A complex electromagnetic signal demodulation method based on multi-task learning

Abstract

The application discloses a complex electromagnetic signal demodulation method based on multi-task learning, relates to the technical field of electromagnetic signal demodulation, and comprises the following steps: a hybrid architecture of ResNet-18 and Transformer is adopted as a backbone network, a multi-task learning model is constructed in combination with a dynamic weight distribution mechanism; and a to-be-measured complex electromagnetic signal is input into the multi-task learning model to obtain a modulation type identification result and a symbol sequence prediction result, the collaborative optimization of two tasks is realized, and a negative transfer effect caused by task conflict is effectively inhibited through the dynamic weight distribution mechanism. In the electromagnetic signal demodulation process, the modulation mode adopted by the signal is first judged, and then the original symbol sequence is predicted, the modulation type identification is the basis of the symbol sequence prediction, the accuracy of the modulation type identification directly influences the final result of the symbol sequence prediction, the multi-task learning model improves the accuracy of the modulation type identification through the collaborative optimization of the two tasks, and meanwhile, the accuracy of the symbol sequence generation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electromagnetic signal demodulation, in particular to a complex electromagnetic signal demodulation method based on multi-task learning. BACKGROUND

[0002] Electromagnetic signals are information carriers that propagate through electromagnetic waves and are widely used in communication, radar, navigation, remote sensing, and other fields. Modulation and demodulation of electromagnetic signals are key technologies for information transmission and reception. Traditional demodulation techniques are mainly based on the frequency domain or time domain characteristics of signals, and are implemented through hardware devices such as filters and demodulators. However, with the increasing complexity of the electromagnetic environment, traditional demodulation techniques have obvious limitations in dealing with noise interference, multipath propagation, and signal overlap. In recent years, artificial intelligence and machine learning techniques have been widely applied in the field of electromagnetic signal demodulation. Deep learning-based demodulation methods, such as convolutional neural networks and recurrent neural networks, can effectively extract signal features and improve the accuracy and robustness of demodulation. However, existing deep learning methods still have limitations in dealing with signal overlap, noise interference, and multipath propagation in complex electromagnetic environments, and are mostly focused on single tasks, lacking joint optimization of multiple tasks, which makes it difficult for the model to meet the needs of multiple tasks simultaneously, and the conflict and dependency between tasks are not fully utilized. In addition, existing methods usually use padding or truncation when dealing with variable-length sequence data, which not only increases the computational overhead but also may cause position encoding offset, affecting the performance of the model.

[0003] Therefore, a complex electromagnetic signal demodulation method based on multi-task learning is provided to solve the above problems by introducing the conflict and dependency between multiple tasks to improve the accuracy of prediction. SUMMARY

[0004] To solve the above problems, the present application provides a complex electromagnetic signal demodulation method based on multi-task learning, which realizes the collaborative optimization of modulation type classification and code sequence generation by constructing a multi-task learning framework with ResNet-18 and Transformer as the backbone, effectively suppresses the negative transfer effect caused by task conflict through a dynamic weight distribution mechanism, and improves the accuracy of modulation type recognition and the accuracy of code sequence generation through collaborative optimization.

[0005] To achieve the above purpose, the present application provides a complex electromagnetic signal demodulation method based on multi-task learning, comprising the following steps:

[0006] S1: Obtain signal data sets under different modulation modes, and divide the signal data sets under different modulation modes into training sets and validation sets;

[0007] S2: adopting a hybrid architecture of ResNet-18 and Transformer as a backbone network, combining a dynamic weight distribution mechanism to construct a multi-task learning model;

[0008] S3: training the multi-task learning model by using the training set and the validation set to obtain a trained multi-task learning model;

[0009] S4: inputting the complex electromagnetic signal to be tested into the trained multi-task learning model to obtain a modulation type recognition result and a symbol sequence prediction result.

[0010] Preferably, S1 further comprises pre-processing the signal data in the data set, the pre-processing comprising: removing noise and clutter in the signal data by using a band-pass filtering method, and visualizing the denoising result by constellation diagram.

[0011] Preferably, S2 specifically comprises:

[0012] The backbone network adopts a hybrid architecture of ResNet-18 and Transformer, the ResNet-18 branch is used to extract local time sequence features and modulation type discriminative patterns, and the Transformer branch captures long-term spectral correlation and modulation parameter evolution law through a multi-head self-attention mechanism;

[0013] The dynamic weight distribution mechanism is used to adaptively adjust the loss weights of the modulation type recognition task and the symbol sequence prediction task;

[0014] A cross-branch feature interaction gate unit is set to dynamically adjust the fusion ratio of ResNet-18 deep features and Transformer sequence features.

[0015] Preferably, the dynamic weight distribution mechanism in S2 specifically comprises a dynamic EMA weight balancing mechanism, a sequence alignment and bit error rate integration, and a composite loss quantification system;

[0016] The dynamic EMA weight balancing mechanism is used to dynamically adjust the loss weight of each task by using an exponential moving average (EMA) algorithm;

[0017] The sequence alignment and bit error rate integration are used to use a dynamic sequence alignment algorithm to realize length matching of the predicted sequence and the target sequence through permute operation and zero padding, and embed the bit error rate as a soft constraint into the loss function;

[0018] The composite loss quantification system is used to weight the modulation type classification loss, the sequence generation probability distribution loss, and the bit error rate loss scaled by a temperature coefficient to form a total loss function.

