STBC-OFDM signal blind recognition method and device based on deep learning and fourth-order lag moment spectrum

Abstract

The present application proposes a method for blind recognition of STBC-OFDM signals based on deep learning and fourth-order lagged moment spectrum, which relates to the technical field of blind signal recognition, including: calculating the fourth-order lagged moment of the received signal and generating a fourth-order lagged moment vector; using two-dimensional vector concatenation to merge the fourth-order lagged moment vector into a fourth-order lagged moment spectrum; constructing an attention-guided multi-scale dilated convolution network, wherein the attention-guided multi-scale dilated convolution network includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer, inputting the fourth-order lagged moment spectrum into the attention-guided multi-scale dilated convolution module, outputting multi-scale guided features, and then inputting the multi-scale guided features into the feature fusion layer and the residual layer, and outputting the recognition results through a fully connected layer with softmax as the activation function. The recognition performance of the present application has been significantly improved, and it has good adaptability to strong interference environments. Moreover, the present application does not require prior information such as channels and noise, and is more suitable for non-cooperative communications than existing algorithms.

Description

Technical Field

[0001] The present application relates to the technical field of blind signal recognition, and in particular to a method and device for blind recognition of STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum. Background Art

[0002] Blind signal identification (BSI) has a wide range of applications in many military and civilian fields, including radio monitoring, communication reconnaissance, electromagnetic countermeasures, and spectrum sensing. For example, when conducting communication reconnaissance on transmitter signals, it is often necessary to identify the signal type from the received data under the condition that the prior information such as the channel and noise are unknown, so as to further decode and restore the original information of the transmitter. With the shortage of spectrum resources, the combination of space-time block code (STBC) and orthogonal frequency division multiplexing (OFDM) technology has been increasingly widely used due to its advantages of high frequency band utilization and strong anti-multipath interference ability. As an important channel coding method, in-depth research on blind identification of STBC-OFDM signals under non-cooperative conditions will further expand its application scope in the field of radio communications and has broad and far-reaching significance.

[0003] Most of the existing algorithms are traditional recognition methods based on feature extraction. Their performance depends on professional knowledge and experience, and they often require more prior information. They have many restrictions on channels and noise, and are not suitable for signal recognition under non-cooperative communication conditions. According to the correlation differences of STBC-OFDM signal coding methods, traditional algorithms obtain statistical feature quantities that can reflect the essential characteristics of the signal, further establish a decision tree, and perform hypothesis tests item by item to complete the recognition. Based on the spatiotemporal redundancy of the signal, the cross-correlation function between different receiving antennas can be calculated and used as the identification feature to achieve effective recognition of SM-OFDM (Special Multiplexing-Orthogonal Frequency Division Multiplexing) and AL-OFDM (Alamouti-Orthogonal Frequency Division Multiplexing) signals. Based on the second-order signal cyclostationarity, the signal type can be confirmed by comparing the second-order cyclic statistics of the received signal with the threshold. According to the correlation differences of different coding matrices, the type of the transmitted STBC-OFDM signal can be identified by calculating the fourth-order lag moment of the received signal. Traditional algorithms have been studied in depth, but because they require manual feature extraction and setting of test thresholds, there are still problems such as difficulty in feature selection, complex parameter adjustment process, and sensitivity to noise. They are also less adaptable to non-cooperative communication scenarios with a lack of prior information.

[0004] In recent years, with the rapid development of deep learning technology in the field of computer vision (CV), thanks to the improvement of GPU parallel computing capabilities, deep learning models have achieved stronger classification performance with their powerful mapping capabilities for massive data. Combined with the performance advantages of this technology, domestic and foreign scholars have gradually applied it to the field of BSI, especially in the fields of modulation recognition, individual radiation source recognition and radar signal recognition, deep learning methods have been widely used. Compared with the intuitive and visual features of visual images, the intrinsic features of signals are often more difficult to be directly discovered by deep learning models. Therefore, many of the above algorithms need to pre-process the received signal first, and after obtaining its constellation diagram, cyclic spectrum, time-frequency image and other features, one or more of them are used as the input of the deep learning model and combined with ensemble learning to achieve training and recognition. The successful application of these methods shows the feasibility of introducing deep learning into STBC-OFDM recognition, but the coding method of STBC-OFDM signals is relatively complex, and it cannot be effectively recognized by using only simple transformation methods or migrating signal transformation methods in other fields (constellation diagram, cyclic spectrum, etc.).

