Automatic modulation identification method based on multi-scale features and improved encoder

Abstract

The invention discloses an automatic modulation identification method [1] [2] based on multi-scale features and an improved encoder, which comprises the following steps: eliminating phase offset influence through phase transformation in a data preprocessing stage, capturing local details and global structure features of a signal by using a multi-scale feature extraction module, reducing an information loss rate, and then, carrying out automatic modulation identification on the signal by using the multi-scale feature extraction module; according to the self-encoder and the encoder module, through time information retention, feature reconstruction and noise suppression, while the reconstruction loss is minimized, the feature reconstruction and recognition capability of a modulation signal is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, and in particular to an automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure. BACKGROUND

[0002] In the real world, the communication environment will be affected by various interferences and noises. Automatic modulation recognition (AMR) technology is widely used to identify the modulation mode of communication signals, as well as in the fields of signal monitoring, fault detection, and security protection. Therefore, to improve the stability and reliability of the communication system, AMR has received extensive attention. Traditional methods such as likelihood-based (LB) or hand-crafted feature-based (FB) recognition schemes rely on prior knowledge and have poor generalization. Although deep learning improves recognition performance, it still has the following defects: (1) Phase sensitivity problem: phase offset causes constellation distortion, reducing recognition accuracy; (2) Insufficient feature expression: single-scale convolution cannot balance local details and global structure; (3) Noise interference: feature discriminability decreases under low signal-to-noise ratio (SNR).

[0003] For example, the CNN-LSTM combination model does not solve the phase offset, and the CNN stacking structure is limited by sequential feature extraction. SUMMARY

[0004] The purpose of the present application is to solve the problems in the prior art and to provide an automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure.

[0005] In order to achieve the above purpose, the present application adopts the following technical solutions: An automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure, comprising: Step 1: converting the IQ format signal into AP format data through an amplitude and phase conversion formula, inputting the P component in the AP format data into a PT module for phase transformation, and inputting the P component and the A component in the AP format data into a multi-scale feature extraction module; Step 2: the multi-scale feature extraction module uses four multi-scale one-dimensional convolution modules to process the phase-transformed P component and A component in parallel to extract features; Step 3: inputting the extracted features in step 2 into a feature denoising and time series modeling module for feature optimization and recognition, performing Softmax classification on the optimized and recognized features, and outputting the results.

[0006] Further, the phase conversion formula is:

[0007]

[0008] wherein, , respectively represent the amplitude and phase components of the AP sample signal, , respectively represent the in-phase component and quadrature component of the IQ signal.

[0009] Further, the step of inputting the P component in the AP format data into the PT module for phase transformation comprises: Step 4: The phase shift parameter is obtained by a large number of sample training of the PT module through the flat layer and the full connection layer ; Step 5: The specific value of the anti-offset factor is obtained through the offset matrix ; Step 6: The specific value of the anti-offset factor is substituted into the AP format data phase offset elimination formula to calculate the P component after phase transformation. Further, the AP format data phase offset elimination formula is:

[0010]

[0011]

[0012] wherein, , represents the AP format data without phase offset, represents the first row parameter of the offset matrix, represents the second row parameter of the offset matrix, and is called an anti-offset factor, represents the coordinates after phase offset.

[0013] Further, the feature input feature denoising and time series modeling module comprises: a self-encoder and a GRU encoder. The self-encoder is used for learning and reconstructing features. The GRU encoder is used for extracting and compressing multi-scale features.

[0014] Further, the GRU encoder is composed of two decreasing CUDNNGRU layers and two Dropout layers, wherein the CUDNNGRU layer processes sequence data through a recurrent neural network structure, and the Dropout layer avoids over-reliance on a certain GRU unit in the training process by randomly shutting down part of the GRU units, effectively preventing overfitting.​

[0015] Further, the output after the Dropout layer processing is:

[0016] wherein, is the reserved probability, is the hidden state after Dropout, represents the mask.