[0019] Preferably, the exponential moving average (EMA) algorithm is expressed as:

[0020] ;

[0021] wherein, is a decay factor, is a loss weight of the t-th time step, is a loss weight of the t-1-th time step, is an original loss value.

[0022] Preferably, the bit error rate is embedded into the loss function as a soft constraint, which is expressed as:

[0023] ;

[0024] wherein, is an indicator function, is a number of valid bits, is a target sequence, is a predicted sequence.

[0025] Preferably, the total loss function adopts a hierarchical multi-task joint modeling framework to decouple the modulation type classification and the sequence generation probability distribution into two parallel sub-tasks; the main path adopts a cross-entropy loss to quantify the modulation type classification error, and the loss weight is adaptively adjusted through a dynamic EMA mechanism; the auxiliary path optimizes the sequence generation probability distribution through a cross-entropy loss and monitors the performance of the model through the bit error rate calculation.

[0026] Preferably, the total loss function is expressed as:

[0027] ;

[0028] wherein, is a modulation type classification loss weight, is a sequence generation probability distribution loss weight, is a temperature coefficient scaled bit error rate loss weight, is a modulation type classification loss, is a sequence generation probability distribution loss, is a temperature coefficient scaled bit error rate loss.

[0029] Preferably, the model training adopts a dynamic batch construction strategy, an optimizer setting, and a multi-dimensional stability enhancement mechanism;

[0030] The dynamic batch construction strategy is used to ensure the consistency of the sequence length within each batch through width clustering and dynamic batch construction;

[0031] The optimizer setting is used to adopt the AdamW optimizer algorithm combined with the cosine annealing learning rate scheduler to balance the momentum update and the adaptive learning rate correction.

[0032] Preferably, the multi-dimensional stability enhancement mechanism includes an adversarial data augmentation strategy, a gradient clipping technique, and a model checkpoint saving strategy.

[0033] The adversarial data augmentation strategy makes the model learn noise-invariant features by adding Gaussian white noise and transient interference pulses at the input end;

[0034] The gradient clipping technique is used to limit the risk of gradient explosion, and the maximum norm constraint is set to 1.0. The joint gradient of the cross-branch feature interaction gating unit is normalized by the clip_grad_norm function;

[0035] The early stopping mechanism and the model checkpoint saving strategy are used to terminate training when the weighted total loss on the validation set does not decrease for 15 consecutive cycles, and the optimal parameter state is preserved.

[0036] Therefore, the above-mentioned complex electromagnetic signal demodulation method based on multi-task learning is adopted, and the modulation mode of the signal is first judged in the electromagnetic signal demodulation process, and then the original symbol sequence is predicted. It can be seen that modulation type identification is the basis for symbol sequence prediction, and the accuracy of modulation type identification directly affects the final result of symbol sequence prediction. Based on this, the multi-task learning model designed by the application improves the accuracy of modulation type identification through the cooperative optimization of the two tasks, and improves the accuracy of symbol sequence generation. The local timing features and modulation type discriminative patterns of the signal are extracted through the ResNet-18 network, and the long-range spectral correlation and modulation parameter evolution law are captured by using the Transformer architecture, forming a dual-task branch of modulation classification and symbol sequence generation. In order to solve the common feature fragmentation problem in multi-task learning, a cross-branch feature interaction gating unit is innovatively designed, which dynamically adjusts the fusion ratio of ResNet-18 deep features and Transformer sequence features, realizes the implicit cooperative optimization of modulation feature learning and sequence generation task. Further, the dynamic weight distribution mechanism effectively suppresses the negative transfer effect caused by task conflict, thereby further improving the adaptability and robustness of the model in complex environments.

[0037] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A flowchart of a complex electromagnetic signal demodulation method based on multi-task learning in the application;

[0039] Figure 2 A model framework diagram in an embodiment of the application;

[0040] Figure 3 An I signal time domain graph in an embodiment of the application;

[0041] Figure 4 Q signal time domain graph in the embodiment of the application;

[0042] Figure 5 I signal time domain graph after filtering in the embodiment of the application;

[0043] Figure 6 Q signal time domain graph after filtering in the embodiment of the application;

[0044] Figure 7 Standard QPSK constellation graph in the embodiment of the application;

[0045] Figure 8 Constellation graph of the I / Q signal after filtering in the embodiment of the application;

[0046] Figure 9 Loss function change line graph with training round number in the embodiment of the application, wherein (a) is a training set loss change line graph, and (b) is a validation set loss change line graph;

[0047] Figure 10 Classification result confusion matrix graph in the embodiment of the application. DETAILED DESCRIPTION

[0048] The following detailed description of embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the application.