[0005] In addition, as the electromagnetic environment and information technology become increasingly dense and complex, signal transmission is increasingly affected by channel conditions and noise interference. The traditional STBC-OFDM signal recognition method based on feature extraction and threshold decision-making can no longer meet the actual needs of accurate and rapid recognition in actual communication environments. It is urgent to further improve the algorithm's adaptability to low signal-to-noise ratios. In addition, the existing methods require manual feature extraction and setting of test thresholds, and their recognition performance is greatly affected by parameter selection. After changing the channel and noise, the decision rules need to be reset, and the robustness to different communication environments is poor. Considering that in actual communication processes, signals often need to be transmitted in complex electromagnetic environments with low signal-to-noise ratios, and signals are greatly affected by noise and channel attenuation. Therefore, how to correctly identify the transmitter signal type in a strong interference environment needs to be solved urgently. Summary of the invention

[0006] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0007] To this end, the first purpose of this application is to propose a blind recognition method for STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum, which solves the technical problems that the existing methods need to manually extract features and set test thresholds, and their recognition performance is greatly affected by parameter selection, and the robustness to different communication environments is poor. At the same time, it solves the technical problem that the existing methods are greatly affected by noise and channel attenuation when transmitting in a complex electromagnetic environment with a low signal-to-noise ratio and cannot be accurately and quickly recognized. It has excellent comprehensive recognition performance, a short recognition time for a single sample, and significantly improves the recognition accuracy under low signal-to-noise ratio, and has good adaptability to strong interference environments. In addition, the deep learning model can directly identify the FOLMS (fourth order lag moment spectrum) samples obtained by preprocessing, so this method does not require prior information such as channels and noise, and is more suitable for non-cooperative communication than existing algorithms.

[0008] The second objective of this application is to propose a STBC-OFDM signal blind recognition device based on deep learning and fourth-order lag moment spectrum.

[0009] A third object of the present application is to provide a non-transitory computer-readable storage medium.

[0010] To achieve the above-mentioned purpose, the first embodiment of the present application proposes a blind recognition method for STBC-OFDM signals based on deep learning and fourth-order lagged moment spectrum, including: calculating the fourth-order lagged moment of the received signal and generating a fourth-order lagged moment vector; using two-dimensional vector splicing to merge the fourth-order lagged moment vector into a fourth-order lagged moment spectrum; constructing an attention-guided multi-scale dilated convolution network, taking the fourth-order lagged moment spectrum as input, and outputting a recognition result, wherein the attention-guided multi-scale dilated convolution network includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer, taking the fourth-order lagged moment spectrum as input, and outputting a recognition result, including: inputting the fourth-order lagged moment spectrum into the attention-guided multi-scale dilated convolution module, outputting multi-scale guided features, and then inputting the multi-scale guided features into the feature fusion layer and the residual layer, and outputting the recognition result via a fully connected layer with softmax as the activation function.

[0011] Optionally, in one embodiment of the present application, the fourth-order lag moment of the received signal is calculated, and a fourth-order lag moment vector is generated, wherein the fourth-order lag moment is expressed as:

[0012]

[0013] Among them, for the i-th receiving antenna, N b OFDM blocks Define y(q,τ) to represent the fourth-order lag moment when the delay parameter is (0,τ,0,τ),

[0014] The fourth-order lag moment vector is expressed as:

[0015] V=[E[y(0,τ)],E[y(1,τ)],…,E[y(N s -1,τ)]]

[0016] Among them, E[y(N s-1 ,τ)] represents the fourth-order hysteresis moment y(N s-1 ,τ).

[0017] Optionally, in one embodiment of the present application, the fourth-order lag moment spectrum is expressed as:

[0018]

[0019] in, represents the kth fourth-order lag moment vector.

[0020] Optionally, in one embodiment of the present application, inputting the fourth-order lag moment spectrum into an attention-guided multi-scale dilated convolution module and outputting a multi-scale guided feature comprises the following steps:

[0021] The input fourth-order lag moment spectrum is subjected to multi-scale dilated convolution to obtain multi-scale features, and obtains targeted feature representations of the fourth-order lag moment spectra of different STBC-OFDM signals;

[0022] A standard convolutional layer is used to further extract deep features of the multi-scale feature map;

[0023] Finally, the multi-scale deep features are guided by the convolutional block attention module to generate multi-scale guided features.

[0024] Optionally, in one embodiment of the present application, the multi-scale deep features are guided by a convolutional block attention module to generate multi-scale guided features, including:

[0025] Input the multi-scale deep features into the channel attention module to generate channel attention features;

[0026] The product of the channel attention feature and the multi-scale deep feature is used as the input of the spatial attention module, and the final multi-scale guided feature is output.

