A lightweight modulation recognition method and system based on noise perception and multi-task learning
By employing a signal modulation recognition method based on noise perception and multi-task learning, a signal modulation recognition model with an LSTM module and a noise perception module is constructed. By combining adversarial training strategies and phased fine-tuning, the robustness and computational efficiency issues of existing modulation recognition methods in complex electromagnetic scenarios are resolved, achieving high accuracy and low complexity in signal modulation recognition.
Patent Information
- Application Number
- CN202511362530.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-23
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-09-23
AI Technical Summary
Existing modulation recognition methods lack robustness in complex electromagnetic scenarios and have low computational efficiency. In particular, the recognition accuracy drops significantly under low signal-to-noise ratio conditions, and there is a lack of a unified framework with high accuracy, low complexity, and strong anti-interference capabilities.
A lightweight modulation recognition method based on noise perception and multi-task learning is adopted. A signal modulation mode recognition model is constructed, including an input layer, a shared feature extraction layer and an output layer. The LSTM module is used to extract time-dependent features, and the noise perception module is combined to suppress noise. The model is trained using an adversarial training strategy and a staged fine-tuning strategy.
It significantly improves the accuracy and robustness of signal modulation recognition, especially performing well in low signal-to-noise ratio environments, and achieves lightweight and anti-interference capabilities of the model, adapting to complex communication environments.
Smart Images

Figure CN122741295B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of communication signal processing technology, and in particular to a lightweight modulation recognition method and system based on noise perception and multi-task learning. Background Technology
[0002] In modern communications and electromagnetic countermeasures applications, signal modulation pattern identification is crucial for improving communication network performance, optimizing spectrum resource utilization, and ensuring communication security. Traditional modulation identification methods rely on manual extraction of features such as cyclic spectra and higher-order cumulants, which have significant drawbacks, including complex feature extraction processes, reliance on prior knowledge, and difficulty in adapting to complex signal environments. In particular, under low signal-to-noise ratio conditions, the feature extraction capability drops significantly, leading to a substantial decrease in identification accuracy.
[0003] In recent years, deep learning technology has provided new approaches for signal modulation pattern recognition. For example, Convolutional Neural Networks (CNNs) perform well in processing complex-valued time radio signals, and Recurrent Neural Networks (RNNs) have certain advantages in time series signal processing. However, CNNs lack time sensitivity, and RNNs consume large amounts of computational resources, limiting their practical applications. While existing deep learning-based methods have made some progress in complex electromagnetic scenarios, they still suffer from insufficient robustness and low computational efficiency. With the advent of the 6G era, radio signal modulation methods are becoming increasingly complex, communication environments are becoming more challenging, and low signal-to-noise ratio situations are common, placing higher demands on modulation recognition.
[0004] Current methods for addressing the robustness and lightweighting challenges in modulation scheme recognition include introducing lightweight designs such as group convolution to compress parameter size, or utilizing adversarial training strategies to improve recognition rates in low signal-to-noise ratio (SNR) scenarios. However, these solutions often only address single problems and lack a unified framework for integrating and optimizing recognition accuracy, model computational efficiency, and anti-interference capabilities. Therefore, in resource-constrained edge devices and complex electromagnetic countermeasures scenarios, there is an urgent need for a signal modulation recognition scheme that simultaneously possesses high accuracy, low complexity, and strong anti-interference capabilities. Summary of the Invention
[0005] This invention provides a lightweight modulation recognition method and system based on noise perception and multi-task learning to overcome the above-mentioned technical problems.
[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A lightweight modulation recognition method based on noise perception and multi-task learning includes the following steps: S1: Modulate the acquired raw communication signal to obtain the modulated IQ signal, and preprocess the IQ signal to obtain time series data; S2: Construct a signal modulation pattern recognition model based on noise perception and multi-task learning; Furthermore, the signal modulation method recognition model includes an input layer, a shared feature extraction layer with adjustable feature dimensions, and an output layer; Time series data is input into the shared feature extraction layer through the input layer; The shared feature extraction layer includes several cascaded shared feature extraction modules, and each shared feature extraction module includes an LSTM module and a noise perception module connected in sequence. The time dependency features in the time series data are extracted using the LSTM module; the noise perception module is used to suppress noise in the time dependency features to obtain noise perception features. Based on the output layer SNR network task branch and the modulation type MT network task branch, multi-task joint optimization is performed according to noise perception characteristics to achieve signal modulation mode identification. S3: Based on the adversarial training strategy and the phased fine-tuning strategy, the signal modulation mode recognition model is trained according to the time series data to obtain a lightweight signal modulation recognition model, and the signal modulation mode is recognized according to the lightweight signal modulation recognition model.
[0007] Furthermore, in S2, the method for obtaining noise-perceived features by performing noise suppression processing on the time-dependent features through the noise perception module specifically includes: S21: Obtain the time dependency feature set H output by the LSTM module; H= ,in Indicates time step The corresponding time dependency features; S22: Using the feature extraction module, global features of the time series data are extracted based on the time dependency feature set H, and the formula for obtaining the global features is as follows: , In the formula: Represents global features; Indicates batch size; T This indicates the sequence length of the time series data; D Indicates feature dimension; Indicates the first t Input at each time step ; S23: Obtain the noise mask vector based on the extracted global features through the constructed noise estimation network; And the expression for the noise estimation network is: ; , In the formula: The noise level is represented by the linear layer and activation function; σ represents... function; and Represents the weight matrix; Indicates the bias term; Represents the noise mask vector; S24: Obtain the gated vector based on the output of the LSTM module through a gated generation network; And the formula for obtaining the gating vector is: , In the formula: Represents the gate vector; and This represents the gating layer parameters of the gated generation network; S25: Obtain the feature enhancement vector based on the output of the LSTM module through the feature enhancement network; And the formula for obtaining the feature enhancement vector is: , In the formula: Represents the feature enhancement vector; and These represent the parameters of the enhancement layers in the feature enhancement network; S26: Fuse the feature enhancement vector, the gate vector, the output of the LSTM module, and the noise mask vector to obtain the feature fusion vector; And the expression for the feature fusion vector is: , In the formula: This represents the feature fusion vector.
