An adaptive robust modulation identification method and system for a complex electromagnetic environment

Abstract

The application belongs to the technical field of modulation identification, and particularly relates to a self-adaptive robust modulation identification method and system for a complex electromagnetic environment, which performs data preprocessing on original IQ signals, reshapes the preprocessed signals into a model adaptation format, performs data enhancement on the preprocessed signals to simulate noise interference and transmission deviation of the complex electromagnetic environment, adopts an enhanced multi-scale CNN architecture to perform feature extraction and classification on the enhanced signals, monitors a CNN training state in real time through a PPO reinforcement learning module, performs adaptive learning rate scheduling based on verification accuracy feedback, and predicts a test set by using a trained CNN model. The application monitors the CNN training state in real time through the PPO reinforcement learning module, dynamically and adaptively adjusts learning rate, batch size and other hyperparameters, replaces a traditional artificial trial-and-error parameter adjustment mode, and significantly reduces the manpower and time cost of model training.

Description

Technical Field

[0001] This invention belongs to the field of modulation recognition technology, specifically relating to an adaptive robust modulation recognition method and system for complex electromagnetic environments. Background Technology

[0002] With the rapid development of wireless communication technology and the increasing scarcity of spectrum resources, intelligent identification and classification of radio signals have become key technologies in fields such as communication security, spectrum regulation, and military reconnaissance. Especially when dealing with unauthorized signals such as "black flight" drones and illegal communication devices, rapid and accurate signal identification capabilities are crucial. Automatic Modulation Recognition (AMR), as a cutting-edge intersection of signal processing and machine learning, aims to classify, identify, and track signal sources by analyzing the modulation patterns of received signals. However, the actual electromagnetic environment is complex and variable, and signals are often affected by factors such as multipath fading, noise interference, and frequency shift. Especially under low signal-to-noise ratio conditions, signal characteristics become blurred, and inter-class differences are small, posing a significant challenge to traditional identification methods.

[0003] Traditional AMR methods primarily rely on manual feature extraction and classical classifiers, such as those based on maximum likelihood estimation, higher-order statistics, and wavelet transform. However, these methods are limited in dynamic wireless environments and low signal-to-noise ratio (SNR) conditions, and suffer from high computational complexity and poor adaptability. With the successful application of deep learning in image and speech recognition, researchers have begun to incorporate it into AMR tasks, achieving automatic extraction and classification of signal features through end-to-end learning, significantly improving recognition performance.

[0004] In recent years, deep learning-based AMR methods have shown a trend towards multi-architecture, multi-modal, lightweight, and robustness-enhanced development. To simultaneously capture the spatiotemporal features of signals, many studies have adopted hybrid architectures of CNNs and RNNs (especially LSTMs). For example, Cheng R et al. proposed CVCNN-LSTM, using a complex network structure to preserve the phase integrity of I / Q signals; Tang H et al. designed a parallel feature extraction network, using CNNs and LSTMs respectively to extract spatial and temporal features, avoiding feature degradation problems. Transformers, with their powerful global modeling capabilities, are widely used in AMR tasks. Li G et al. proposed a robust AMR method based on Vision Transformers, improving stability under adversarial attacks through temporal correlation modeling. Kong W et al. designed a two-stage Transformer with mask training and fine-tuning to achieve efficient representation and recognition of multimodal signals. To address complex electromagnetic environments and domain differences, multimodal fusion and domain adaptation techniques have attracted attention. Yan S et al. proposed a Hybrid Modal Contrast Fusion (HMCF-AMR) method, combining image and sequence modalities for self-supervised pre-training and feature fusion. Zhang M et al. proposed a semi-supervised domain adaptive method (SSDA-AMR) to achieve cross-domain feature alignment through adversarial training. Lightweight network design has become a research hotspot due to the computational efficiency requirements in practical deployments. Zhu Z et al. proposed MSGNet, which uses depthwise separable convolutions and gated recurrent units to reduce the number of parameters and computational complexity. Wang S et al. designed a multi-source heterogeneous network (MSHNet) that integrates temporal, image, and graph representations, improving feature diversity while maintaining performance. Improving model robustness is particularly important in the face of adversarial attacks. Chen Z et al. proposed an adversarial multi-distillation mechanism (AMD) to transfer classification and defense knowledge through knowledge distillation. Li G et al. introduced noise-adaptive adversarial training on top of the Transformer, significantly improving recognition accuracy under real attacks.

