Signal modulation recognition method based on complex-valued partial convolutional neural network

Abstract

The present invention discloses a signal modulation mode recognition method based on a complex-valued partial convolutional neural network. The recognition neural network used includes a plurality of stage parts and complex-valued embedding layers corresponding to the stage parts one by one. The complex-valued embedding layers are arranged at intervals with the stage parts and connected in sequence. Each complex-valued embedding layer is located before the corresponding stage part. The output features of the last stage part are processed by a complex-valued pooling module, a fully connected layer and an output layer in sequence to obtain a probability distribution vector. Each stage part includes a plurality of feature extraction blocks connected in sequence. The feature extraction block includes a complex-valued partial convolution module. Only a part of the input features is convolved in the complex-valued partial convolution module. The present invention can reduce the amount of calculation, reduce the computational complexity, and improve the performance of the network. It also has the advantages of high accuracy and light weight.

Description

Technical Field

[0001] The present invention belongs to the technical field of convolutional networks, and in particular relates to a signal modulation mode recognition method based on a complex-valued partial convolutional neural network. Background Art

[0002] Automatic Modulation Classification (AMC) technology focuses on the identification of signal modulation types and plays an important role in the field of wireless communications, especially in industrial wireless communication systems. With the rapid development of wireless communication technology, spectrum resources have become increasingly scarce. Cognitive Radio (CR) technology has become the key to ensuring communication quality and effectively utilizing spectrum resources, and AMC technology is an important part of CR technology.

[0003] Traditional AMC methods include likelihood-based and feature-based methods. Feature-based methods are favored due to their stability and robustness in complex and changing environments. They accurately identify modulation modes by designing signal features (such as cyclic spectrum, frequency components, etc.) and combining them with advanced machine learning models. Although this method has a clear physical explanation for the features, it requires extensive expertise to be effectively applied to real-world communication environments.

[0004] In recent years, advanced methods represented by real-valued convolutional neural networks have brought innovative solutions to AMC. However, this type of model treats the two components of the complex-valued modulated signal as discrete real-valued inputs, which will cause structural phase damage to the original signal, thereby reducing the performance and interpretability of the model. Therefore, those skilled in the art have further proposed a complex-valued convolutional network, which retains the amplitude and phase information of the I and Q signals through complex-valued operations, further improving network performance and providing a new solution for AMC. However, complex-valued operations greatly increase the computational complexity and the timeliness cannot meet the requirements. Summary of the invention

[0005] The present invention proposes a signal modulation mode recognition method based on a complex-valued partial convolutional neural network, the purpose of which is to solve the problems of high computational complexity and poor timeliness in the existing signal modulation mode recognition method based on a complex-valued convolutional network.

[0006] The technical solution of the present invention is as follows:

[0007] A signal modulation mode recognition method based on a complex-valued partial convolutional neural network, wherein a complex-valued signal is input into a recognition neural network to obtain a probability distribution vector, and the signal modulation type corresponding to the largest probability value in the probability distribution vector is the recognition result, wherein the recognition neural network comprises a plurality of stage parts and complex-valued embedding layers corresponding to the stage parts one by one, wherein the complex-valued embedding layers are arranged at intervals with the stage parts and are sequentially connected, and each complex-valued embedding layer is located before the corresponding stage part; the output features of the last stage part are sequentially processed by a complex-valued pooling module, a fully connected layer and an output layer to obtain the probability distribution vector;

[0008] The complex-valued embedding layer is used to downsample and expand the channels of the input features;

[0009] Each stage part includes a plurality of feature extraction blocks connected in sequence, wherein the feature extraction block includes a complex-valued partial convolution module, and the complex-valued partial convolution module only performs convolution on a part of the input feature.

[0010] As a further improvement of the signal modulation mode recognition method based on the complex-valued partial convolutional neural network, the specific processing method of the complex-valued partial convolution module is:

[0011] Assume that the input feature of the complex-valued partial convolution module is , is the real part of the input feature, is the imaginary part of the input feature, and the complex convolution parameter is , is the real part of the complex convolution parameter, is the imaginary part of the complex convolution parameter, then first and Split into two parts along the channel in the same proportion and , and , then and The plural form pass Perform complex-valued convolution to get ,in , ,Will and The concatenated result is used as the real part of the partial convolution result. , and The concatenated result is used as the imaginary part of the partial convolution result. , thus obtaining the result of complex-valued partial convolution .