[0049] EMBODIMENT

[0050] A complex electromagnetic signal demodulation method based on multi-task learning, as shown in Figures 1-2 includes the following steps:

[0051] S1: Obtain signal data sets under different modulation modes, and divide the signal data sets under different modulation modes into a training set and a validation set;

[0052] The data set is derived from the third "Electromagnetic Big Data Challenge", and the data set has 10 known modulation signal sequence samples, and the ID mapping relationship is as follows: 1: BPSK, 2: QPSK, 3: 8PSK, 4: MSK, 5: 8QAM, 6: 16-QAM, 7: 32-QAM, 8: 8-APSK, 9: 16-APSK, 10: 32-APSK, the signal generation process covers different signal lengths, a wide range of pulse widths, different symbol widths, and simulates different signal-to-noise ratio SNR conditions and multipath, fading and other interference environments, and the answer of the sequence sample is the modulation type, the symbol width and the modulation code sequence.

[0053] The sampling rate of the training data setting signal is 20 MHz, and there are 10 folders in total, each folder has 1.8W different signal length, symbol width, SNR and multipath sample data, and the data format is csv file, wherein the first two columns are simulation IQ waveform, the third column is simulation code sequence, the fourth column is simulation modulation type, and the fifth column is simulation symbol width. The data sample format is shown in Table 1.

[0054] Table 1 Data sample format

[0055] ;

[0056] The data set simulates a real wireless environment, and the data classification difficulties are as follows: (1) There are many modulation methods, including various common digital modulation methods; (2) The symbol width and symbol sequence length are not fixed, resulting in large differences in I / Q signal sequence length; (3) The electromagnetic data constellation diagram difference is not obvious, in order to increase the classification difficulty, the multipath, fading and other interference environment are added, and the noise and other actual influencing factors are added. The above three factors bring certain difficulty to modulation identification, and are also the problems that the field tries to solve.

[0057] Due to the limitation of computing power, only the data sets corresponding to the modulation methods of 1:BPSK, 2:QPSK, 3:8PSK, 4:MSK and 5:8QAM in the data set are taken for training in the embodiment. Meanwhile, the I / Q signal sequence and the symbol sequence are divided according to a certain length, so as to reduce the length of the sequence. The training set and the verification set are divided according to the ratio of 8:2.

[0058] S1 also includes pre-processing the signal data in the data set, and the pre-processing includes: removing noise and clutter in the signal data by using a band-pass filtering method, and visualizing the denoising result by constellation diagram.

[0059] First, the data is pre-processed by using a traditional filtering method to remove noise, clutter and other influencing factors in the original I / Q signal data, and the denoising result is visualized by constellation diagram.

[0060] Take a data modulated by QPSK as an example. QPSK (Quadrature Phase Shift Keying) is a digital modulation method that transmits data by changing the phase of the carrier signal. QPSK carries two bits of data on each symbol, so its spectrum utilization efficiency is higher, and it is a common modulation technology, especially widely used in wireless communication. Its carrier phase has four possible values, corresponding to 00, 01, 10 and 11 four bit combinations.

[0061] From the constellation diagram of the original I / Q signal data as shown in FIG. 1, it can be seen that the original I / Q signal data is not very clean, and there are some noise and clutter in the data. Figures 3-4The I / Q signal time domain diagram shown can see that the I / Q signal waveform fluctuation is very obvious, which shows that the original signal sequence contains high noise.

[0062] First, the data is simply processed by using a band-pass filter. A band-pass filter is a filter that only allows intermediate frequencies to pass through. It filters out high and low frequency signals above and below a certain range by retaining only signals with frequencies within that range, so that the passing signal only contains intermediate frequency components. The frequency response characteristic of such a filter is relatively flat in the middle frequency band, and rapidly decays in the high and low frequency bands. The lower cutoff frequency of the band-pass filter is set to 10 MHz, the upper cutoff frequency is set to 50 MHz, and the filter order is set to 100. The filtered I / Q signal time domain diagram is shown in Figures 5-6 As can be seen from the figure, the band-pass filter effectively filters out some noise in the original signal sequence, making the signal sequence more orderly, but some noise is still not filtered out.

[0063] The constellation diagram of the filtered I / Q signal is shown in Figure 8 The signal sequence contains a lot of noise. The constellation diagram of the standard QPSK modulated signal is shown in Figure 7 All signal points converge around four points; while in the constellation diagram drawn according to the data set, the signal points are scattered and there is no clear convergence center, indicating that there is a large amount of noise data. Therefore, it can be seen that ordinary filtering methods are difficult to directly process the noise in the data set, and deep learning and other methods need to be used to classify the data set.

[0064] S2: A hybrid architecture of ResNet-18 and Transformer is used as the backbone network, combined with a dynamic weight distribution mechanism, to build a multi-task learning model;

[0065] In the electromagnetic signal demodulation process, the modulation method used by the signal needs to be determined first, and then the original symbol sequence is predicted. Therefore, modulation type identification is the basis for symbol sequence prediction, and the accuracy of modulation type identification directly affects the final result of symbol sequence prediction. Based on this, a multi-task learning model is designed in this embodiment, which improves the accuracy of modulation type identification through the cooperative optimization of two tasks, and improves the accuracy of symbol sequence generation.

