switching threshold be... ,when At this point, the blockchain node determines that the loss function of the global model is in a period of rapid decline, and the blockchain node collects the local models submitted by each edge server. Then, based on the loss function values ​​on the common validation set... Calculate the quality proof score : ； Quality proof score This enables blockchain nodes to achieve initial and rapid convergence of the global model by assigning higher aggregation weights to low-loss local models. Diversity score: when the conditions are met At this point, each blockchain node determines that the global model training has entered a plateau period and selects a local model diversity score. As an evaluation metric, the diversity score during the stable fine-tuning phase for: ； in, The arithmetic mean accuracy of the local model across all modulation categories. The standard deviation of the accuracy for each category is given. For the parameters of the local model, This is a preset dispersion penalty coefficient; Each blockchain node assigns a corresponding aggregation weight to each local model based on the quality proof score or diversity score, and generates a new round of global model, thereby completing the global model update in a decentralized environment.

10. The automatic modulation classification and recognition method based on ring topology and multi-layer splitting federated learning according to claim 1, characterized in that, The process for global model update and recognition result output in step (7) is as follows: For the newly generated global model, each blockchain node first performs performance verification and stores the evidence on the blockchain. Each blockchain node will quickly evaluate the newly generated model on the public verification set. The resulting key performance indicators, along with the hash values ​​of the local model parameters, will be used as metadata for this aggregation and recorded in the new block. The public verification set is logically a unified performance metric. During physical deployment, it is divided into multiple non-overlapping verification subsets and distributed to each blockchain node. Each node exchanges the evaluation results based on its own subset, and finally establishes a globally consistent aggregation weight through a consensus algorithm. The immutability of the blockchain ensures the auditability of the entire decision-making and execution process from weight calculation to global model generation. Each blockchain node performs a secure distribution operation of global model parameters to achieve a logical closed loop in the training process. After the complete block of the latest aggregation result is confirmed by each blockchain node, the blockchain consensus layer securely distributes the latest global model parameters to each edge server. Subsequently, each edge server synchronizes the received parameters to the underlying terminal monitoring station equipment under its jurisdiction and uses it as the starting point of the next training cycle. If the performance of the obtained global model is verified to meet the preset requirements, the blockchain node uses the current global model as the final output to complete the modulation recognition task.

Images