[0012] As a further improvement of the signal modulation mode recognition method based on complex-valued partial convolutional neural network: the complex-valued embedding layer includes a one-dimensional complex-valued convolution module and a one-dimensional complex-valued batch normalization layer connected in sequence.

[0013] As a further improvement of the signal modulation mode identification method based on the complex-valued partial convolutional neural network: the feature extraction block includes the complex-valued partial convolution module, the complex-valued processing module and the complex-valued point convolution module connected in sequence.

[0014] As a further improvement of the signal modulation mode recognition method based on complex-valued partial convolutional neural network: the complex-valued processing module includes a complex-valued batch normalization layer and a ReLU activation layer.

[0015] As a further improvement of the signal modulation mode recognition method based on the complex-valued partial convolutional neural network: the recognition neural network includes three stage parts, the first stage part includes two feature extraction blocks, the second stage part includes three feature extraction blocks, and the third stage part includes three feature extraction blocks.

[0016] As a further improvement of the signal modulation mode recognition method based on complex-valued partial convolutional neural network: the number of channels of the three complex-valued embedding layers are 16, 32, and 64 respectively.

[0017] As a further improvement of the signal modulation mode recognition method based on complex-valued partial convolutional neural network: the output layer includes a Softmax function.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. The present invention uses complex-valued partial convolution to significantly reduce the amount of calculation, reduce the computational complexity, and solve the problem of poor timeliness.

[0020] 2. The present invention introduces complex-valued point convolution and complex-valued partial convolution in the feature extraction block, which cooperate with each other to promote information fusion between channels, solve the problem that complex-valued partial convolution cannot well fuse information between channels, and improve the performance of the network.

[0021] 3. When the number of partial feature extraction blocks in each stage of the present invention is configured as 2, 3, and 3, it can not only ensure the optimal network performance, but also ensure the rationality of the number of parameters.

[0022] 4. The network of the present invention also has the advantage of being lightweight, and also solves the confusion problem between QAM16 and QAM64. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 To identify the architecture diagram of the neural network;

[0024] Figure 2 Schematic diagram of the process of complex-valued partial convolution calculation;

[0025] Figure 3 The average confusion matrix obtained from the visualization experiment;

[0026] Figure 4 This is the confusion matrix obtained from the visualization experiment when the signal-to-noise ratio is 18dB. DETAILED DESCRIPTION

[0027] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0028] A signal modulation mode recognition method based on complex-valued partial convolutional neural network, the complex-valued signal is input into the recognition neural network to obtain a probability distribution vector, and the signal modulation type corresponding to the largest probability value in the probability distribution vector is the recognition result.

[0029] like Figure 1 , the recognition neural network includes several stage parts and complex-valued embedding layers corresponding to the stage parts one by one, the complex-valued embedding layers are arranged at intervals with the stage parts and connected in sequence, and each complex-valued embedding layer is located before the corresponding stage part. In this embodiment, the recognition neural network includes 3 stage parts. The output features of the last stage part are processed by the complex-valued pooling module, the fully connected layer and the output layer (Softmax function) in sequence to obtain the probability distribution vector.

[0030] The complex-valued embedding layer is used to downsample and expand the channels of the input features. Specifically, the complex-valued embedding layer includes a one-dimensional complex-valued convolution module (size is 2, stride is 2) and a one-dimensional complex-valued batch normalization layer connected in sequence. In this embodiment, the number of channels of the three complex-valued embedding layers are 16, 32, and 64, respectively.

[0031] Each stage part includes a plurality of feature extraction blocks connected in sequence. In this embodiment, the first stage part includes 2 feature extraction blocks, the second stage part includes 3 feature extraction blocks, and the third stage part includes 3 feature extraction blocks. The stage parts are used for feature extraction and will not cause transformation in channels and dimensions.