[0005] Despite significant progress in AMR (Automatic Mapping), deep learning still faces challenges such as performance degradation under low signal-to-noise ratio and complex channel conditions, and insufficient generalization ability across scenarios and devices. To address these issues, reinforcement learning (RL) has shown great potential in optimizing the deep learning training process in recent years. RL learns optimal decision-making policies through interaction with the environment and is suitable for dynamic optimization problems. Proximal policy optimization (PPO), as a stable and efficient RL algorithm, has achieved success in areas such as robot control and game decision-making, but its application in signal recognition tasks is still in the exploratory stage. Summary of the Invention

[0006] To address the technical problems of existing automatic modulation identification methods, this invention provides an adaptive robust modulation identification method and system for complex electromagnetic environments.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An adaptive robust modulation identification method for complex electromagnetic environments includes the following steps: S1. Perform data preprocessing on the original IQ signal and reshape the preprocessed signal into a model-adaptive format; S2. Perform data enhancement on the preprocessed signal to simulate noise interference and transmission deviation in a complex electromagnetic environment; S3. An enhanced multi-scale CNN architecture is used to extract features and classify the enhanced signal. The convolution operation satisfies the following formula: Where: y is the output feature map of the convolutional layer. It is the ReLU activation function, used to introduce non-linearity and enhance feature representation; W is the convolution kernel weight matrix. It is a convolution operation, where x is the input signal and b is the bias term; S4. Monitor the CNN training state in real time using the PPO reinforcement learning module, construct a 10-dimensional state vector, and perform adaptive hyperparameter adjustment based on a multi-objective weighted reward function. The reward function satisfies the following equation: in: This is the final reward value; It is the difference between the current and previous rounds of validation accuracy, weighted... The target is 300, to encourage improved accuracy. It is the difference between the current accuracy and the historical best accuracy, rewarding breakthrough performance; stability is the stability of the accuracy in the last 5 rounds of validation, avoiding metric fluctuations; penalty is the overfitting penalty, which is triggered when the difference between the training accuracy and the validation accuracy is greater than 0.1, thereby suppressing overfitting. S5. Based on the verification accuracy feedback, perform adaptive learning rate scheduling. The learning rate update satisfies the following formula: in: This is the updated learning rate; This is the current learning rate; This is the attenuation coefficient, set to 0.5 to control the attenuation magnitude each time; It is the lower bound of the learning rate; S6. Input the trained CNN model into the test set for prediction and output the signal modulation type identification result.

[0008] In step S1, the preprocessed 2×128 signal is reshaped into a (2,128,1) format.

[0009] In step S2, the data augmentation specifically includes: superimposing Gaussian noise with a standard deviation of 0.005-0.02, shifting the time axis by ±3 units, scaling the amplitude by a random coefficient of 0.9-1.1, and masking 10% of the time point data with a random mask.

[0010] In step S3, the enhanced multi-scale CNN architecture uses multi-scale convolutional kernels of (1,8) and (1,4) sizes. After max pooling, global average pooling, and two fully connected layers, it outputs probability values ​​corresponding to 11 signal types.

[0011] In step S4, the PPO model includes a policy network and a value network with the same structure; the policy network outputs the probability distribution of five parameter tuning actions: increasing the learning rate, decreasing the learning rate, maintaining the learning rate, adjusting the batch size, and maintaining the batch size; the 10-dimensional state vector includes: training progress, current validation accuracy, current validation loss, recent average accuracy, accuracy fluctuation, recent average loss, loss fluctuation, overfitting degree, historical best accuracy, and the number of steps away from the best epoch.

[0012] In step S4, the overfitting penalty term is triggered when the difference between the training accuracy and the validation accuracy is greater than 0.1.