[0017] Further, the mask generation formula is: .

[0018] Further, the autoencoder is composed of a decoder and a GRU encoder, wherein the decoder adopts a TimeDistributed Dense structure, and the two encapsulated Dense layers have 32 neurons respectively.

[0019] Compared with the prior art, the present application has the advantages that: 1. The present application eliminates phase offset interference through phase transformation (PT), reduces information loss through a multi-scale feature extraction module (MSFE), and suppresses noise through an autoencoder and encoder (AE) module. Under the synergistic action of multiple links, the algorithm can stably perform in different data sets and different signal-to-noise ratio environments, and has stronger anti-interference ability and environmental adaptability.

[0020] 2. The multi-scale feature extraction module (MSFE) used in the present application can simultaneously capture local details and global structural features of the signal, avoiding the information one-sidedness caused by single-scale extraction. At the same time, the autoencoder and encoder (AE) module can preserve time information and realize high-quality feature reconstruction, effectively reducing information loss and providing a more complete and accurate feature basis for subsequent recognition. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The present application proposes a PT module structure diagram in the automatic modulation recognition method based on multi-scale features and improved encoder.

[0022] Figure 2 The present application proposes a network structure diagram of the automatic modulation recognition method based on multi-scale features and improved encoder.

[0023] Figure 3 The present application proposes an autoencoder structure diagram in the automatic modulation recognition method based on multi-scale features and improved encoder.

[0024] Figure 4 The present application proposes a GRU encoder structure diagram in the automatic modulation recognition method based on multi-scale features and improved encoder. DETAILED DESCRIPTION

[0025] The application will be further explained in connection with the accompanying drawings.

[0026] The application provides an automatic modulation recognition method based on multi-scale features and an improved encoder, comprising: Step 1: converting the IQ format signal into AP format data through an amplitude and phase conversion formula, inputting the P component in the AP format data into a PT module for phase transformation, and inputting the P component and the A component in the AP format data into a multi-scale feature extraction module.

[0027] Step 2: the multi-scale feature extraction module adopts four multi-scale one-dimensional convolution modules to process the phase-transformed P component and A component in parallel to extract features.

[0028] Step 3: inputting the features extracted in step 2 into a feature denoising and time series modeling module for feature optimization and recognition, performing Softmax classification on the optimized and recognized features, and outputting the results.

[0029] In step 1, when the original IQ format data is input, the complexity of the two-dimensional complex value time series format data makes it difficult for a lightweight deep learning model (such as LSTM) to capture appropriate time series characteristics and correlations in the signal. When the AP format data is used as input, the recognition performance of the LSTM is greatly improved in the full SNR range.

[0030] IQ format data can be converted into amplitude and phase (AP) data by the following formula.

[0031] , (1) wherein, 、 represent the amplitude and phase components of the AP sample signal, 、 represent the in-phase component and the quadrature component of the IQ signal.

[0032] Since the signal is prone to phase shift during wireless transmission, the phase shift will change the position of the constellation point of the signal, thereby causing the receiving end to misjudge the signal modulation type. The phase shift in the IQ coordinate system can be represented as: (2) wherein, represents the coordinate after phase shift, represents the coordinate before phase shift, represents the offset matrix, which is represented by .

[0033] Combining formula (1), the AP format data eliminating the influence of phase offset can be expressed as: (3) (4) wherein, , represents the AP format data without phase offset, represents the offset matrix first row parameter, represents the offset matrix second row parameter, and is called the anti-offset factor.

[0034] As can be seen from formula (3) and (4), the A component is not affected by the phase offset, so it can be directly input into the network after format conversion. The P component will be affected by the phase offset, and can be eliminated by phase transformation (PT) before being input into the network. The specific structure is shown in Figure 1 .

[0035] The PT module is first trained by a large number of samples through a flatten layer and a fully connected layer to obtain the phase offset parameter , that is, the offset matrix , to obtain the specific value of the anti-offset factor , and then obtain the P component without the influence of the phase offset.