[0032] The feature extraction block includes a complex-valued partial convolution module, a complex-valued processing module and a complex-valued point convolution module which are connected in sequence.

[0033] like Figure 2 In the complex-valued partial convolution module, considering the redundancy of features between channels, complex-valued convolution is performed on only some channels, and the remaining channels are directly retained. The specific processing method is:

[0034] Assume that the input feature of the complex-valued partial convolution module is , is the real part of the input feature, is the imaginary part of the input feature, and the complex convolution parameter is , is the real part of the complex convolution parameter, is the imaginary part of the complex convolution parameter, then first and Split into two parts along the channel in the same proportion and , and , then and The plural form pass Perform complex-valued convolution to get ,in , ,Will and The concatenated result is used as the real part of the partial convolution result. , and The concatenated result is used as the imaginary part of the partial convolution result. , thus obtaining the result of complex-valued partial convolution .

[0035] The complex-valued processing module includes a complex-valued batch normalization layer and a ReLU activation layer.

[0036] The working process of the recognition neural network is as follows: suppose the complex-valued signal input to the recognition neural network contains 128 sampling points, each sampling point is complex-valued, and after the first complex-valued embedding layer, the feature with 16 channels and 64 dimensions is obtained. After the feature extraction of the first stage, the feature shape obtained is still , that is, the stage part does not change the channel and dimension of the feature. The feature shape obtained after the second complex-valued embedding layer is , that is, the channel is expanded to 2 times and the dimension is reduced to half of the original. After the second stage and the third complex-valued embedding layer, the feature shape obtained is , and then the feature shape obtained after complex value pooling is , and finally the probability distribution vector is obtained through the fully connected layer and Softmax function.

[0037] The data sets used for training and testing in this embodiment are RML2016.10a and RML2016.10b, which simulate the modulation signals generated in the channel environment with additive Gaussian white noise, selective fading, center frequency offset, and sampling rate offset. The former contains 8 digital modulation modes: (PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK), 3 analog modulation modes (AM-DSB, AM-SSB, WBFM), a total of 220,000 modulation data of 11 modulation types; the latter contains 10 modulation modes (AMSSB is missing), containing 1,200,000 modulation data. One signal of both data sets is in the format of 2×128, where 2 represents the I and Q signals, and 128 represents the data of 128 sampling points. Before use, the real I and imaginary Q signals are synthesized into one complex value signal. 80% of the data sets are randomly sampled as the training set, and the other 20% are used as the test set.

[0038] The training hardware and software environment is: NVIDIA RTX3070s, python3.11 and pytorch2.2.1. The batch size is set to 512. In order to train the optimal model, the model early stopping and dynamic learning rate technology are used, the number of iterations is set to 200 generations, the initial learning rate is 0.001, and the learning rate is halved when the loss value does not decrease after 10 generations. When it does not decrease after more than 20 generations, the training is stopped.

[0039] The following comparative experiments verify the capabilities of this model:

[0040] 1. Performance comparison.

[0041] In order to analyze the performance, the trained recognition neural network model (hereinafter referred to as CPCNet, i.e. Complex PartialConvolution) was compared with other models. Because this model is a complex-valued lightweight network based on convolution, the models selected for comparison include the following three categories:

[0042] The first category is MCLDNN, which has the best accuracy at present;

[0043] The second category includes three lightweight networks (MCNet, SCNN, and PET_CGDNN);

[0044] The third category includes two complex-valued neural networks (CDSN and CDSCNN).

[0045] Table 1 - Performance comparison of different network models

[0046]

[0047] The parameters are shown in Table 1. It can be seen that in the RML2016.10a dataset, although MCLDNN has the best performance, CPCNet is 0.09% lower than MCLDNN in terms of accuracy, but the FLOPs (Floating point Operations) of CPCNet is only 1 / 26 of that of MCLDNN.