[0027] Optionally, in one embodiment of the present application, inputting multi-scale deep features into a channel attention module includes:

[0028] The input feature map first undergoes global average pooling and global maximum pooling on the channel dimension to produce two channel attention maps;

[0029] The two generated channel attention maps are fed into a shared multilayer perceptron, and then the features output by the shared multilayer perceptron are summed up element-wise, and then subjected to a sigmoid activation operation to generate the final channel attention features.

[0030] Optionally, in one embodiment of the present application, the product of the channel attention feature and the multi-scale deep feature is used as the input of the spatial attention module to generate the final multi-scale guided feature, including:

[0031] Perform global average pooling and global maximum pooling on the spatial dimension of the product of the channel attention feature and the multi-scale deep feature to obtain the first feature map and the second feature map, and then perform a splicing operation on the first feature map and the second feature map;

[0032] The concatenated feature map is subjected to standard convolution operation and sigmoid activation function to generate spatial attention features;

[0033] Multiply the spatial attention feature and the input feature of the module to obtain the final output feature.

[0034] To achieve the above-mentioned purpose, the second embodiment of the present application proposes a STBC-OFDM signal blind recognition device based on deep learning and fourth-order lag moment spectrum, including: a vector generation module, a splicing module, and a result generation module, wherein:

[0035] A vector generation module, used for calculating the fourth-order lag moment of the received signal and generating a fourth-order lag moment vector;

[0036] A splicing module, used for merging the fourth-order lag moment vector into a fourth-order lag moment spectrum by using two-dimensional vector splicing;

[0037] The result generation module is used to construct an attention-guided multi-scale dilated convolutional network, which takes the fourth-order lag moment spectrum as input and outputs the recognition result, where:

[0038] The attention-guided multi-scale dilated convolutional network includes an attention-guided multi-scale dilated convolutional module, a feature fusion layer, and a residual layer. It takes the fourth-order lag moment spectrum as input and outputs the recognition results, specifically:

[0039] The fourth-order lagged moment spectrum is input into the attention-guided multi-scale dilated convolution module to output the multi-scale guided features. The multi-scale guided features are then input into the feature fusion layer and the residual layer, and the recognition results are output through the fully connected layer with softmax as the activation function.

[0040] In order to achieve the above-mentioned objectives, the third aspect embodiment of the present application proposes a non-temporary computer-readable storage medium. When the instructions in the storage medium are executed by a processor, a STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum can be executed.

[0041] The STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum, the STBC-OFDM signal blind recognition device based on deep learning and fourth-order lagged moment spectrum and the non-temporary computer-readable storage medium of the embodiment of the present application solve the technical problems that the existing method needs to manually extract features and set test thresholds, its recognition performance is greatly affected by parameter selection, and the robustness to different communication environments is poor. At the same time, it solves the technical problem that the existing method cannot accurately and quickly identify the signal when it is transmitted in a complex electromagnetic environment with a low signal-to-noise ratio and is greatly affected by noise and channel attenuation. It has excellent comprehensive recognition performance, a short recognition time for a single sample, and significantly improves the recognition accuracy under low signal-to-noise ratio, and has good adaptability to strong interference environments. In addition, the deep learning model can directly identify the FOLMS samples obtained by preprocessing, so this method does not require prior information such as channels and noise, and is more suitable for non-cooperative communication than existing algorithms.

[0042] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0044] Figure 1 A flowchart of a method for blind recognition of STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum provided in Example 1 of the present application;

[0045] Figure 2 A block diagram of an STBC-OFDM transmitter of an STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

[0046] Figure 3 A structural diagram of the FOLMS / AMDC-net method of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum in an embodiment of the present application;

[0047] Figure 4 A schematic diagram of fourth-order delayed moment spectra of four STBC-OFDM signals of a blind identification method of STBC-OFDM signals based on deep learning and fourth-order delayed moment spectrum in an embodiment of the present application;

[0048] Figure 5 It is a schematic diagram of the overall structure of AMDC-net of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum in an embodiment of the present application;

[0049] Figure 6 A schematic diagram of a multi-scale dilated convolution of a blind recognition method for STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

[0050] Figure 7 A schematic diagram of a convolutional block attention module scheme of a STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application;

[0051] Figure 8 A schematic diagram of an attention-guided multi-scale dilated convolution module of a blind recognition method for STBC-OFDM signals based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application;

[0052] Fig. 9A schematic diagram of the position of an attention-guided multi-scale dilated convolution module in an attention-guided multi-scale dilated convolution network of a blind recognition method for STBC-OFDM signals based on deep learning and fourth-order lagged moment spectrum according to an embodiment of the present application;