[0008] Furthermore, in S2, the method of performing multi-task joint optimization based on the output layer SNR network branch and the modulation type MT network branch to achieve modulation signal recognition according to noise perception characteristics specifically includes: S100: Obtain the Nth time-step output of the noise perception module in the last shared feature extraction module of the shared feature extraction layer. , respectively, serve as inputs to the SNR network task branch and the modulation type MT network task branch; S200: Through the SNR network task branch and the modulation type MT network task branch, based on the output Perform multi-task joint optimization to obtain prediction results including modulation recognition task loss and signal-to-noise ratio classification task loss; and realize the identification of signal modulation mode based on the prediction results.
[0009] Furthermore, S3 specifically includes the following steps: S31: Construct the loss function for training the signal modulation pattern recognition model; S32: Define the adversarial training strategy, which is to obtain adversarial samples based on time series data using a loss function combined with the Fast Gradient Sign Method (FGSM). And the formula for obtaining adversarial examples is: , In the formula: This represents the original input signal, i.e., time series data; Indicates the magnitude of the disturbance; Represents a symbolic function; Represents the loss function For input The gradient; This indicates the acquired adversarial example; A label indicating the modulation scheme of the signal corresponding to the time series data; Indicates learnable parameters; S32: Based on a phased fine-tuning strategy, a lightweight signal modulation recognition model is obtained by training the signal modulation recognition model through adversarial examples. The model training specifically includes: S321: Randomly divide the adversarial examples into two training sets and two test sets; S322: Keep the model parameters of each noise perception module in the signal modulation mode recognition model constant, and use any set of training data to perform the first stage of model training for the signal modulation mode recognition model. Then, based on any set of test data and the loss function, confirm whether the output of the signal modulation mode recognition model after model training has converged. The signal modulation scheme recognition model corresponding to the confirmed output convergence is used as the optimization model; otherwise, the model parameters of the trained signal modulation scheme recognition model are adaptively adjusted based on the backpropagation method, and the first stage of model training process is repeated. S323: Set the model parameters of each noise perception module in the optimization model to be learnable parameters; The second stage of model training was performed using another set of training data and the output of the optimized model was confirmed to have converged based on another set of test data and the loss function. The optimized model corresponding to the confirmed output convergence is used as the lightweight signal modulation recognition model; otherwise, the model parameters of the optimized model are adaptively adjusted based on the backpropagation method, and the second stage of model training process is repeated.
[0010] Furthermore, the loss function constructed in S31 has the following expression: , , , In the formula: This represents the total loss after weighting; Indicated for balance Loss and The weighting factor for the loss; This represents the final loss for the signal-to-noise ratio classification task; Represents the cross-entropy loss function; and This represents a single-layer fully connected network. express Learnable parameters of network task branches; This indicates the corresponding SNR tag; This represents the final loss of the signal modulation type classification task; Learnable parameters representing the modulation type of the MT network task branch; The set of labels representing the signal modulation type classification task.
[0011] Furthermore, the expression for the LSTM module in S2 is: , , , , , , In the formula: They represent the first Layer LSTM at time steps The activation values of the input gate, forget gate, and output gate; Indicates the state of candidate cells; This indicates the updated cell state; Indicates the connection to the previous hidden state. The weights to the input gate; This indicates the connection to the cell state at the previous time step of the current layer. The weights to the input gate; This represents the bias term of the input gate; This represents the weight matrix connecting the previous hidden state to the forget gate; This represents the weight matrix connecting the cell state of the previous time step in the current layer to the forget gate; Indicates the forget gate bias term; This represents the weight matrix connecting the hidden state of the previous layer to the candidate cell state; Bias terms representing the state of candidate cells; This represents the weight matrix connecting the hidden state of the previous layer to the output gate; This represents the weight matrix connecting the current cell state to the output gate; Indicates the output gate bias term; express Activation function; This represents element-wise multiplication. This represents the hyperbolic tangent activation function.
[0012] Furthermore, the formula for preprocessing the IQ signal to obtain time series data in S1 is as follows: , In the formula: and These represent the minimum and maximum values in the original signal sample, i.e., the IQ signal; The first one represents the original signal sample. each sample value and ; Indicates the number of samples; This represents the normalized IQ signal; And record the time series data as .
[0013] A lightweight modulation recognition system based on noise perception and multi-task learning includes a signal modulation module, a data preprocessing module, a model building module, a model training module, and a model deployment module. The signal modulation module is used to perform different types of signal modulation on the original communication signal based on a preset signal modulation method to obtain IQ signals; The data preprocessing module is used to preprocess the IQ signal to obtain time series data; The model building module is used to build a signal modulation mode recognition model based on noise perception and multi-task learning; The model training module is used to construct a loss function and, based on an adversarial training strategy and a phased fine-tuning strategy, train the signal modulation mode recognition model according to time series data to obtain a lightweight signal modulation recognition model. The model deployment module is used to deploy the lightweight signal modulation recognition model to a pre-set device or server to achieve the recognition of signal modulation mode.
[0014] This invention provides a lightweight modulation recognition method and system based on noise perception and multi-task learning, with the following advantages: (1) Construct a signal modulation mode recognition model based on noise perception and multi-task learning. Through the designed noise perception module, signal features can be accurately extracted and noise interference can be effectively suppressed. Furthermore, by combining the LSTM module with the noise perception module, the signal time dependence can be captured, enhancing the model's ability to perceive noise, thus laying a good foundation for signal modulation recognition.