[0048] Due to the significant increase in the amount of data in the RML2016.10b dataset, it can be regarded as a data-enhanced version of the RML2016.10a dataset. After data enhancement, the overall accuracy (OA) of different networks has been improved. It is particularly noteworthy that the CPCNet model significantly surpasses MCLDNN in OA, achieving an improvement of 2.26% and 2.84% compared with MCLDNN and PET_CGDNN, respectively. In addition, the OA of CPCNet is significantly higher than other lightweight networks such as MCNet, while the number of parameters (Params) is significantly lower than that of complex-valued convolutional networks such as CDSN. These results further confirm that complex-valued operations can fully utilize the physical correlation of IQ signals and significantly improve the recognition ability of the network. The excellent performance of the CPCNet model shows that through carefully designed lightweight structures and complex-valued operations, more accurate signal recognition and classification can be achieved while reducing the complexity of the model. This optimization not only improves the efficiency of the model, but also maintains a high accuracy, showing the great potential of applying lightweight complex-valued networks in the field of automatic modulation recognition.

[0049] 2. Effectiveness of complex-valued partial convolution modules.

[0050] In order to verify the effectiveness of the complex-valued partial convolution module (hereinafter referred to as CPC, i.e., Complex PartialConvolution) proposed by the present invention, this embodiment conducts experimental analysis on the RML2016.10a and RML2016.10b datasets. The specific operation details and results of the experiment are shown in Table 2. The experiment uses the above-mentioned CPCNet as the basic architecture, modifies or replaces the complex-valued partial convolution module (CPC) and the complex-valued point convolution module (CPW, Complex-Valued PointwizeConvolution), and conducts corresponding comparative experiments. CDW in Table 2 refers to complex-valued depthwise separable convolution, and 1d_cc refers to one-dimensional complex convolution.

[0051] Table 2 - Analysis of the effectiveness of complex-valued partial convolution modules

[0052]

[0053] By comparing the first two sets of experiments, it can be observed that the channel fusion operation implemented by CPW can significantly improve the overall accuracy (OA) of the model. Although this operation leads to an increase in the number of parameters (Params) to a certain extent, its performance improvement proves its value. This shows that CPW plays a positive role in enhancing the model's feature fusion capability, helping to capture richer signal features, thereby improving the accuracy of modulation recognition.

[0054] The main purpose of the third group of experiments is to compare and analyze the two lightweight operators, CPC and CDW, proposed in this invention, to evaluate their advantages and disadvantages. The experimental results show that although the number of parameters (Params) of CPC+CPW is higher than that of CDW+CPW, the overall accuracy (OA) on the RML2016.10b dataset is significantly improved. This finding particularly highlights the potential of CPC in processing more complex datasets. It should be pointed out that the number of parameters is an indirect measure of the computational complexity of the model. In fact, due to the structural characteristics of CDW, it inevitably leads to higher memory access time, which cannot be reflected by the number of parameters alone. Therefore, it is not comprehensive to evaluate the complexity of the two operators only by Params and FLOPs, and other indicators are needed for in-depth comparative analysis. For this reason, the batch size (Batchsize) is set to 512 in this experiment, and the speed of the four groups of experimental models in the table is measured. The forward transmission time (Inference Time) of each model is shown in Table 2. It can be proved that although CDW has fewer Params, its characteristics inevitably increase the memory access time, and the actual running speed is not as good as CPC.

[0055] The last set of experiments aims to compare with the first set to verify the effect of using the designed lightweight operator CPC to replace the conventional complex-valued convolution operation. The experimental results confirm that this replacement can significantly reduce the number of parameters (Params) of the model, while the overall accuracy (OA) of the model can also be improved. This finding shows that CPC not only helps to reduce the computational complexity of the model, but also has the potential to maintain or improve the performance of the model. It also confirms that there is a large redundancy in the signal information between different channels.

[0056] 3. Experimental analysis of the optimal form.