[0053] Fig.10 A structural schematic diagram of the feature fusion and residual block parts of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application;

[0054] Fig.11 A schematic diagram showing the performance comparison of AMDC-net and ASC-net for multi-scale feature evaluation according to an embodiment of the present application;

[0055] Fig.12 Schematic diagram of confusion matrix of AMDC-net and ASC-net in different environments of the embodiment of the present application;

[0056] Fig.13 A schematic diagram showing the performance comparison of four networks used for attention mechanism evaluation according to an embodiment of the present application;

[0057] Fig.14 A schematic diagram of the confusion matrix of the four networks in the embodiment of the present application at -14 decibels;

[0058] Fig.15 It is a schematic diagram of the recognition performance of two fusion methods in the embodiment of the present application;

[0059] Fig.16 A schematic diagram showing a comparison of recognition performances of different methods according to an embodiment of the present application;

[0060] Fig.17 A schematic diagram of the structure of a STBC-OFDM signal blind recognition device based on deep learning and fourth-order lag moment spectrum provided in Example 2 of the present application. DETAILED DESCRIPTION

[0061] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0062] As an important feature extraction and transformation method, spectrum analysis has a wide range of applications in signal processing and recognition. With the rise of deep learning in the field of computer vision (computer version, CV), the idea of ​​converting signals into images through methods such as cyclic spectrum analysis, constellation analysis, and time-frequency transformation, and further using neural networks to achieve recognition has been widely used. Benefiting from the essential characteristics of two-dimensional signals displayed by spectrum analysis and the rapid development of deep learning, new methods in the field of modulation recognition and radar signal recognition have been able to significantly improve the performance of existing algorithms and shorten the recognition time of signal samples (thanks to the acceleration of the calculation process by GPU parallel computing), thus having a broad and profound impact on future research and development in this field.

[0063] Multi-scale convolution is widely used in many information processing fields such as image processing because it can fully extract feature information of different scales and fully utilize the complementarity of features. However, while using multi-scale convolution to extract features brings performance improvements, it also leads to a significant increase in the computing resources it consumes. Considering that the dilated convolution can obtain a larger receptive field (Receptive Field) to ensure that the size of the output feature map remains unchanged, this means that the features of the image in a larger area can be obtained with as few parameters to be trained as possible. Therefore, multi-scale dilated convolution, which combines the advantages of both, came into being and achieved significant performance improvements in depth estimation, image correction, and image classification.

[0064] Inspired by human perception and visual systems, attention mechanisms are widely used in the field of computer vision. They selectively invest computing resources in the target area of ​​focus, thereby guiding the attention of the acquired deep features. Zhou et al. used a new attention mechanism to improve the U-shape network and proposed a hierarchical U-shape attention network (HUAN) with strong robustness and high performance, which can greatly reduce memory consumption and significantly improve mask quality. In order to further improve the feature guidance performance of the attention mechanism, Woo et al. proposed a lightweight and efficient convolutional block attention module (CBAM) to infer attention maps according to channel and spatial dimensions. This module can be embedded in any CNN architecture and has universality and broad application prospects. Based on CBAM, Zhao et al. proposed a complex convolutional block attention module (CCBAM) to construct more information features to improve the representation ability of the complex-valued convolutional layer. By adding an attention mechanism module at the appropriate position in the constructed deep learning framework, it is possible to effectively guide the correct attention on the extracted features and automatically highlight and strengthen the differences in identification features, thereby further improving the utilization of computing resources and recognition performance.

[0065] The following describes the STBC-OFDM signal blind recognition method and device based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application with reference to the accompanying drawings.

[0066] Figure 1 A flowchart of a method for blind recognition of STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum provided in Example 1 of the present application.

[0067] like Figure 1 As shown, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum includes the following steps:

[0068] Step 101, calculating the fourth-order lag moment of the received signal and generating a fourth-order lag moment vector;

[0069] Step 102, using two-dimensional vector concatenation to merge the fourth-order lag moment vectors into a fourth-order lag moment spectrum;

[0070] Step 103, construct an attention-guided multi-scale dilated convolutional network, take the fourth-order lagged moment spectrum as input, and output the recognition result, wherein the attention-guided multi-scale dilated convolutional network includes an attention-guided multi-scale dilated convolutional module, a feature fusion layer, and a residual layer, takes the fourth-order lagged moment spectrum as input, and outputs the recognition result, including inputting the fourth-order lagged moment spectrum into the attention-guided multi-scale dilated convolutional module, outputting multi-scale guided features, and then inputting the multi-scale guided features into the feature fusion layer and the residual layer, and outputting the recognition result via a fully connected layer with softmax as the activation function.