[0015] (2) Compared with existing classical models, the signal modulation mode recognition model proposed in this invention significantly improves the recognition accuracy under different signal-to-noise ratio conditions, especially in low signal-to-noise ratio environments. An adversarial training strategy is adopted, which generates adversarial examples through the FGSM method and combines it with a phased training fine-tuning strategy to train the model, thereby improving the robustness and generalization ability of the model and ensuring stable convergence of training.
[0016] (3) Introducing a multi-task learning architecture: namely, a framework formed by a signal modulation mode recognition model, an adversarial training strategy, and a phased fine-tuning strategy; this invention suppresses noise through a noise perception module, achieves model lightweighting through multi-task training and learning and optimizing the network structure, and the multi-task and adversarial training strategies can synergistically improve model performance. At the same time, it has significant advantages in robustness, lightweighting and integrated optimization, providing a solution for modulation recognition in complex environments. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0018] Figure 1 This is a flowchart of the lightweight modulation recognition method based on noise perception and multi-task learning according to the present invention. Figure 2 This is a schematic diagram of the signal modulation mode recognition model based on noise perception and multi-task learning constructed in this embodiment. Figure 3 This is a schematic diagram of the working principle of the noise sensing module in this embodiment; Figure 4 This is a schematic diagram illustrating the changes in training loss and validation loss during the first training phase of the model in this embodiment; Figure 5 This is a schematic diagram illustrating the change in accuracy during the first training phase of the model in this embodiment; Figure 6 This is a schematic diagram illustrating the change in accuracy during the second training phase of the model in this embodiment; Figure 7This is a schematic diagram comparing the performance of different models under different SNRs in this embodiment; Figure 8 This is a schematic diagram of the normalized confusion matrix of each model in this embodiment; Figure 9 This is a schematic diagram comparing the noise resistance performance of the LSTM variant model in this embodiment; Figure 10 This is a schematic diagram of parameter sensitivity analysis in this embodiment. Figure 11 This is a schematic diagram showing the performance comparison under different SNR values in this embodiment. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] This embodiment provides a lightweight modulation recognition method based on noise perception and multi-task learning, such as Figure 1 As shown, it includes the following steps: S1: Modulate the acquired raw communication signal to obtain the modulated IQ signal, and preprocess the IQ signal to obtain time series data; Specifically, signal modulation is the process or method of changing certain characteristics of one waveform according to another waveform or signal. Modulation can shift the spectrum of the original communication signal, and the modulated communication signal is called the modulated signal, i.e., the IQ signal. The modulated signal carries information and is suitable for transmission in the channel. In particular, signal modulation is to use the baseband signal to control the change of one or more parameters of the carrier signal, and to load information onto it to form a modulated signal for transmission. Demodulation is the reverse process of modulation, which can recover the original baseband signal from the parameter changes of the modulated signal.
[0021] Specifically, the formula for obtaining time series data by preprocessing the IQ signal is as follows: , In the formula: and These represent the minimum and maximum values in the original signal sample, i.e., the IQ signal; The first one represents the original signal sample. each sample value and ; Indicates the number of samples; This represents the normalized IQ signal; And record the time series data as .
[0022] S2: Construct a signal modulation pattern recognition model based on noise perception and multi-task learning; The signal modulation mode identification model is denoted as IANA-MTLSTM, such as Figure 2 As shown, the signal modulation mode recognition model includes an input layer, a shared feature extraction layer with adjustable feature dimensions, and an output layer; Time series data is input into the shared feature extraction layer through the input layer; The shared feature extraction layer includes several cascaded shared feature extraction modules. The number of shared feature extraction modules in the layer is adjusted according to the required feature dimension. Each shared feature extraction module includes an LSTM module and a noise perception module connected in sequence. In this embodiment, a shared feature extraction layer comprising an LSTM module and a noise perception module is used to extract features from time series data, including: Extract time dependency features from time series data using the LSTM module; The noise perception module performs noise suppression processing on the time-dependent features to obtain noise perception features. Specifically, the shared feature extraction layer of the model receives time-series data and processes it through multiple LSTM units and a noise sensing block; the LSTM units process the IQ signal data sampled at each time step. As input, get the output. h i and will and h i As the input to the next LSTM unit, it can effectively model the temporal information of the signal, while the noise-sensing block takes the output of the LSTM layer at each time step. As input, these features are processed to obtain a new representation, namely noise-aware features, which helps enhance the model's robustness to noise. The LSTM module is used to capture the temporal dependencies of sequence data. Output of layer LSTM Output from the previous layer and the cell state at the current time step Together, we calculate that the expression for the LSTM module is: , , , , , , In the formula: They represent the first Layer LSTM at time steps The activation values of the input gate, forget gate, and output gate; Indicates the state of candidate cells; This indicates the updated cell state; Indicates the connection to the previous hidden state. The weights to the input gate; This indicates the connection to the cell state at the previous time step of the current layer. The weights to the input gate; This represents the bias term of the input gate; This represents the weight matrix connecting the previous hidden state to the forget gate; This represents the weight matrix connecting the cell state of the previous time step in the current layer to the forget gate; Indicates the forget gate bias term; This represents the weight matrix connecting the hidden state of the previous layer to the candidate cell state; Bias terms representing the state of candidate cells; This represents the weight matrix connecting the hidden state of the previous layer to the output gate; This represents the weight matrix connecting the current cell state to the output gate; Indicates the output gate bias term; express Activation function; This represents element-wise multiplication. This represents the hyperbolic tangent activation function.