[0013] In step S5, the attenuation coefficient Set the lower bound of the learning rate to 0.5. Set to 1e-6; when the accuracy does not exceed the historical best value in 15 consecutive rounds of validation, perform learning rate decay.

[0014] In step S6, the argmax function is used to select the category index with the highest probability to determine the signal modulation type.

[0015] An adaptive robust modulation recognition system for complex electromagnetic environments includes a data preprocessing module, a data augmentation module, a CNN model training module, a PPO model enhancement module, and a prediction module. The signal output terminal of the data preprocessing module is connected to the signal input terminal of the data augmentation module, and the signal output terminal of the data augmentation module is connected to the signal input terminal of the CNN model training module. The PPO model enhancement module is bidirectionally connected to the CNN model training module, and the adaptive learning rate scheduling unit is connected to both the CNN model training module and the PPO model enhancement module. The output of the CNN model training module is connected to the input of the prediction module; The data preprocessing module is used to receive the raw IQ signal and convert it into standardized input data; The data augmentation module is used to apply Gaussian noise, time offset, amplitude scaling, and random masking to the preprocessed signal. The prediction module is used to output the signal modulation type identification result based on the trained CNN model.

[0016] The CNN model training module adopts an enhanced multi-scale CNN architecture, which includes three convolutional blocks connected sequentially to a max pooling layer, a global average pooling layer, a 512-neuron fully connected layer, and a 256-neuron fully connected layer. The policy network and value network of the PPO model enhancement module are both constructed from fully connected layers, batch normalization layers, and dropout layers.

[0017] Compared with the prior art, the beneficial effects of this invention are: 1. This invention uses the PPO reinforcement learning module to monitor the CNN training status in real time and dynamically and adaptively adjust hyperparameters such as learning rate and batch size, replacing the traditional manual trial-and-error parameter tuning method, which significantly reduces the manpower and time cost of model training; at the same time, combined with a multi-objective weighted reward mechanism, it effectively suppresses overfitting and avoids the model from getting stuck in local optima, making the training process more efficient and convergence more stable.

[0018] 2. This invention adopts an enhanced CNN architecture with multi-scale convolutional kernels, which can simultaneously capture the wide range of signal trends and local detailed features. Combined with global average pooling and channel weight optimization of fully connected layers, it significantly improves the feature discrimination ability of blurred signals under low signal-to-noise ratio conditions. It can still maintain high recognition accuracy in complex electromagnetic environments such as multipath fading, noise interference, and frequency shift, solving the problem of performance degradation of traditional methods under harsh channels.

[0019] 3. This invention uses multi-dimensional data enhancement, including Gaussian noise, time offset, amplitude scaling, and random masking, to simulate various interferences and deviations in real electromagnetic transmission, thus enriching sample diversity. Combined with adaptive learning rate scheduling based on verification accuracy, it further enhances the model's adaptability to signals from different scenarios and devices, enabling the model to have stronger generalization in cross-scenario and cross-domain signal recognition. Attached Figure Description

[0020] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0021] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0022] Figure 1 This is a diagram of the network model architecture of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. These descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the claims of the present invention. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] I. Network Model Architecture The adaptive robust modulation identification method and system architecture for complex electromagnetic environments proposed in this embodiment are as follows: Figure 1 As shown, the system comprises five core modules: data preprocessing, data augmentation, CNN model training, PPO model enhancement, and prediction. The system takes the raw IQ signal as input, preprocesses and enhances it before feeding it into the CNN for feature extraction and classification. The PPO agent monitors the training status in real time and outputs hyperparameter adjustment actions, achieving end-to-end adaptive optimization.

[0026] The core objective of this architecture is to automatically identify the modulation type of radio signals (such as common BPSK, AM-DSB, etc.), while solving the problems of "large workload for parameter optimization and low recognition accuracy under low signal-to-noise ratio" faced by traditional methods. It has the following three key functions.