[0036] In step 2, the multi-scale feature extraction module (MSFE) module takes A and P dual-channel data as input, which can avoid the difference in amplitude or phase characteristics when taking AP format as single-channel input. The specific structure is shown in Figure 2 . Four multi-scale one-dimensional convolution modules are used to process the input data in parallel, which mainly solves the limitation problem of single-scale convolution kernel in feature extraction. The model can capture local details and global structure features of the signal at the same time, avoid feature loss, and enhance feature diversity. Multi-scale parallel processing also improves the robustness of the model to noise, distortion and channel interference, reduces information loss, and maintains high parameter efficiency through shared weights and parallel computing. This design enables the model to fully cover the multi-scale features of the signal, and improves the performance and robustness of the complex signal modulation recognition task.

[0037] The working principle of the multi-scale feature extraction module is as follows: For the input sequence , when the convolution kernel is , the one-dimensional convolution operation can be expressed as: (5) wherein, denotes a convolution kernel, denotes a convolution kernel size, denotes a sequence length, denotes an output.

[0038] The multi-scale one-dimensional convolution can be expressed as: (6) wherein, denotes a multi-scale one-dimensional convolution module output, denotes a number of one-dimensional convolutions.

[0039] In step 3, the feature denoising and time series modeling module is mainly composed of an autoencoder and encoders (AE), and its specific structure is shown in FIG. 2B. The autoencoder is responsible for feature learning and reconstruction, and the encoders are used for multi-scale feature extraction and compression. Figure 3 、 Figure 4

[0040] The GRU encoder is composed of two decreasing CUDNN GRU layers and two Dropout layers. The CUDNN GRU layer processes sequence data through a recurrent neural network structure, and its core idea is to control the flow of information through a gating mechanism. Its operation logic expression is: (7) (8) (9) (10) wherein, denotes a current unit input, denotes a current unit output, denotes a previous time step unit output, denotes an update gate, denotes a reset gate, is a weight matrix, and denote sigmoid and hyperbolic tangent functions, respectively.

[0041] The Dropout layer avoids over-reliance on a certain GRU unit during training by randomly turning off part of the GRU units, effectively preventing overfitting. The mask generation formula is: (11) The output after Dropout processing is​ (12) where, is the reserved probability, is the hidden state after Dropout, denotes the mask.

[0042] The high-dimensional multi-scale features can be mapped to a low-dimensional feature sequence by using the GRU encoder, which removes the redundancy and interference noise of high-dimensional features and helps to identify and extract key high-weight features in the modulated signal.

[0043] The decoder adopts the TimeDistributed Dense structure, in which the two encapsulated Dense layers have 32 neurons respectively. The TimeDistributed layer ensures the correlation between time steps by applying the same weight matrix and bias vector to each time step, avoiding the loss of time dimension information. In the feature reconstruction process, the nonlinear transformation of the feature space is realized through the fully connected layer, and its mathematical expression is: (13) where, and are the weight matrix and bias vector of the decoder, is the output of the th time step.

[0044] The reconstruction loss of the entire AE module is usually defined as the mean square error (MSE), and the formula is: (14) where, is the sequence length, is the input of the th time step, is the reconstruction output of the th time step. By minimizing the reconstruction loss , the AE module can effectively extract high-weight features in the input sequence and filter redundant information and noise in the reconstruction process.

[0045] The AE module realizes learning and efficient reconstruction of signal features through a self-encoder, and simultaneously uses a GRU encoder to map high-dimensional multi-scale features to low-dimensional sequences, remove redundant information and noise interference, and extract key high-weight features in the modulated signal. The decoder adopts a TimeDistributed Dense structure to ensure the relevance between time steps, avoid loss of time dimension information, and realize nonlinear transformation of the feature space through a fully connected layer. Finally, the module realizes efficient feature extraction, noise suppression and preservation of time dimension information while minimizing the reconstruction loss, significantly improving the feature reconstruction and recognition ability of the modulated signal.