[0023] like Figure 3 As shown, in the output of the LSTM module Based on this, a noise perception module performs noise suppression processing on the features to obtain noise-perceived features. Specifically, the noise perception module achieves adaptive filtering and original signal feature enhancement through noise level estimation, thereby enhancing the robustness of the model's recognition. The expression for the noise perception module is: ; In the formula: These represent the parameters of the noise sensing module; Represents the noise perception function; Indicates the first The layer is obtained after processing by the noise sensing module. Characteristic representation of time-series signals; like Figure 3 As shown, the specific processing steps of the noise perception module include: S21: Obtain the time dependency feature set H output by the LSTM module; H= , in, Indicates time step The corresponding time dependency features; S22: Using the feature extraction module, global features of the time series data are extracted based on the time dependency feature set H, and the formula for obtaining the global features is as follows: , In the formula: Represents global features; Indicates batch size; T This indicates the sequence length of the time series data; D Indicates feature dimension; Indicates the first t Input at each time step ; S23: Obtain the noise mask vector based on the extracted global features through the constructed noise estimation network; And the expression for the noise estimation network is: ; , In the formula: The noise level is represented by the linear layer and activation function; σ represents... function; and Represents the weight matrix; Indicates the bias term; Represents the noise mask vector; S24: Obtain the gated vector based on the output of the LSTM module through a gated generation network; And the formula for obtaining the gating vector is: , In the formula: Represents the gate vector; and This represents the gating layer parameters of the gated generation network; S25: Obtain the feature enhancement vector based on the output of the LSTM module through the feature enhancement network; And the formula for obtaining the feature enhancement vector is: , In the formula: Represents the feature enhancement vector; and These represent the parameters of the enhancement layers in the feature enhancement network; S26: Fuse the feature enhancement vector, the gate vector, the output of the LSTM module, and the noise mask vector to obtain the feature fusion vector; And the expression for the feature fusion vector is: , In the formula: Represents the feature fusion vector; Based on the output layer's SNR network task branch and modulation type MT network task branch, multi-task joint optimization is performed according to noise perception characteristics to achieve signal modulation mode identification, specifically including: S100: Obtain the Nth time-step output of the noise perception module in the last shared feature extraction module of the shared feature extraction layer. , respectively, serve as inputs to the SNR network task branch and the modulation type MT network task branch; S200: Through the SNR network task branch and the modulation type MT network task branch, based on the output Perform multi-task joint optimization to obtain prediction results including modulation recognition task loss and signal-to-noise ratio classification task loss, and realize the identification of signal modulation mode based on the prediction results; In this embodiment, the signal modulation method recognition model uses two task heads. and To achieve multi-task joint optimization, specifically: and The output at the Nth time step of the last noise-sensing module in the shared feature extraction layer. As input, the prediction value for the corresponding task is generated through an independent neural network mapping; the method or technical principle of generating the prediction value for the corresponding task through an independent neural network mapping is existing well-known technology and will not be elaborated on here. S3: Based on adversarial training and phased fine-tuning strategies, the signal modulation mode recognition model is trained using time-series data to obtain a lightweight signal modulation mode recognition model. The recognition of signal modulation modes is then achieved using this lightweight model. Specifically, the following steps are included: S31: Construct the loss function for training the signal modulation pattern recognition model; Specifically, the constructed loss function is as follows: , , , In the formula: This represents the total loss after weighting; Indicated for balance Loss and The weighting factor of the loss and ; This represents the final loss for the signal-to-noise ratio classification task; Represents the cross-entropy loss function; and This represents a single-layer fully connected network. express Learnable parameters of network task branches; This indicates the corresponding SNR tag; This represents the final loss of the signal modulation type classification task; Learnable parameters representing the modulation type of the MT network task branch; The set of labels representing the signal modulation type classification task.
[0024] S32: Define the adversarial training strategy, which is to obtain adversarial samples based on time series data using a loss function combined with the Fast Gradient Sign Method (FGSM). This embodiment employs an adversarial training strategy to enhance the model's robustness and generalization ability. By introducing the FGSM strategy during training to generate adversarial examples, the model can learn more robust feature representations. Specifically: first, the model is forward-propagated using the original samples to calculate the loss function value of the original samples; then, adversarial examples are generated based on the gradient information of the original samples using the FGSM attack function. The formula for obtaining the adversarial examples is as follows: , In the formula: This represents the original input signal, i.e., time series data; Indicates the magnitude of the disturbance; Represents a symbolic function; Represents the loss function For input The gradient; This indicates the acquired adversarial example; A label indicating the modulation scheme of the signal corresponding to the time series data; Indicates learnable parameters; In this embodiment, the generated adversarial sample has the same label as the original sample. By performing forward propagation on the model again, the loss function value of the adversarial sample can be calculated. The loss of the adversarial sample is combined with the loss of the original sample, and the model parameters are updated through backpropagation. S32: Based on a phased fine-tuning strategy, a lightweight signal modulation recognition model is obtained by training the signal modulation recognition model through adversarial examples. In this embodiment, the phased fine-tuning strategy is divided into two phases: The first stage involves freezing the parameters of the noise estimation submodule of the noise perception module in the signal modulation mode identification model, training only the main part of the model, and quickly learning basic features and data patterns. The second stage involves unfreezing the signal modulation method and identifying the parameters of the noise estimation submodule of the noise perception module in the model. This stage also involves jointly optimizing all parameters of the entire model to reduce the learning rate and improve the ability to perceive noise characteristics. Specifically, the following steps are included: S321: Randomly divide the adversarial examples into two training sets and two test sets; S322: Keep the model parameters of each noise perception module in the signal modulation mode recognition model constant, and use any set of training data to perform the first stage of model training for the signal modulation mode recognition model. Then, based on any set of test data and the loss function, confirm whether the output of the signal modulation mode recognition model after model training has converged. The signal modulation scheme recognition model corresponding to the confirmed output convergence is used as the optimization model; otherwise, the model parameters of the trained signal modulation scheme recognition model are adaptively adjusted based on the backpropagation method, and the first stage of model training process is repeated. S323: Set the model parameters of each noise perception module in the optimization model to be learnable parameters; The second stage of model training was performed using another set of training data and the output of the optimized model was confirmed to have converged based on another set of test data and the loss function. The optimized model corresponding to the confirmed output convergence is used as the lightweight signal modulation recognition model; otherwise, the model parameters of the optimized model are adaptively adjusted based on the backpropagation method, and the second stage of model training process is repeated.