[0057] In order to verify that the recognition neural network model in this embodiment is the best form under the same architecture, an experimental analysis was conducted on the RML2016.10b dataset. The purpose of this experiment is to explore the specific impact of the number of repetitions of feature extraction blocks (2, 3, and 3 in this embodiment) at different stages in the network on network performance. This experiment specifically focuses on adjusting the number of block repetitions in the second and third stages. , because usually deeper features contain richer semantic information, which has a more significant impact on network performance. The experimental results are shown in Table 3:

[0058] Table 3 - Effect of different feature extraction block repetition times on network performance

[0059]

[0060] From the data in Table 3, we can observe that there is basically a positive correlation between the overall accuracy (OA) and the number of parameters (Params), which is consistent with the general law of neural networks. However, when the number of repetitions of the block is increased to version E7, it can be found that OA no longer increases with the increase in the number of parameters. In particular, when the number of repetitions of the block is set to version E6, OA reaches the maximum value. It can be concluded that in the proposed network, the number of repetitions of the blocks in each stage should be set to 2, 3, and 3 respectively. Such a configuration can ensure the optimal performance of the network while maintaining the rationality of the number of parameters.

[0061] 4. Visualization Experiment

[0062] Figure 3 and Figure 4 The following are the average confusion matrix obtained from the visualization experiment and the confusion matrix when the signal-to-noise ratio is 18dB. The horizontal axis in the figure is the true label, the vertical axis is the predicted result, and each square is the corresponding prediction probability. The higher the probability on the diagonal line, the better the classification performance.

[0063] There are two classic problems in this field, namely the confusion between QAM16 and QAM64 and between WBFM and AM-DSB. The main reason for the confusion in the former is that the distribution of constellation points is very similar, and the latter is because they are both continuous analog modulation. Figure 3 and Figure 4 It can be seen that CPCNet perfectly solves the confusion problem between QAM16 and QAM64, and the confusion problem between WBFM and AM-DSB has also been improved to a certain extent.

[0064] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A signal modulation mode recognition method based on complex-valued partial convolutional neural network, wherein a complex-valued signal is input into a recognition neural network to obtain a probability distribution vector, and the signal modulation type corresponding to the largest probability value in the probability distribution vector is the recognition result, characterized in that: The recognition neural network includes a plurality of stage parts and complex-valued embedding layers corresponding to the stage parts one by one, the complex-valued embedding layers are arranged alternately with the stage parts and connected in sequence, and each complex-valued embedding layer is located before the corresponding stage part; the output features of the last stage part are processed in sequence by the complex-valued pooling module, the fully connected layer and the output layer to obtain the probability distribution vector; The complex-valued embedding layer is used to downsample and expand the channels of the input features; Each stage part includes a plurality of sequentially connected feature extraction blocks, each of which includes a complex-valued partial convolution module, and each of which performs convolution on only a part of the input feature; The specific processing method of the complex-valued partial convolution module is: Assume that the input feature of the complex-valued partial convolution module is , is the real part of the input feature, is the imaginary part of the input feature, and the complex convolution parameter is , is the real part of the complex convolution parameter, is the imaginary part of the complex convolution parameter, then first and Split into two parts along the channel in the same proportion and , and , then and The plural form pass Perform complex-valued convolution to get ,in , ,Will and The concatenated result is used as the real part of the partial convolution result. , and The concatenated result is used as the imaginary part of the partial convolution result. , thus obtaining the result of complex-valued partial convolution ; The feature extraction block includes the complex-valued partial convolution module, the complex-valued processing module and the complex-valued point convolution module connected in sequence; The recognition neural network includes three stages, the first stage includes two feature extraction blocks, the second stage includes three feature extraction blocks, and the third stage includes three feature extraction blocks.

2. The signal modulation mode recognition method based on complex-valued partial convolutional neural network according to claim 1, characterized in that: The complex-valued embedding layer includes a one-dimensional complex-valued convolution module and a one-dimensional complex-valued batch normalization layer connected in sequence.

3. The signal modulation mode recognition method based on complex-valued partial convolutional neural network according to claim 1, characterized in that: The complex-valued processing module includes a complex-valued batch normalization layer and a ReLU activation layer.

4. The signal modulation mode recognition method based on complex-valued partial convolutional neural network according to claim 1, characterized in that: The number of channels of the three complex-valued embedding layers are 16, 32, and 64 respectively.

5. The signal modulation mode recognition method based on complex-valued partial convolutional neural network according to any one of claims 1 to 3, characterized in that: The output layer includes a Softmax function.

Images