[0066] Specifically, the backbone network adopts a hybrid architecture of ResNet-18 and Transformer, the ResNet-18 branch is used to extract local time sequence features and modulation type discriminative patterns, and the Transformer branch captures long-term spectral correlation and modulation parameter evolution law through multi-head self-attention mechanism.

[0067] ResNet-18 is composed of 4 residual block groups, each containing a different number of residual blocks, each containing multiple convolutional layers, batch normalization layers, and activation functions. For example, in a common two-layer residual block, the input x is first passed through a convolutional layer, then batch normalized and activated with ReLU, followed by a second convolutional layer and batch normalization, resulting in a residual F(x). Finally, the residual F(x) is added to the input x, and activated with ReLU to obtain the output of the residual block. When the dimensions of the input and residual are not consistent, an additional convolutional layer (usually 1x1 convolution) is needed to adjust the dimensions of the input x so that it can be added. ResNet-18, as a deep neural network model, has strong feature extraction and classification capabilities. In practical applications, the network structure can be optimized by adjusting the number of residual blocks, the parameters of convolutional layers, etc. to improve the performance of the model. In addition, other techniques such as data augmentation, regularization, etc. can be combined to further improve the generalization ability of the model, which is very suitable for feature extraction of I / Q signal sequences.

[0068] Transformer is a sequential data modeling architecture with long-distance feature capture capability, feature extraction capability, and efficient parallel computing capability. The Transformer architecture introduces an attention mechanism, allowing the model to consider all positions in the input sequence simultaneously, rather than processing them sequentially. The self-attention mechanism allows the model to weight each position based on the relationship between different positions in the input sequence, capturing global context information and being widely used in sequence-to-sequence tasks. The self-attention mechanism generates three vector representations of query (Query), key (Key), and value (Value) for each element in the input sequence, constructs a similarity measurement matrix across positions, and then dynamically integrates context information through weighted aggregation.

[0069] Mathematically, given an input sequence where is the sequence length, is the feature dimension, the self-attention mechanism first generates a query matrix , a key matrix and a value matrix , where , , are learnable parameter matrices, and are the dimensions of the query and key, and the value, respectively. Then, the scaled dot-product attention score of the query and the key is calculated:

[0070] ;

[0071] where The softmax function normalizes scores into probability distribution, implementing soft selection of attention weights. This mechanism explicitly models the similarity of all position pairs in the sequence through matrix operation, and then generates the context-aware representation of each position through weighted summation.

[0072] To enhance the model's representation ability of multi-granularity information, the Transformer introduces multi-head attention mechanism, i.e., splitting the input features into independent subspaces, performing the above attention calculation in each subspace in parallel, then concatenating the results and fusing them through linear transformation:

[0073] (3.3);

[0074] where , is the output projection matrix. The multi-head mechanism focuses on different feature dimensions of the input sequence through different subspaces, significantly improving the model's expressiveness and robustness. The advantage of self-attention mechanism lies in its parallelization and global dependency modeling capability. Self-attention allows simultaneous processing of full sequence information in a single forward propagation, and its computational complexity grows quadratically with sequence length, but can be optimized to linear complexity through sparse attention patterns. In addition, the self-attention mechanism has position awareness, which injects sequence position information into the input vector by adding position encoding, enabling the model to distinguish the context relationships of different positions and capture the mutual relationships between sampling data at different times in the I / Q signal sequence.

[0075] The dynamic weight distribution mechanism is used to adaptively adjust the loss weights of the modulation type recognition task and the symbol sequence prediction task; the dynamic weight distribution mechanism specifically includes a dynamic EMA weight balancing mechanism, sequence alignment and bit error rate integration, and a composite loss quantification system;

[0076] The dynamic EMA weight balancing mechanism is used to dynamically adjust the loss weight of each task through the exponential moving average (EMA) algorithm;

[0077] To address the negative transfer problem in multi-task learning, a dynamic weight distribution system based on exponential moving average (EMA) is designed. Three learnable normalization factors mod_loss_scale, seq1_loss_scale, and seq2_loss_scale are fused through the EMA algorithm to integrate historical loss statistics:

[0078] ;

[0079] where is the decay factor, which can be 0.1, is the loss weight of the t-th time step, is the loss weight of the t-1-th time step, is the original loss value. This mechanism enables the loss weight to adapt to the data distribution changes of different tasks and the characteristics of the training stage. Compared with static weight setting, EMA dynamic adjustment improves the sensitivity of task weight adjustment of the model in the dynamic electromagnetic interference scene, effectively suppressing the gradient cancellation phenomenon caused by task conflict.

[0080] Sequence alignment and bit error rate integration is used to utilize the dynamic sequence alignment algorithm to realize the length matching of the predicted sequence and the target sequence through permute operation and zero padding, and to embed the bit error rate as a soft constraint into the loss function;

[0081] For the variable-length sequence generation task, a dynamic sequence alignment algorithm is designed to realize the length matching of the predicted sequence and the target sequence through permute operation and zero padding, while minimizing the interpolation error while maintaining the time sequence continuity. Further innovatively, the bit error rate (BER) is embedded as a soft constraint into the loss function:

[0082] ;

[0083] wherein, is the indicator function, is the number of valid bits (filter padding value), is the target sequence, is the predicted sequence. This design enables the model to optimize the sequence generation probability distribution while directly aligning the bit-level performance indicators, solving the scale difference problem between traditional cross-entropy loss and actual bit error rate.