[0025] A lightweight modulation recognition system based on noise perception and multi-task learning includes a signal modulation module, a data preprocessing module, a model building module, a model training module, and a model deployment module. The signal modulation module is used to perform different types of signal modulation on the original communication signal based on a preset signal modulation method to obtain IQ signals; The data preprocessing module is used to preprocess the IQ signal to obtain time series data; The model building module is used to build a signal modulation mode recognition model based on noise perception and multi-task learning; The model training module is used to construct a loss function and, based on an adversarial training strategy and a phased fine-tuning strategy, train the signal modulation mode recognition model according to time series data to obtain a lightweight signal modulation recognition model. The model deployment module is used to deploy the lightweight signal modulation recognition model to a pre-set device or server to achieve the recognition of signal modulation mode.
[0026] The beneficial effects of the method and system described in this embodiment are as follows: (1) Construct a signal modulation mode recognition model based on noise perception and multi-task learning. Through the designed noise perception module, signal features can be accurately extracted and noise interference can be effectively suppressed. Furthermore, by combining the LSTM module with the noise perception module, the signal time dependence can be captured, enhancing the model's ability to perceive noise, thus laying a good foundation for signal modulation recognition.
[0027] (2) Compared with existing classical models, the signal modulation mode recognition model proposed in this invention significantly improves the recognition accuracy under different signal-to-noise ratio conditions, especially in low signal-to-noise ratio environments. An adversarial training strategy is adopted, which generates adversarial examples through the FGSM method and combines it with a phased training fine-tuning strategy to train the model, thereby improving the robustness and generalization ability of the model and ensuring stable convergence of training.
[0028] (3) Introducing a multi-task learning architecture: namely, a framework formed by a signal modulation mode recognition model, an adversarial training strategy, and a phased fine-tuning strategy; this invention suppresses noise through a noise perception module, achieves model lightweighting through multi-task training and learning and optimizing the network structure, and the multi-task and adversarial training strategies can synergistically improve model performance. At the same time, it has significant advantages in robustness, lightweighting and integrated optimization, providing a solution for modulation recognition in complex environments.
[0029] The method and system proposed in this embodiment will be further illustrated through specific experiments: A. Experimental Dataset: The RML2016.10a dataset used in the experiments described in this embodiment is a benchmark dataset widely used in radio signal classification tasks, as shown in Table 1. It contains radio signal samples with various modulation schemes, covering modulation types from low to high order, and can effectively simulate the signal characteristics of different modulation schemes in real communication environments.
[0030] Table 1. RML2016.10a Dataset Composition
[0031] As shown in Table 1, the dataset contains 11 different digital and analog modulation schemes, including amplitude modulation (AM), frequency modulation (FM), phase modulation (PM), and various quadrature amplitude modulation (QAM) schemes. Each modulation scheme is sampled under multiple signal-to-noise ratio (SNR) conditions, with SNR ranging from -20dB to 30dB in 2dB intervals, comprehensively reflecting the signal characteristics under different levels of noise interference. The dataset is organized as a dictionary, where key-value pairs correspond to different modulation schemes and SNR combinations. Each combination corresponds to 1000 signal samples, and each sample has 128 sampling points. Each sampling point contains two signal data streams: in-phase (I) and quadrature (Q) components.
[0032] B. Experimental comparison model and model training: Several classic machine learning and deep learning models were selected as baseline models for comparative experiments with the proposed IANA-MTLSTM signal modulation scheme identification model in this embodiment. These baseline models cover different network structures and learning strategies, comprehensively reflecting the advantages of the proposed model in the radio signal modulation scheme identification task, including LSTM networks, gated recurrent units (GRUs), autoencoders (AEs), denoising autoencoders (DAEs), PETCGDNN models, CNN networks, and multi-convolutional deep neural networks (MCLDNNs). The proposed IANA-MTLSTM signal modulation scheme identification model was trained, and the training parameters are shown in Table 2.
[0033] Table 2. Model Training Parameters
[0034] C. Comparison of experimental results: The total training cycles are set to 150, with the first phase consisting of 100 training cycles.
[0035] like Figure 4 As shown, where Figure 4 (a) is a schematic diagram of the training loss. Figure 4 (b) is a schematic diagram of the validation loss. In the early stages of training, the loss value is high, indicating a poor fit between the model and the training and validation data. As training progresses, the model gradually learns the characteristics of the data, and both the training and validation losses decrease rapidly, indicating that the model's performance is continuously improving. In the later stages of training, the training loss continues to decrease slowly, while the validation loss tends to stabilize after a certain number of rounds, indicating that the model's performance on the validation set is stabilizing. Figure 5 As shown, where Figure 5 (a) Schematic diagram showing the changes in training accuracy and validation accuracy for modulation scheme classification. Figure 5(b) is a schematic diagram illustrating the changes in training accuracy and validation accuracy for signal-to-noise ratio (SNR) classification. In this embodiment, the trends in accuracy and loss of the model on the training and validation sets correspond during training. Furthermore, the modulation scheme classification accuracy is higher than that of the SNR classification task during training, indicating that the model has a strong ability to identify modulation schemes. This is because the SNR classification task itself is more difficult, requiring more training rounds and more complex network parameter settings.