[0027] First, regarding automatic signal classification, an enhanced CNN network is proposed. In recent years, significant progress has been made in CNN-based automatic modulation classification research. O'Shea et al. applied CNN to radio signals, demonstrating the effectiveness of deep temporal learning. Hermawan et al. proposed IC-AMCNet, improving accuracy and efficiency through network structure optimization to meet B5G requirements. Huynh-The et al. designed MCNet, utilizing asymmetric convolutional kernels to learn spatiotemporal correlations to optimize model accuracy. Therefore, this embodiment designs an enhanced CNN network architecture to identify radio signal types.

[0028] Secondly, regarding automatic parameter tuning during training, a PPO-based algorithm reinforcement module is proposed. Traditional model training requires manual and repeated adjustments to parameters such as learning rate and batch size. This embodiment introduces the PPO reinforcement learning algorithm, allowing the system to automatically optimize parameters based on training progress, significantly improving the model's accuracy in recognizing unknown signals.

[0029] Finally, the raw radio signals obtained from the public dataset are processed into standardized data that the model can directly learn from. This addresses the issues of mismatched raw data format with model input and uneven data distribution leading to training bias, facilitating subsequent model learning. Since signals in real-world environments are susceptible to noise interference, this embodiment actively adds Gaussian noise to the raw data to simulate random interference in signal transmission, such as electromagnetic noise. A "time offset" is added to simulate the small time delay during signal transmission, such as the time difference between the signal traveling from an unauthorized flying device to the receiver. The enhanced signal allows the model to better adapt to new noise environments during subsequent learning.

[0030] II. Key Module Design 2.1 Enhanced CNN Design Radio signals exhibit blurred features at low signal-to-noise ratios (SNR), necessitating the design of targeted CNN architectures to enhance feature discrimination capabilities. This embodiment presents an enhanced CNN architecture employing multi-scale convolutions and fully connected layers to strengthen the capture of effective features.

[0031] The CNN model training module serves as the "core module for signal recognition" of the system. Signal format conversion first reshapes the preprocessed "2×128" signal into a "(2,128,1)" format. Three "convolutional blocks" are used to extract the "key features" of the signal layer by layer. Multi-scale convolutional kernels of sizes (1,8) and (1,4) are used, corresponding to wide-range trends and local detail features, respectively. A convolutional kernel of size (1,8) represents a kernel with a height of 1 and a width of 8, equivalent to covering 8 time steps in the time series. A convolutional kernel of size (1,4) represents a kernel with a height of 1 and a width of 4, equivalent to covering 4 time steps in the time series. Convolutional operations capture signal features of different granularities; the core convolution is shown in formula (1). y is the output feature map of the convolutional layer. It is the ReLU activation function, used to introduce non-linearity and enhance feature representation. W is the convolution kernel weight matrix. This is a convolution operation, where x is the input signal and b is the bias term. Max pooling "compresses and simplifies" the extracted features, retaining the most important information and reducing computational cost.

[0032] (1) Employing a fully connected mechanism enhances the weights of effective feature channels. First, global average pooling is performed on the convolutional output to compress the spatial dimension while preserving channel information. Then, two fully connected layers learn the channel importance weights. Flattening the features transforms the "multi-dimensional features" output by the convolutional block into a "one-dimensional vector." The fully connected layers compute further features through two layers: the first with 512 neurons and the second with 256 neurons, finally outputting "11 probability values" corresponding to the probabilities of 11 signal types. The highest probability value represents the signal type identified by the model.

[0033] 2.2 PPO-driven adaptive training mechanism The PPO model consists of a policy network (Actor) and a value network (Critic). These two networks have identical structures. The policy network (Actor) outputs the probability distribution of five parameter tuning actions based on the state vector: increasing the learning rate, decreasing the learning rate, maintaining the learning rate, adjusting the batch size, and maintaining the batch size. The value network (Critic) evaluates the value of the current state, representing the expected accuracy achievable through subsequent training, assisting the policy network in making more reasonable decisions. Both networks are constructed using fully connected layers, batch normalization, and dropout to ensure stable decisions and prevent overfitting. After parameter tuning, if the CNN's accuracy improves and overfitting is reduced, the reward mechanism gives the PPO a positive reward; if accuracy decreases and overfitting worsens, a negative penalty is incurred. The PPO adjusts its policy based on the "reward," prioritizing parameter tuning actions that yield high rewards when encountering similar states in the future, gradually learning the most suitable parameter tuning method for the current training.