[0084] Composite loss quantification system is used to weight the modulation type classification loss, sequence generation probability distribution loss, and temperature coefficient scaled bit error rate loss to form the total loss function.

[0085] The total loss function is represented as:

[0086] ;

[0087] wherein, is the modulation type classification loss weight, is the sequence generation probability distribution loss weight, is the temperature coefficient scaled bit error rate loss weight, is the modulation type classification loss, is the sequence generation probability distribution loss, is the temperature coefficient scaled bit error rate loss.

[0088] The loss function adopts a hierarchical multi-task joint modeling framework to decouple the modulation type classification (discrete label prediction) and the simulation code sequence reconstruction (continuous sequence generation) into two parallel sub-tasks. The main path adopts CrossEntropyLoss to quantify the modulation classification error, and its loss weight is adaptively adjusted by a dynamic EMA mechanism. The auxiliary path innovatively introduces a double constraint mechanism: on the one hand, it directly optimizes the sequence generation probability distribution through CrossEntropyLoss, and on the other hand, it realizes bit-level performance monitoring through BER calculation. This triple loss architecture breaks through the limitations of traditional single-objective optimization and forms an implicit collaboration between modulation feature learning and sequence generation tasks.

[0089] A cross-branch feature interaction gating unit is set up to dynamically adjust the fusion ratio of ResNet-18 deep features and Transformer sequence features.

[0090] To solve the feature fragmentation problem in multi-task learning, a cross-branch feature interaction gating unit is designed to dynamically adjust the fusion ratio of ResNet-18 deep features and Transformer sequence features, realizing implicit collaborative optimization of modulation classification tasks and code sequence generation tasks.

[0091] This embodiment constructs a dual-flow multi-task network based on cross-modal feature interaction and dynamic task coordination. Based on the shared I / Q orthogonal signal time-frequency feature space, a hierarchical multi-task collaborative mechanism is adopted. The input signal is processed by a dual-channel preprocessing module to generate time-domain waveform features and frequency-domain features, and the cross-modal feature fusion network is used to realize the complementary of time-frequency domain information.

[0092] When predicting the length of the symbol sequence, the symbol width and the sampling rate are regarded as known constants, and the product of the two is used as the length of the token to divide the I / Q signal sequence, forming a [batch_size, seq_len, token_len, 2] dimensional dataset, and the long sequence prediction task is converted into a task of predicting one symbol with one token. Then the information of modulation type is added to learn the different mapping methods corresponding to different modulation types. And through the transformer, the relationship between tokens is learned to capture the mutual relationship between symbols, reduce the influence of noise, and improve the accuracy of symbol sequence prediction, realizing the dynamic prediction of the length of the symbol sequence.

[0093] Multi-task learning improves the system's generalization performance and robustness on a single task by constructing a shared parameterized model architecture and achieving knowledge transfer and representation learning during joint training of multiple related tasks. The theoretical basis of this method can be traced back to the joint risk minimization principle in statistical learning theory, which involves jointly optimizing the loss functions of multiple tasks to mitigate overfitting risk by leveraging the inherent correlation between tasks, while simultaneously mining potential shared feature representations in the data. Mathematically, multi-task learning can be formulated as a multi-objective optimization problem, where the objective function is typically a weighted combination of the loss functions of each task, i.e.:

[0094] ;

[0095] in Indicates shared parameters. These are private parameters for task t. and and represent the task weights and regularization coefficients, respectively, with R being the model complexity constraint. This parametric design allows the model to capture common features across tasks through shared layers, while adapting to the differentiated needs of each task through task-specific layers.

[0096] The core of multi-task learning lies in parameter sharing strategies, loss function design, and model architecture innovation. Parameter sharing strategies encompass two paradigms: hard sharing and soft sharing. Hard sharing extracts shared features through the first few layers of a unified neural network, followed by a task-specific output layer. Soft sharing, on the other hand, achieves implicit feature interaction through parameterized latent variable models (such as multi-task Gaussian processes or variational autoencoders), which has the advantage of avoiding negative transfer caused by task conflicts. Loss function design focuses on the dynamic adjustment mechanism of task weights. Classic methods include uncertainty weighting and gradient normalization. The former optimizes weights by minimizing the weighted variance of uncertainty across tasks, while the latter avoids the problem of gradient suppression of secondary tasks by equalizing gradient magnitudes. Model architecture innovation is reflected in heterogeneous network design and dynamic network routing. These methods improve collaborative learning efficiency by enhancing the information exchange paths between tasks.

[0097] The multi-task learning framework effectively facilitates knowledge transfer between modulation type identification and code sequence generation tasks by sharing the time-frequency features of I / Q signals. The local time-frequency features extracted by the ResNet-18 branch provide modulation type-related contextual information for the Transformer branch, while the Transformer's long-range dependency modeling capability, in turn, enhances the temporal correlation analysis of the modulation classification task. This bidirectional feature enhancement mechanism makes the model more robust to complex electromagnetic interference.