[0071] The STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum in the embodiment of the present application calculates the fourth-order lag moment of the received signal and generates a fourth-order lag moment vector; uses two-dimensional vector splicing to merge the fourth-order lag moment vector into a fourth-order lag moment spectrum; constructs an attention-guided multi-scale dilated convolution network, takes the fourth-order lag moment spectrum as input, and outputs a recognition result, wherein the attention-guided multi-scale dilated convolution network includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer, takes the fourth-order lag moment spectrum as input, and outputs a recognition result, including inputting the fourth-order lag moment spectrum into the attention-guided multi-scale dilated convolution module, outputting multi-scale guided features, and then inputting the multi-scale guided features into the feature fusion layer and the residual layer, and outputting the recognition result via a fully connected layer with softmax as the activation function. Therefore, it can solve the technical problems that the existing methods need to manually extract features and set test thresholds, their recognition performance is greatly affected by parameter selection, and their robustness to different communication environments is poor. At the same time, it can solve the technical problems that the existing methods cannot accurately and quickly identify signals when they are transmitted in a complex electromagnetic environment with a low signal-to-noise ratio and are greatly affected by noise and channel attenuation. It has excellent comprehensive recognition performance, a short recognition time for a single sample, and significantly improves the recognition accuracy under low signal-to-noise ratio, and has good adaptability to strong interference environments. In addition, the deep learning model can directly identify the FOLMS (fourth order lag moment spectrum) samples obtained by preprocessing, so this method does not require prior information such as channels and noise, and is more suitable for non-cooperative communication than existing algorithms.

[0072] This application introduces deep learning into the field of STBC-OFDM signal recognition, and proposes a recognition scheme based on a fourth-order lag moment spectrum and an attention-guided multi-scale dilated convolutional network. First, the fourth-order lag moment of the received signal is calculated and a fourth-order lag moment vector is generated. The generated vector is further merged into a fourth-order lag moment spectrum as the network input by two-dimensional vector splicing. Then, on the basis of fully extracting the detailed information of different scales of the image using multi-scale dilated convolution, the convolution block attention module is further introduced to construct an attention-guided multi-scale dilated convolution module, so that network resources are more concentrated on the target area that needs to be focused on. Finally, the multi-scale guided features are fused to enhance the complementarity of the network features, and the residual layer is added to further enhance the utilization and representation ability of the deep fusion features, and the recognition results are output through the fully connected layer with softmax as the activation function.

[0073] In order to introduce deep learning (DL) into STBC-OFDM recognition, this application designs an identification feature that can fully reflect the essential characteristics of the signal and is suitable for network training, and constructs a fourth-order lag moment spectrum as the input sample of the network to realize the transformation of the signal recognition problem into the image recognition problem. At the same time, in order to extract multi-scale guided features, an attention-guided multi-scale dilated convolution module is designed. Specifically, a multi-scale dilated convolution algorithm with different expansion rates is used to achieve a larger receptive field (Receptive Field) while keeping the size of the feature map unchanged, and at the same time, it can fully extract the detailed information of different scales of the image, avoiding the problem of image information loss caused by extracting only a single feature. Furthermore, the convolution block attention module is used to guide the extracted multi-scale features to obtain multi-scale guided features, which has a stronger feature representation capability and can more efficiently reflect the essential characteristics of the input fourth-order lag moment spectrum.

[0074] In order to make full use of the complementarity of different types of features, the multi-scale guided features are used to retain the detailed information of each channel image and improve the representation ability of the deep fusion features. The introduction of the residual layer enables the multi-scale fusion features to flow across layers. By fusing the detailed information of the shallow network and the high-dimensional features mapped by the deep network, the network structure is made more compact and easier to train, and the performance degradation caused by the increase in the number of network layers is effectively alleviated.

[0075] Consider a system with n t transmit antennas and n rIn a wireless communication system with STBC-OFDM and multiple receiving antennas, data symbols are first encoded by STBC (Space-time Block Code) to generate multiple data streams, and then they are respectively processed by inverse fast Fourier transform (IFFT) and adding cyclic prefix to realize the generation of OFDM (Orthogonal Frequency Division Multiplexing) modulation. Assuming that the length of the OFDM block is N, each OFDM block can be expressed as:

[0076]

[0077] in, Represents the j-th data symbol of the Ub+u-th OFDM block of the f-th antenna, U is the length of the coding matrix (for example, U=1 for SM code and U=2 for AL code), b and u represent the u-th OFDM block in the b-th coding matrix block, u=0,1,…,U-1.