[0046] The AE module in the above embodiments preferably adopts a GRU encoder.

[0047] In the data preprocessing stage, the present application eliminates the influence of phase offset through phase transformation, and uses a multi-scale feature extraction module (MSFE) to capture local details and global structural features of the signal, reduce the information loss rate, then, the self-encoder and the encoder module realize time information preservation, feature reconstruction and noise suppression, while minimizing the reconstruction loss, significantly improving the feature reconstruction and recognition ability of the modulated signal.

[0048] From the common general knowledge, the present application can be realized through other embodiments without departing from the spirit or essential characteristics thereof. Therefore, the above disclosed embodiments are merely illustrative in all aspects, and are not the only ones. All changes within the scope of the present application or within the scope equivalent to the present application are included in the present application.

Claims

1. An automatic modulation recognition method based on multi-scale features and an improved encoder, characterized in that, include: Step 1: Convert the IQ format signal into AP format data using amplitude and phase transformation formulas. Input the P component of the AP format data into the PT module for phase transformation, and input it together with the A component of the AP format data into the multi-scale feature extraction module. Step 2: The multi-scale feature extraction module uses four multi-scale one-dimensional convolution modules to process the P and A components after phase transformation in parallel to extract features; Step 3: Input the features extracted in Step 2 into the feature denoising and temporal modeling module for feature optimization and recognition. Perform Softmax classification on the optimized and recognized features and output the results.

2. The automatic modulation recognition method based on multi-scale features and an improved encoder according to claim 1, characterized in that: The phase transformation formula is as follows: in, , These represent the amplitude and phase components of the AP sample signal, respectively. , These represent the in-phase and quadrature components of the IQ signal, respectively.

3. The automatic modulation recognition method based on multi-scale features and an improved encoder according to claim 1, characterized in that: The specific steps for inputting the P component of AP format data into the PT module for phase transformation include: Step 4: In the PT module, the phase shift parameters are obtained through training with a large number of samples using flattened layers and fully connected layers. ; Step 5: Phase offset parameters Through the offset matrix Obtain the anti-split factor The specific value; Step 6: Adjust the anti-split factor The specific value is substituted into the formula for eliminating the phase shift effect in AP format data to calculate the P component after phase transformation.

4. The automatic modulation recognition method based on multi-scale features and an improved encoder according to claim 3, characterized in that: The formula for eliminating the phase shift effect in AP format data is: in, , This represents AP format data without phase shift. This represents the parameters of the first row of the offset matrix. This represents the parameters of the second row of the offset matrix, and... This is called the anti-split factor. This indicates the coordinates after the phase shift has occurred.

5. The automatic modulation recognition method based on multi-scale features and an improved encoder according to claim 1, characterized in that: The feature input denoising and temporal modeling module includes: an autoencoder and a GRU encoder. Autoencoders are responsible for learning and reconstructing features; The GRU encoder is used for multi-scale feature extraction and compression.

6. The automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure according to claim 5, characterized in that: The GRU encoder consists of two decreasing CUDNNNGRU layers and two Dropout layers. The CUDNNNGRU layers process sequence data through a recurrent neural network structure, while the Dropout layers avoid over-reliance on a particular GRU unit during training by randomly disabling some GRU units, thus effectively preventing overfitting.

7. The automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure according to claim 6, characterized in that: The output after processing by the Dropout layer is: in, To preserve probability, This is the hidden state after Dropout. Indicates a mask.

8. The automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure according to claim 7, characterized in that: The mask generation formula is: 。 9. The automatic modulation recognition method based on multi-scale feature extraction and improved encoder structure according to claim 5, characterized in that: The autoencoder consists of a decoder and a GRU encoder. The decoder adopts a TimeDistributed Dense structure, in which the two encapsulated Dense layers each have 32 neurons.

Images