[0036] like Figure 6 As shown, where Figure 6 (a) is a schematic diagram showing the change in accuracy of the signal modulation type prediction model. Figure 6 (b) is a schematic diagram illustrating the change in the accuracy of the SNR value prediction model; from Figure 6 (a) It can be seen that during the overall parameter training phase of training phase two, the modulation accuracy increased significantly in the early stage of training, reaching its highest level at 105 rounds, and remained relatively stable in subsequent training, with an overall trend of gradual decline. For example... Figure 6 As shown in (b), the signal-to-noise ratio (SNR) accuracy rapidly increased from 0.22 to 0.28 in the early stages of training (approximately 100-110 epochs), then the growth rate slowed down, with an overall steady upward trend. As training progressed, the SNR accuracy continued to gradually increase, indicating that the model has a stronger adaptability to complex tasks. Figure 7 As shown, the recognition accuracy of the IANA-MTLSTM model significantly improves with increasing SNR, and outperforms other baseline models at most SNR values. Particularly under SNR conditions, this model exhibits a clear advantage, demonstrating its ability to better extract signal features from noise, thus achieving more accurate recognition. In contrast, traditional models such as LSTM, CNN, and DAE show lower recognition accuracy at low SNR. While their accuracy improves with increasing SNR, the improvement is not as significant as that of the IANA-MTLSTM model. Furthermore, under high SNR conditions, the recognition accuracy of the IANA-MTLSTM model tends to stabilize and remain at a high level, while other models exhibit some fluctuations. This indicates that the proposed IANA-MTLSTM model can consistently demonstrate its superior recognition performance under different signal-to-noise ratio environments, validating the model's effectiveness and robustness.
[0037] like Figure 8 As shown, where Figure 8 (a) is a CNN; Figure 8 (b) is DAE; Figure 8 (c) is a GRU; Figure 8 (d) is an LSTM; Figure 8 (e) is IANA-MTLSTM; Figure 8(f) represents PETCGDNN; the darkness of the diagonal elements in the confusion matrix shows that the IANA-MTLSTM model has a significant advantage in signal modulation type classification tasks. Darker diagonal elements indicate higher classification accuracy. In contrast, traditional models such as CNN, DAE, and GRU show more pronounced misclassification among similar modulation types. For example, the CNN model cannot completely distinguish between signals under QAM16 and QAM64 modulation modes; PETCGDNN, LSTM, and MCLDNN models show some performance improvement, but their overall performance is still inferior to the IANA-MTLSTM model; MCLDNN is currently the best performing model on the RML201610.a dataset, but its performance in classifying the easily misclassified QAM16 and QAM64 modulation modes differs significantly from IANA-MTLSTM. Table 3 shows the overall signal-to-noise ratio performance of different models for the two higher-order modulation modes, QAM16 and QAM64. In summary, the IANA-MTLSTM model exhibits superior performance in high-order modulation signal recognition, effectively reducing misclassification and improving classification accuracy and reliability.
[0038] Table 3. Performance of different models on the overall signal-to-noise ratio under QAM16 and QAM64.
[0039] like Figure 9As shown, ablation experiments were conducted to compare the classification performance of various LSTM variant models under different SNRs. Compared with the basic LSTM model (blue dotted line), adding adversarial training modules (PGD, FGSM, TRADES) and multi-task learning frameworks can effectively improve the classification ability of LSTM models under different noise environments. Among them, the multi-task LSTM+FGSM combination scheme has the best effect, demonstrating the synergistic effect of adversarial training and multi-task learning, and providing strong support for the application of the model under complex noise conditions. By layering different modules and tuning parameters, the performance of the original model steadily improved. Specifically, the LSTM+PGD model with added PGD adversarial training (orange triangle) showed a significant performance improvement in the low to medium signal-to-noise ratio (-10dB to 10dB) range, with an average accuracy increase of about 3% to 4% compared to the basic LSTM. The LSTM+PGD+TRADES model (green square) combined with the TRADES method showed stronger stability in the high signal-to-noise ratio region (above 10dB), with smaller fluctuations in its accuracy curve, indicating enhanced generalization ability. The LSTM+FGSM model using FGSM (purple diamond) performed well in the high signal-to-noise ratio (10dB to 18dB) range. The multi-task LSTM model (red asterisk) exhibits significant advantages in the signal-to-noise ratio (SNR) range, achieving an accuracy approximately 2% higher than the basic LSTM. The multi-task LSTM model (red asterisk) within the multi-task learning framework performs exceptionally well across the entire SNR range, maintaining high classification accuracy in both low-noise and high-noise environments. The multi-task LSTM+FGSM model (brown dotted line), combining multi-task learning and FGSM, demonstrates the best performance among all variants. In the medium-to-high SNR region (0dB to 18dB), its accuracy remains consistently around 0.94, further approaching 1.0 at high SNR, showcasing excellent robustness and adaptability. However, its performance is subpar in the -12dB to -10dB range, indicating room for further improvement.