[0098] S3: Train the multi-task learning model using the training set and validation set to obtain a trained multi-task learning model;

[0099] The model training adopts a dynamic batch construction strategy, an optimizer setting, and a multi-dimensional stability enhancement mechanism.

[0100] In view of the variable length characteristic of the I / Q signal sequence, a dynamic batch construction strategy is implemented in the data loading stage. In the preprocessing stage, the original signal is clustered by width, and samples with the same signal width (i.e., the same token division length) are allocated to the same batch to ensure the consistency of the sequence length in each batch, thereby avoiding the additional computational overhead and position encoding offset problem caused by dynamic padding. This strategy can improve the GPU memory utilization while maintaining the global modeling effectiveness of the self-attention mechanism of the Transformer branch.

[0101] The model training adopts AdamW optimizer as the core optimization algorithm, and the parameter setting follows the characteristics of the mixed task. The initial learning rate is set to 3×10 −4 , the weight decay coefficient is configured to 1×10 −4 , and the beta parameter is set to (0.9, 0.999) to balance the momentum update and adaptive learning rate correction. To cope with the gradient scale difference of multi-task learning, a cosine annealing learning rate scheduler is introduced, and the cycle length is set to 50 epochs. The learning rate is automatically reduced to 10% of the initial value in the later training stage, which effectively alleviates the oscillation phenomenon of the modulation classification and sequence generation tasks in the parameter space update.

[0102] Multi-dimensional stability enhancement mechanism: First, for the noise robustness cultivation in complex electromagnetic environment, an adversarial data augmentation strategy is implemented, and Gaussian white noise and transient interference pulse are added at the input end to force the model to learn noise invariant features. Second, the gradient clipping technique is used to limit the risk of gradient explosion, and the maximum norm constraint is set to 1.0. The joint gradient of the cross-modal feature interaction gating unit is normalized by the clip_grad_norm function. Finally, an early stopping mechanism and model checkpoint saving strategy are designed. When the weighted total loss on the validation set does not decrease for 15 consecutive epochs, the training is terminated, and the optimal parameter state is preserved.

[0103] S4: input the complex electromagnetic signal to be tested into the trained multi-task learning model to obtain the modulation type recognition result and the symbol sequence prediction result.

[0104] Embodiment 1

[0105] In this embodiment, classification accuracy, bit error rate, and cosine similarity are used as evaluation indicators.

[0106] Classification accuracy is a commonly used evaluation metric in machine learning and statistical classification tasks, used to measure the consistency between the model's predictions and the true labels. Its core definition is the ratio of the number of correctly predicted samples to the total number of samples. Assuming a dataset containing N samples, the model's prediction for each sample is... The real label is The accuracy rate is then expressed as:

[0107] ;

[0108] in, It is an indicator function that takes the value 1 when the condition is met, and 0 otherwise.

[0109] Bit Error Rate (BER) is a core metric used in communication systems and digital transmission to measure the reliability of data transmission. It represents the proportion of erroneous bits detected by the receiver out of the total number of transmitted bits. The BER is defined as:

[0110] ;

[0111] For example, if transmission If 100 bits of the bit data are corrupted, then... (That is, 1 bit error per billion bits).

[0112] Cosine similarity measures the directional similarity between two vectors by calculating the cosine of the angle between them in space. It is expressed as:

[0113] ;

[0114] in, For vector dot product, and Let be the magnitude of the vector. The range is [-1, 1]. The closer to 1, the more consistent the directions; the closer to 0, the more orthogonal; and the closer to -1, the opposite the directions.

[0115] In the modulation signal recognition and symbol sequence prediction tasks, this embodiment compares the performance of three deep learning architectures on the two sub-tasks of modulation type classification and symbol sequence prediction. The results are shown in Table 2:

[0116] Table 2

[0117] ;

[0118] The experimental results show that the Resnet18-Transformer hybrid model fusing convolutional neural network and self-attention mechanism achieves a classification accuracy of 56.51% in the modulation type classification task, which is 2.71 percentage points higher than the Resnet18 benchmark model with pure convolutional architecture, and also shows better symbol sequence prediction capability, with a bit error rate 3.53% lower (a relative decrease of 6.87%) than the Transformer benchmark model and a cosine similarity index of 0.072 (a relative increase of 11.57%). It is worth noting that although the pure Transformer architecture performs well in the sequence similarity measurement of symbol prediction (cosine similarity of 0.6225), its classification accuracy decreases by 24.55 percentage points compared with the hybrid model, reflecting the limitations of single attention mechanism in local feature extraction. The traditional Resnet18 model performs second in the classification task, with a bit error rate of symbol prediction 21.83% higher than the hybrid model, indicating that the pure convolutional architecture has shortcomings in time series dependency modeling. The above comparisons verify the advantages of the hybrid architecture in the joint task, which effectively captures the local spectral features through convolutional layers and establishes long-time dependency relationships through self-attention mechanism, achieving a synergistic improvement in classification accuracy and sequence prediction performance.