[0078] Assumption d Xb+x Denotes the OFDM blocks transmitted in each STBC coding matrix C (for SM coding matrix X = n t , AL X=2, STBC3 X=3, STBC4 X=4), X is the number of OFDM blocks contained in each STBC coding matrix C, x is the sequence number of the OFDM block in each STBC coding matrix C, x=0,1,…,X-1. d Xb+x The data symbols in are uncorrelated, that is:

[0079] E[d Xb+x (k)d Xb+x (k′)]=0

[0080] This application considers the four types of signals that are most widely used in STBC-OFDM wireless communication systems in most existing algorithms, namely SM-OFDM, AL-OFDM, STBC3-OFDM and STBC4-OFDM.

[0081] Considering the SM-OFDM signal with two transmitting antennas, its coding matrix can be expressed as:

[0082]

[0083] The coding matrix of AL-OFDM signal can be expressed as:

[0084]

[0085] The coding matrix of STBC3-OFDM signal can be expressed as:

[0086]

[0087] The coding matrix of STBC4-OFDM signal can be expressed as:

[0088]

[0089] At the transmitter, for each OFDM block Perform an N-point inverse fast Fourier transform (N-IFFT) to obtain the OFDM block in the time domain:

[0090]

[0091] Going further, Add a cyclic prefix. Assuming the length of the cyclic prefix is ​​ν, the OFDM block after adding the cyclic prefix can be expressed as:

[0092]

[0093] The signal satisfies:

[0094]

[0095] Therefore, the signal sequence transmitted on the fth transmitting antenna can be expressed as:

[0096]

[0097] At the receiving end, assume that the kth signal in () is s (f) (k), the kth received signal of the i-th receiving antenna is:

[0098]

[0099] Among them, L h is the number of transmission paths, h fi (l) is the channel coefficient, w (i) (k) means the mean is 0 and the variance is Gaussian white noise, that is

[0100] In the field of STBC-OFDM recognition, most existing algorithms often use cyclic cross-correlation function (CCF) and other features as identification features, and compare them with appropriate thresholds to achieve STBC-OFDM signal recognition by constructing hypothesis testing methods. However, it is not feasible to directly use the identification features of traditional feature extraction algorithms as samples for deep learning algorithms for training. The main reasons for the difficulty of network training are as follows. First, the threshold set by the hypothesis test is a parameter rather than a vector, which means that the final identification feature has only one value. For neural networks, it is impossible to use this identification feature for training in terms of data volume and input sample dimension. On the other hand, if the final feature is not used as a training sample, and the intermediate feature of a certain step in obtaining the identification feature is used as the network input sample, it is also necessary to try to meet the requirement that the neural network input is often a two-dimensional matrix. After obtaining a two-dimensional identification feature with sufficient parameter volume as the network input, the identification of this complex signal can be achieved by constructing a corresponding suitable network. In fact, in the two sub-steps of generating FOLMS, generating a fourth-order lag moment vector realizes the role of obtaining intermediate features, and two-dimensional vector concatenation realizes the requirement of meeting the input sample dimension of the neural network.

[0101] Assume that the received signal on the i-th receiving antenna is R (i) , combined with formula (), we can know

[0102]

[0103] Among them, N b is the number of OFDM blocks received, represents the jth OFDM block received on the i-th receiving antenna

[0104]

[0105] Further, in the embodiment of the present application, the fourth-order lag moment of the received signal is calculated, and a fourth-order lag moment vector is generated, wherein the fourth-order lag moment is expressed as:

[0106]

[0107] Among them, for the i-th receiving antenna, N b OFDM blocks Define y(q,τ) to represent the fourth-order lag moment when the delay parameter is (0,τ,0,τ),

[0108] The fourth-order lag moment vector is expressed as:

[0109] V=[E[y(0,τ)],E[y(1,τ)],…,E[y(N s-1,τ)]]

[0110] Among them, E[y(N s-1 ,τ)] represents the fourth-order hysteresis moment y(N s-1 ,τ).

[0111] Taking AL-OFDM signal as an example, since STBC-OFDM signal uses different coding matrices when performing STBC coding, and and its internal elements are correlated, while the vectors and elements between different space-time block codes are uncorrelated, such as and The same rule exists for other STBC-OFDM signals. In addition, because the length of the coding matrix of different signals is different during STBC coding, the fourth-order lag moment vectors calculated for the four types of STBC-OFDM signals have different peak distributions, and this distribution characteristic can be used as a key identification feature.