[0040] like Figure 10 As shown, the weights of the modulation type prediction task in a multi-task learning model are illustrated. The impact on the relationship between signal-to-noise ratio (SNR) and accuracy. Experiments show that as SNR increases, all... The accuracy curves corresponding to the values all show an upward trend and tend to stabilize in the high signal-to-noise ratio region, indicating that the model can still be adjusted in environments with high noise interference. The optimized values demonstrate a certain level of anti-interference capability and good performance robustness. It is worth noting that when the SNR is below -5dB in the low signal-to-noise ratio range, different... The accuracy curves corresponding to these values show significant differences; however, as the SNR increases above -5dB, the curves gradually converge and eventually reach similar accuracy levels (approximately 0.85-0.9). Within the -15dB to 0dB range, The value primarily affects model performance by adjusting the initial slope of the curve and the magnitude of its increase before the inflection point. Specifically, in the low signal-to-noise ratio (-15dB to -10dB) region, low... Values (such as 0.5) exhibit stronger adaptability, with a relatively flat accuracy curve but higher stability; while high values... Values (e.g., 0.99) show a steeper initial growth trend, indicating that they are more sensitive to accuracy improvements at low signal-to-noise ratios. When the SNR approaches the inflection point (-5dB), the slope of all curves decreases significantly, at which point high... The rate of increase in the accuracy of low values has begun to slow down, while the rate of increase in low values has begun to slow down. The performance advantage of these values gradually diminishes. Within the -5dB to 0dB range, different... The curve of the value gradually approaches parallel development. Despite the high... The value (0.99) still maintains a lead of approximately 0.12 at 0 dB, but the relative gap between the curves has narrowed significantly (e.g., =0.99 and The accuracy difference at 0 dB (=0.5) decreases from 0.15 at -10 dB to 0.08, indicating that as the signal-to-noise ratio approaches its inflection point, the parameter... The impact on model performance gradually weakens, and eventually, at 0dB, the accuracy of all curves converges to similar levels. When SNR reaches 0dB, each... The accuracy of the values is close to saturation, indicating that the model performance no longer significantly depends on... The setting of the value. Further analysis shows that even The performance differences remained statistically insignificant even when the values varied between 0.5 and 0.99. This result validates the model's effectiveness. The parameters are highly insensitive, exhibit stable performance, and demonstrate good performance robustness.
[0041] like Figure 11As shown, the accuracy of the IANA-MTLSTM model proposed in this embodiment is higher than that of the MCLDNN model in the signal-to-noise ratio (SNR) range [8, -18], and MCLDNN is one of the best-performing models on this public dataset. Therefore, the IANA-MTLSTM model in this invention significantly improves the recognition accuracy under different SNR conditions, especially in low SNR environments. Furthermore, the IANA-MTLSTM model proposed in this embodiment is significantly superior to the MCLDNN model in terms of parameter size. The IANA-MTLSTM model proposed in this embodiment has 189,058 parameters, while the MCLDNN model has as many as 405,175 parameters, a reduction of approximately 50%. A comparison of the specific number of model parameters is shown in Table 4. This indicates that the IANA-MTLSTM model maintains high performance while significantly reducing model complexity and computational resource requirements. This lightweight design not only reduces hardware resource consumption but also significantly accelerates model training and inference speed, improving model response efficiency. Furthermore, lightweight models offer greater advantages in practical applications, especially in resource-constrained environments (such as mobile devices or embedded systems), better adapting to the demands of real-time signal processing and rapid decision-making. The optimization of parameter scale in the IANA-MTLSTM model not only demonstrates its superior computational efficiency but also provides greater feasibility and flexibility for deployment in a wide range of real-world applications.
[0042] Table 4. Comparison of the number of model parameters
[0043] In summary, this embodiment provides a lightweight modulation recognition method and system based on noise perception and multi-task learning, which solves the problems of low recognition accuracy and poor robustness of traditional methods in complex signal environments. At the same time, it meets the requirements for lightweight signal modulation recognition models in special scenarios, achieves efficient and accurate recognition of signal modulation methods, and makes the model framework portable.
[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A lightweight modulation recognition method based on noise perception and multi-task learning, characterized in that, Includes the following steps: S1: Modulate the acquired raw communication signal to obtain the modulated IQ signal, and preprocess the IQ signal to obtain time series data; S2: Construct a signal modulation pattern recognition model based on noise perception and multi-task learning; Furthermore, the signal modulation method recognition model includes an input layer, a shared feature extraction layer with adjustable feature dimensions, and an output layer; Time series data is input into the shared feature extraction layer through the input layer; The shared feature extraction layer includes several cascaded shared feature extraction modules, and each shared feature extraction module includes an LSTM module and a noise perception module connected in sequence. Extract time dependency features from time series data using the LSTM module; The noise perception module performs noise suppression processing on the time-dependent features to obtain noise perception features. Based on the output layer SNR network task branch and the modulation type MT network task branch, multi-task joint optimization is performed according to noise perception characteristics to achieve signal modulation mode identification. S3: Based on adversarial training and phased fine-tuning strategies, the signal modulation mode recognition model is trained using time-series data to obtain a lightweight signal modulation mode recognition model. The recognition of signal modulation modes is then achieved using this lightweight model. Specifically, the following steps are included: S31: Construct the loss function for training the signal modulation pattern recognition model; S32: Define the adversarial training strategy as obtaining adversarial samples based on time series data using a loss function combined with the Fast Gradient Sign Method (FGSM). And the formula for obtaining adversarial examples is: In the formula: This represents the original input signal, i.e., time series data; Indicates the magnitude of the disturbance; Represents a symbolic function; Represents the loss function For input The gradient; This indicates the acquired adversarial example; A label indicating the modulation scheme of the signal corresponding to the time series data; Indicates learnable parameters; S32: Based on a phased fine-tuning strategy, a lightweight signal modulation recognition model is obtained by training the signal modulation recognition model through adversarial examples. The model training specifically includes: S321: Randomly divide the adversarial examples into two training sets and two test sets; S322: Keep the model parameters of each noise perception module in the signal modulation mode recognition model constant, and use any set of training data to perform the first stage of model training for the signal modulation mode recognition model. Then, based on any set of test data and the loss function, confirm whether the output of the signal modulation mode recognition model after model training has converged. The signal modulation scheme recognition model corresponding to the confirmed output convergence is used as the optimization model; otherwise, the model parameters of the trained signal modulation scheme recognition model are adaptively adjusted based on the backpropagation method, and the first stage of model training process is repeated. S323: Set the model parameters of each noise perception module in the optimization model to be learnable parameters; The second stage of model training was performed using another set of training data and the output of the optimized model was confirmed to have converged based on another set of test data and the loss function. The optimized model corresponding to the confirmed output convergence is used as the lightweight signal modulation recognition model; otherwise, the model parameters of the optimized model are adaptively adjusted based on the backpropagation method, and the second stage of model training process is repeated.