[0119] Embodiment 2

[0120] In the ablation experiment for the multi-task joint optimization framework, this embodiment analyzes the synergistic mechanism in the modulation type classification and symbol sequence prediction tasks through a systematic module removal strategy. The experimental design includes three key configurations: simultaneous execution of dual tasks, only focusing on the classification task, and only focusing on the sequence prediction task, and the experimental results are shown in Table 3.

[0121] Table 3 Experimental results

[0122] ;

[0123] The model shows significant advantages in the multi-task cooperative training mode, with a modulation type classification accuracy of 56.51%, which is 2.71 percentage points higher than only performing the classification task; the bit error rate of symbol sequence prediction is as low as 0.4960, and the cosine similarity is 0.6943, which is 3.53% and 11.57% lower and higher than the only sequence prediction task, respectively.

[0124] The experiment verifies the superiority of the multi-task learning framework from the perspective of task coupling. It achieves feature reuse through parameter sharing mechanism, maintains performance advantage in classification task, and achieves greater similarity improvement in prediction task compared with single task model. It is worth noting that the single task framework does not reach the joint optimization level of multi-task model in dedicated task, indicating that the synergy of multi-task not only does not cause negative transfer between tasks, but also enhances the generalization representation ability of the model through feature interaction.

[0125] Embodiment 3

[0126] The preprocessed signal data is input into the multi-task model built, and the loss function is plotted against the number of training rounds, as shown in Figure 9 The loss function value shows a clear trend of change with the increase of training rounds. In the early stage of training, the loss function value decreases rapidly, and the model is rapidly learning the basic features and patterns in the data, and the prediction error is significantly reduced. The rapid decline in this stage reflects the preliminary adaptation of the model to the training data and the formation of feature extraction ability.

[0127] As the training goes deeper, the decline speed of the loss function value gradually slows down and starts to fluctuate. This fluctuation is due to the model starting to learn more complex features, which may be affected by noise, uneven data distribution and other factors. Under the framework of multi-task learning, the model needs to optimize the modulation classification and symbol sequence generation tasks simultaneously, increasing the complexity of training. The feature fragmentation and task conflict between different tasks may cause the loss function value to fluctuate. In addition, the adjustment of the dynamic EMA weight allocation mechanism during the training process may also affect the loss function value, so that the model dynamically allocates resources between different tasks to seek the overall optimum.

[0128] In the later stage of training, the loss function value tends to be stable and gradually converges to a smaller value. This indicates that the model has learned most of the useful information in the data and can achieve good performance on the validation or test set. The achievement of the convergence state means that the model has reached a relatively stable state in the training process and can better generalize to new data.

[0129] The variation in loss function values ​​is closely related to the model architecture and dynamic weight allocation mechanism. The joint demodulation model based on multi-task learning optimizes both modulation classification and symbol sequence generation tasks simultaneously through the design of a composite loss function. This design requires the model to balance the learning progress of different tasks during training, thus increasing the training difficulty. However, the introduction of the dynamic EMA weight allocation mechanism allows the model to adaptively adjust the loss weights for different tasks, ensuring that the model always focuses on the task most in need of optimization during training. This mechanism helps the model better balance the feature fragmentation and task conflict problems in multi-task learning, thereby improving overall performance.

[0130] Plot a confusion matrix on the classification results, such as Figure 10 As shown in the figure, the algorithm in this study achieves high accuracy in recognizing BPSK and MSK, but it easily identifies other types as BPSK, indicating that the algorithm cannot effectively distinguish the relationship between other modulation types and BPSK. PSK in BPSK indicates the use of phase-shift keying (PSK), a form of phase modulation used to represent a series of discrete states. BPSK corresponds to 2 states, QPSK to 4 states, and 8PSK to 8 states. Therefore, BPSK, QPSK, and 8PSK are closely related, causing the model to fail to distinguish them well. Thus, further research is needed to improve the model's accuracy in differentiating different PSK modulation methods.

[0131] Therefore, the application adopts the above-mentioned complex electromagnetic signal demodulation method based on multi-task learning, extracts local time sequence features and modulation type discriminative patterns of the signal through a ResNet-18 network, captures long-range spectral correlation and modulation parameter evolution law by using a Transformer architecture, and forms dual-task branches of modulation classification and code sequence generation. In solving the common feature fragmentation problem in multi-task learning, the cross-branch feature interaction gate unit is designed innovatively to dynamically adjust the fusion ratio of ResNet deep features and Transformer sequence features, realizing the implicit collaborative optimization of modulation feature learning and sequence generation tasks. In the dynamic weight distribution mechanism, the dynamic weight balance system based on the exponential moving average (EMA) is proposed, which realizes the adaptive adjustment of task weight by fusing historical loss statistics. Compared with the static weight setting, the EMA mechanism improves the sensitivity of task weight adjustment of the model in the dynamic electromagnetic interference scene, effectively inhibiting the gradient cancellation phenomenon caused by task conflict. For the variable-length code sequence generation task, a dynamic sequence alignment algorithm and bit error rate integration strategy are designed, which directly associates the traditional cross-entropy loss with the bit-level performance index, so that the model can optimize the sequence generation probability distribution while realizing the explicit constraint of the bit error rate. The embodiment designs a multi-task learning model, which improves the accuracy of modulation type recognition and the accuracy of code sequence generation through the collaborative optimization of dual tasks.