[0112] The generated length is N s The fourth-order lag moment vector of has a regularly changing peak value, but because the peak distribution of the one-dimensional vector is relatively sparse, the training effect of directly inputting the fourth-order lag moment vector into the network is poor. At the same time, in the field of computer vision (Computer Version, CV), the image dimensions of the input layer often use a square structure with the same side length, and the side length is often a power of 2, so we use N s (When generating STBC-OFDM signal N s That is, the default selection is a power of 2) with a length of N s The fourth-order lag moment vectors of are taken as a group and merged into the fourth-order lag moment spectrum by two-dimensional vector concatenation.

[0113] Further, in the embodiment of the present application, the fourth-order lag moment spectrum is expressed as:

[0114]

[0115] in, represents the kth fourth-order lag moment vector.

[0116] kth column vector It can be expressed as:

[0117]

[0118] The fusion of multi-scale information and the mining of relationship information between different pixels in the image can effectively improve the utilization rate of input feature detail information and make full use of the complementarity between different types of features. The dilated convolution can ensure that the size of the output feature map remains unchanged while obtaining a larger receptive field, which means that the features of the image in a larger area can be obtained with as few parameters to be trained as possible. Therefore, the use of multi-scale dilated convolution can achieve the purpose of extracting FOLMS detail information and avoid the loss of image information caused by extracting only a single feature.

[0119] Specifically, when selecting the dilation rate of the multi-scale dilated convolution part, considering that the FOLMS of the input STBC-OFDM signal has peak-trough-peak and peak-trough-trough-trough-peak characteristic distributions (AL-OFDM signal and STBC3-OFDM signal), dilated convolutions with dilation rates of 2 and 3 are used to better extract the effective features of FOLMS. Assuming that the dilation rate of the dilated convolution is d, the receptive field size of the nth layer convolution kernel under this dilation rate is:

[0120]

[0121] Among them, RF n-1 represents the receptive field size of the n-1th layer convolution kernel, k n is the size of the convolution kernel in the nth layer, stride i Represents the step size of the i-th convolution kernel. From this formula, we can see that when the convolution kernel size and step size are constant, increasing the expansion rate can effectively expand the receptive field scale to obtain image feature information in a wider area. Furthermore, adding a standard convolution layer to further extract the deep features of the multi-scale feature map makes the features obtained by the network have stronger representation capabilities.

[0122] Furthermore, in an embodiment of the present application, the fourth-order lag moment spectrum is input into an attention-guided multi-scale dilated convolution module, and the multi-scale guided features are output, which includes the following steps:

[0123] The input fourth-order lag moment spectrum is subjected to multi-scale dilated convolution to obtain multi-scale features, and obtains targeted feature representations of the fourth-order lag moment spectra of different STBC-OFDM signals;

[0124] A standard convolutional layer is used to further extract deep features of the multi-scale feature map;

[0125] Finally, the multi-scale deep features are guided by the convolutional block attention module to generate multi-scale guided features.

[0126] Considering that the extracted fourth-order lag moment spectrum does not have peak features in all regions, for example, for STBC4-OFDM, the characteristic distribution of its FOLMS is trough-trough-trough-trough-peak-peak-peak-peak, and its features are concentrated in the peak area. Therefore, in order to make the network devote more attention resources to the target area that needs to be focused on, so as to obtain more detailed information of the target that needs to be focused on, and suppress other useless information, the convolutional block attention module is used to further improve the recognition performance of the network. Compared with other attention mechanism modules, such as the squeeze-excitation module, CBAM has shown significant excellent performance in both classification recognition (ImageNet dataset) and target detection problems (MS COCO dataset), and it can also be used in combination with any CNN structure, with good adaptability and generalization. CBAM consists of two sub-modules: channel attention module and spatial attention module.

[0127] Furthermore, in an embodiment of the present application, the multi-scale deep features are guided by the convolutional block attention module to generate multi-scale guided features, including:

[0128] Input the multi-scale deep features into the channel attention module to generate channel attention features;

[0129] The product of the channel attention feature and the multi-scale deep feature is used as the input of the spatial attention module, and the final multi-scale guided feature is output.

[0130] Furthermore, in the embodiment of the present application, the multi-scale deep features are input into the channel attention module, specifically:

[0131] The input feature map first undergoes global average pooling and global maximum pooling on the channel dimension to produce two channel attention maps;

[0132] The two generated channel attention maps are fed into a shared multilayer perceptron, and then the features output by the shared multilayer perceptron are summed up element-wise, and then subjected to a sigmoid activation operation to generate the final channel attention features.