2. The lightweight modulation recognition method based on noise perception and multi-task learning according to claim 1, characterized in that, In S2, the method for obtaining noise-perceived features by performing noise suppression processing on time-dependent features through a noise perception module includes: S21: Obtain the time dependency feature set H output by the LSTM module; H= ,in Indicates time step The corresponding time dependency features; S22: Using the feature extraction module, global features of the time series data are extracted based on the time dependency feature set H, and the formula for obtaining the global features is as follows: In the formula: Represents global features; Indicates batch size; T This indicates the sequence length of the time series data; D Indicates feature dimension; Indicates the first t Input at each time step ; S23: Obtain the noise mask vector based on the extracted global features through the constructed noise estimation network; And the expression for the noise estimation network is: ; In the formula: The noise level is represented by the linear layer and activation function; σ represents... function; and Represents the weight matrix; Indicates the bias term; Represents the noise mask vector; S24: Obtain the gated vector based on the output of the LSTM module through a gated generation network; And the formula for obtaining the gating vector is: In the formula: Represents the gate vector; and This represents the gating layer parameters of the gated generation network; S25: Obtain the feature enhancement vector based on the output of the LSTM module through the feature enhancement network; And the formula for obtaining the feature enhancement vector is: In the formula: Represents the feature enhancement vector; and These represent the parameters of the enhancement layers in the feature enhancement network; S26: Fuse the feature enhancement vector, the gate vector, the output of the LSTM module, and the noise mask vector to obtain the feature fusion vector; And the expression for the feature fusion vector is: In the formula: This represents the feature fusion vector.
3. The lightweight modulation recognition method based on noise perception and multi-task learning according to claim 2, characterized in that, The method in S2, which uses the output layer SNR network branch and the modulation type MT network branch to perform multi-task joint optimization based on noise perception characteristics to achieve modulation signal recognition, specifically includes: S100: Obtain the Nth time-step output of the noise perception module in the last shared feature extraction module of the shared feature extraction layer. , respectively, serve as inputs to the SNR network task branch and the modulation type MT network task branch; S200: Through the SNR network task branch and the modulation type MT network task branch, based on the output Perform multi-task joint optimization to obtain prediction results including modulation recognition task loss and signal-to-noise ratio classification task loss; and realize the identification of signal modulation mode based on the prediction results.
4. The lightweight modulation recognition method based on noise perception and multi-task learning according to claim 3, characterized in that, The loss function constructed in S31 is expressed as follows: In the formula: This represents the total loss after weighting; Indicated for balance Loss and The weighting factor of the loss and ; This represents the final loss for the signal-to-noise ratio classification task; Represents the cross-entropy loss function; and This represents a single-layer fully connected network. express Learnable parameters of network task branches; This indicates the corresponding SNR tag; This represents the final loss of the signal modulation type classification task; Learnable parameters representing the modulation type of the MT network task branch; The set of labels representing the signal modulation type classification task.
5. A lightweight modulation recognition method based on noise perception and multi-task learning according to claim 4, characterized in that, The expression for the LSTM module in S2 is: In the formula: They represent the first Layer LSTM at time steps The activation values of the input gate, forget gate, and output gate; Indicates the state of candidate cells; This indicates the updated cell state; Indicates the connection to the previous hidden state. The weights to the input gate; This indicates the connection to the cell state at the previous time step of the current layer. The weights to the input gate; This represents the bias term of the input gate; This represents the weight matrix connecting the previous hidden state to the forget gate; This represents the weight matrix connecting the cell state of the previous time step in the current layer to the forget gate; Indicates the forget gate bias term; This represents the weight matrix connecting the hidden state of the previous layer to the candidate cell state; Bias terms representing the state of candidate cells; This represents the weight matrix connecting the hidden state of the previous layer to the output gate; This represents the weight matrix connecting the current cell state to the output gate; Indicates the output gate bias term; express Activation function; This represents element-wise multiplication. This represents the hyperbolic tangent activation function.
6. The lightweight modulation recognition method based on noise perception and multi-task learning according to claim 5, characterized in that, The formula for obtaining time series data by preprocessing the IQ signal in S1 is: In the formula: and These represent the minimum and maximum values in the original signal sample, i.e., the IQ signal; The first one represents the original signal sample. each sample value and ; Indicates the number of samples; This represents the normalized IQ signal; And record the time series data as .
7. A system based on the lightweight modulation recognition method based on noise perception and multi-task learning as described in any one of claims 1 to 6, characterized in that, It includes a signal modulation module, a data preprocessing module, a model building module, a model training module, and a model deployment module; The signal modulation module is used to perform different types of signal modulation on the original communication signal based on a preset signal modulation method to obtain IQ signals; The data preprocessing module is used to preprocess the IQ signal to obtain time series data; The model building module is used to build a signal modulation mode recognition model based on noise perception and multi-task learning; The model training module is used to construct a loss function and, based on an adversarial training strategy and a phased fine-tuning strategy, train the signal modulation mode recognition model according to time series data to obtain a lightweight signal modulation recognition model. The model deployment module is used to deploy the lightweight signal modulation recognition model to a pre-set device or server to achieve the recognition of signal modulation mode.
Citation Information
Patent Citations
Blind modulation identification method and device under the condition of unknown noise level based on joint learning
CN112910811A
Low signal-to-noise ratio OTFS signal identification and classification method and identification and classification system
CN120086647A