[0132] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for demodulating complex electromagnetic signals based on multi-task learning, characterized in that: The method comprises the following steps: S1: obtaining signal data sets under different modulation modes, and dividing the signal data sets under different modulation modes into a training set and a validation set; S2: adopting a hybrid architecture of ResNet-18 and Transformer as a backbone network, combining a dynamic weight distribution mechanism to construct a multi-task learning model; specifically comprising: The backbone network adopts a hybrid architecture of ResNet-18 and Transformer, the ResNet-18 branch is used for extracting local time sequence features and modulation type discriminative patterns, and the Transformer branch captures long-term spectral correlation and modulation parameter evolution law through a multi-head self-attention mechanism; The dynamic weight distribution mechanism is used for adaptively adjusting the loss weights of the modulation type identification task and the symbol sequence prediction task; A cross-branch feature interaction gate unit is set to dynamically adjust the fusion proportion of ResNet-18 deep features and Transformer sequence features; The dynamic weight distribution mechanism specifically comprises a dynamic EMA weight balancing mechanism, a sequence alignment and bit error rate integration, and a composite loss quantification system; The dynamic EMA weight balancing mechanism is used for dynamically adjusting the loss weight of each task through an exponential moving average (EMA) algorithm; The sequence alignment and bit error rate integration are used for utilizing a dynamic sequence alignment algorithm to realize length matching of the predicted sequence and the target sequence through permute operation and zero padding, and embedding the bit error rate as a soft constraint into the loss function; The composite loss quantification system is used for weighting the modulation type classification loss, the sequence generation probability distribution loss and the bit error rate loss scaled by a temperature coefficient to form a total loss function; S3: training the multi-task learning model using the training set and the validation set to obtain a trained multi-task learning model; S4: inputting a complex electromagnetic signal to be tested into the trained multi-task learning model to obtain a modulation type identification result and a symbol sequence prediction result.

2. The complex electromagnetic signal demodulation method based on multi-task learning according to claim 1, wherein: S1 further comprises pre-processing the signal data in the data set, and the pre-processing comprises: removing noise and clutter in the signal data by using a band-pass filter method, and visualizing the denoising result by using a constellation diagram.

3. The method of claim 1, wherein the method is based on multi-task learning. The exponential moving average (EMA) algorithm is represented as: ； wherein, is a decay factor, is a loss weight for the t-th time step, is a loss weight for the t-1-th time step, is an original loss value.

4. The complex electromagnetic signal demodulation method based on multi-task learning according to claim 3, characterized in that: Embedding the bit error rate as a soft constraint into the loss function is represented as: ； wherein, is an indicator function, is the number of valid bits, is the target sequence, is the prediction sequence.

5. The complex electromagnetic signal demodulation method based on multi-task learning according to claim 4, characterized in that: The total loss function adopts a hierarchical multi-task joint modeling framework, decouples modulation type classification and sequence generation probability distribution into two parallel sub-tasks; the main path adopts a cross-entropy loss to quantify modulation type classification error, and adaptively adjusts the loss weight through a dynamic EMA mechanism; the auxiliary path optimizes sequence generation probability distribution through a cross-entropy loss, and realizes performance monitoring of the model through bit error rate calculation.

6. The complex electromagnetic signal demodulation method based on multi-task learning according to claim 5, characterized in that, The total loss function is represented as: ； wherein, is a modulation type classification loss weight, is a sequence generation probability distribution loss weight, is a bit error rate loss weight scaled by a temperature coefficient, is a modulation type classification loss, is a sequence generation probability distribution loss, is a bit error rate loss scaled by a temperature coefficient.

7. The method of claim 6, wherein the method is based on multi-task learning. Model training adopts a dynamic batch construction strategy, an optimizer setting and a multi-dimensional stability enhancement mechanism; The dynamic batch construction strategy is used for making the sequence length consistent within each batch through width clustering and dynamic batch construction; The optimizer setting is used for adopting an AdamW optimizer algorithm, combining a cosine annealing learning rate scheduler to balance momentum update and adaptive learning rate correction.

8. The complex electromagnetic signal demodulation method based on multi-task learning according to claim 7, characterized in that: The multi-dimensional stability enhancement mechanism includes an adversarial data enhancement strategy, a gradient clipping technique, and a model checkpoint saving strategy; The adversarial data enhancement strategy adds Gaussian white noise and transient interference pulses at the input end to enable the model to learn noise-invariant features; The gradient clipping technique is used to limit the risk of gradient explosion, and the maximum norm constraint is set to 1.

0. The joint gradient of the cross-branch feature interaction gating unit is normalized by the clip_grad_norm function. The early stopping mechanism and the model checkpoint saving strategy are used to terminate training when the weighted total loss on the validation set does not decrease for 15 consecutive cycles, and the optimal parameter state is preserved.

Images