[0034] The PPO model enhancement module monitors the CNN training status in real time and dynamically adjusts training parameters, solving the problem of low efficiency in traditional manual parameter tuning and making CNN training more efficient and stable. The state vector consists of 10 key metrics during CNN training. These 10 dimensions include training progress, current validation accuracy, current validation loss, recent average accuracy, accuracy fluctuation, recent average loss, loss fluctuation, overfitting, historical best accuracy, and steps to the best epoch. Recent average accuracy reflects the accuracy trend, accuracy fluctuation reflects the algorithm's stability, and overfitting is measured by calculating the difference between training accuracy and validation accuracy. These metrics allow PPO to clearly understand the current progress and effectiveness of CNN training.

[0035] The reward function is designed with a multi-objective weighting mechanism to balance accuracy improvement, training stability and overfit suppression, as shown in Equation (2). This is the final reward value; It is the difference between the current and previous rounds of validation accuracy, weighted... The target is 300, which encourages improvements in accuracy. This is the difference between the current accuracy and the historical best accuracy, rewarding breakthroughs in performance. Stability is the stability of the accuracy over the last 5 validation runs, avoiding metric fluctuations. Penalty is an overfitting penalty; it is triggered when the difference between the training accuracy and the validation accuracy is greater than 0.1, thus suppressing overfitting.

[0036] (2) 2.3 Training and Evaluation Process The prediction module uses a pre-reserved test set to test the recognition ability of the trained CNN model, outputs the final prediction result, and determines whether the model is qualified. The 11 probability values ​​output by the CNN correspond to 11 signal types. The argmax prediction class index function is used to find the type index with the highest probability, and finally outputs a clear result indicating which type the test signal belongs to.

[0037] Traditional training methods with fixed hyperparameters are ill-suited to the complex convergence process in signal classification tasks, often resulting in slow convergence or overfitting. To address this issue, this embodiment employs a combined strategy of adaptive learning rate scheduling and augmented data.

[0038] For dynamic adjustment of the learning rate, a feedback mechanism based on the validation accuracy is designed. When the validation accuracy of consecutive patience rounds (set to 15) does not exceed the historical best value, the learning rate is automatically reduced to avoid the model getting trapped in local optima. The specific update method is shown in formula (3). This is the updated learning rate; This is the current learning rate; This is the attenuation coefficient, set to 0.5 to control the attenuation magnitude each time; It is the lower limit of the learning rate. Setting it to 1e-6 can prevent the training from stalling due to an excessively low learning rate.

[0039] (3) To mitigate overfitting, multi-dimensional data augmentation enhances sample diversity. Gaussian noise, time shift, amplitude scaling, and random masking are applied sequentially to the original signal. This enriches the training data distribution and improves the model's adaptability to signal variations. When applying Gaussian noise, randomly generated noise with a standard deviation of 0.005-0.02 is superimposed on the signal. During time shifting, the signal is translated ±3 units along the time axis, with zeros padded in the shifted region. During amplitude scaling, the signal amplitude is multiplied by a random coefficient of 0.9-1.1. During random masking, 10% of the time-point data is randomly masked.

[0040] The above description only illustrates the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. An adaptive robust modulation recognition method for complex electromagnetic environments, characterized in that, Includes the following steps: S1. Perform data preprocessing on the original IQ signal and reshape the preprocessed signal into a model-adaptive format; S2. Perform data enhancement on the preprocessed signal to simulate noise interference and transmission deviation in a complex electromagnetic environment; S3. An enhanced multi-scale CNN architecture is used to extract features and classify the enhanced signal. The convolution operation satisfies the following formula: Where: y is the output feature map of the convolutional layer. It is the ReLU activation function, used to introduce non-linearity and enhance feature representation; W is the convolution kernel weight matrix. It is a convolution operation, where x is the input signal and b is the bias term; S4. Monitor the CNN training state in real time using the PPO reinforcement learning module, construct a 10-dimensional state vector, and perform adaptive hyperparameter adjustment based on a multi-objective weighted reward function. The reward function satisfies the following equation: in: This is the final reward value; It is the difference between the current and previous rounds of validation accuracy, weighted... The target is 300, to encourage improved accuracy. It is the difference between the current accuracy and the historical best accuracy, rewarding breakthrough performance; stability is the stability of the accuracy in the last 5 rounds of validation, avoiding metric fluctuations; penalty is the overfitting penalty, which is triggered when the difference between the training accuracy and the validation accuracy is greater than 0.1, thereby suppressing overfitting. S5. Based on the verification accuracy feedback, perform adaptive learning rate scheduling. The learning rate update satisfies the following formula: in: This is the updated learning rate; This is the current learning rate; This is the attenuation coefficient, set to 0.5 to control the attenuation magnitude each time; It is the lower bound of the learning rate; S6. Input the trained CNN model into the test set for prediction and output the signal modulation type identification result.

2. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S1, the preprocessed 2×128 signal is reshaped into a (2,128,1) format.

3. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S2, the data augmentation specifically includes: superimposing Gaussian noise with a standard deviation of 0.005-0.02, shifting the time axis by ±3 units, scaling the amplitude by a random coefficient of 0.9-1.1, and masking 10% of the time point data with a random mask.

4. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S3, the enhanced multi-scale CNN architecture uses multi-scale convolutional kernels of (1,8) and (1,4) sizes. After max pooling, global average pooling, and two fully connected layers, it outputs probability values ​​corresponding to 11 signal types.

5. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S4, the PPO model includes a policy network and a value network with the same structure; the policy network outputs the probability distribution of five parameter tuning actions: increasing the learning rate, decreasing the learning rate, maintaining the learning rate, adjusting the batch size, and maintaining the batch size; the 10-dimensional state vector includes: training progress, current validation accuracy, current validation loss, recent average accuracy, accuracy fluctuation, recent average loss, loss fluctuation, overfitting degree, historical best accuracy, and the number of steps away from the best epoch.

6. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S4, the overfitting penalty term is triggered when the difference between the training accuracy and the validation accuracy is greater than 0.

1.

7. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S5, the attenuation coefficient Set the lower bound of the learning rate to 0.

5. Set to 1e-6; when the accuracy does not exceed the historical best value in 15 consecutive rounds of validation, perform learning rate decay.

8. The adaptive robust modulation identification method for complex electromagnetic environments according to claim 1, characterized in that: In step S6, the argmax function is used to select the category index with the highest probability to determine the signal modulation type.

9. An adaptive robust modulation identification system for complex electromagnetic environments, characterized in that: It includes a data preprocessing module, a data augmentation module, a CNN model training module, a PPO model enhancement module, and a prediction module; The signal output terminal of the data preprocessing module is connected to the signal input terminal of the data augmentation module, and the signal output terminal of the data augmentation module is connected to the signal input terminal of the CNN model training module. The PPO model enhancement module is bidirectionally connected to the CNN model training module, and the adaptive learning rate scheduling unit is connected to both the CNN model training module and the PPO model enhancement module. The output of the CNN model training module is connected to the input of the prediction module; The data preprocessing module is used to receive the raw IQ signal and convert it into standardized input data; The data augmentation module is used to apply Gaussian noise, time offset, amplitude scaling, and random masking to the preprocessed signal. The prediction module is used to output the signal modulation type identification result based on the trained CNN model.

10. The adaptive robust modulation identification system for complex electromagnetic environments according to claim 9, characterized in that: The CNN model training module adopts an enhanced multi-scale CNN architecture, which includes three convolutional blocks connected sequentially to a max pooling layer, a global average pooling layer, a 512-neuron fully connected layer, and a 256-neuron fully connected layer. The policy network and value network of the PPO model enhancement module are both constructed from fully connected layers, batch normalization layers, and dropout layers.

Images