[0133] For the channel attention module, the input feature map First, two channel attention maps are generated by global average-pooling and global max-pooling on the channel axis. and Then, the channel attention map M c1 and M c2 The shared multilayer perceptron is fed into the shared network. At the same time, in order to reduce the number of parameters, the size of the shared network is set to (r is the reduction ratio). Finally, the features output by the shared multi-layer perceptron (MLP) are summed up based on the element level, and then activated by sigmoid to generate the final channel attention feature, i.e., M c In summary, the calculation formula can be expressed as:

[0134]

[0135] in, and represents the shared network weights in the MLP, and represents the average pooling and maximum pooling features, and σ(·) represents the sigmoid activation function.

[0136] Furthermore, in the embodiment of the present application, the product of the channel attention feature and the multi-scale deep feature is used as the input of the spatial attention module to generate the final multi-scale guided feature, including:

[0137] Perform global average pooling and global maximum pooling on the spatial dimension of the product of the channel attention feature and the multi-scale deep feature to obtain the first feature map and the second feature map, and then perform a splicing operation on the first feature map and the second feature map;

[0138] The concatenated feature map is subjected to standard convolution operation and sigmoid activation function to generate spatial attention features;

[0139] Multiply the spatial attention feature and the input feature of the module to obtain the final output feature.

[0140] For the spatial attention module, the input of the module is the output of the channel attention module With CBAM input feature map The product of is expressed as:

[0141]

[0142] The feature map F′ is subjected to global average pooling and global maximum pooling on the spatial axis to obtain the feature map and And the feature map and Perform the splicing operation, expressed as:

[0143]

[0144] Then, the concatenated feature map F concat After a 7×7 dimensional standard convolution operation and a sigmoid activation function, the spatial attention feature is generated, which is expressed as:

[0145]

[0146] Finally, the space attention feature Multiply it with the input feature map F of the module to get the final output feature, expressed as:

[0147]

[0148] Using different fusion methods will have a certain impact on the fusion effect of multi-scale guided features. The splicing fusion method is used to fuse the output features of the convolutional block attention module. Although the spatial complexity of this method is higher than that of the add fusion method, the overall performance of the splicing fusion method is still better when considering the recognition performance and time complexity. After obtaining the splicing fusion features, the residual block is further used to weaken the performance saturation degradation problem caused by the increase in the number of network layers, while improving the utilization rate of the splicing fusion features. The output features of the residual block are:

[0149] y RB =MaxPool(f(Concat(F″ d1 +F″ d2 +F″ d3 )+F(Concat(F″ d1 +F″ d2 +F″ d3 ),W RB )))

[0150] Among them, F d1 , F″ d2 and F″ d3 is the multi-scale guided feature output by the three branches when the network expansion rates are 1, 2, and 3, respectively. Concat(·) represents concatenation fusion, F(·) is the residual mapping function, and f(·) is the ReLU (Rectified Linear Unit) activation function. The multi-scale guided features can be more fully utilized through feature fusion and residual blocks, so that feature information can flow across layers. By fusing the detailed information of the shallow network and the high-dimensional features mapped by the deep network, the network structure is more compact and easier to train.

[0151] Figure 2 This is a block diagram of an STBC-OFDM transmitter of an STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

[0152] like Figure 2 As shown in the figure, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum is considered to have n ttransmit antennas and n r In a wireless communication system with STBC-OFDM and multiple receiving antennas, data symbols are first encoded by STBC to generate multiple data streams, and then OFDM modulation is achieved by two steps: inverse fast Fourier transform (IFFT) and adding a cyclic prefix.

[0153] Figure 3 This is a structural diagram of the FOLMS / AMDC-net method of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum in an embodiment of the present application.

[0154] like Figure 3 As shown, the FOLMS / AMDC-net (attention-guided multi-scale dilated convolution network) scheme of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order delayed moment spectrum includes two main parts: fourth-order delayed moment spectrum for feature extraction and attention-guided multi-scale dilated convolution network for identifying STBC-OFDM type. Specifically, FOLMS includes generating fourth-order lagged moment vectors and two-dimensional vector concatenation. AMDC-net includes attention-guided multi-scale dilated convolution module, feature fusion module, residual block and fully connected layer with Softmax as activation function.

[0155] Figure 4 Schematic diagram of fourth-order delayed matrix spectra of four STBC-OFDM signals of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order delayed matrix spectrum in an embodiment of the present application.

[0156] like Figure 4 As shown, the number of received OFDM blocks N b =8000 as an example, the three-dimensional fourth-order lag moment spectrum under the signal-to-noise ratio of 10dB is obtained. It can be obviously seen from the figure that different STBC-OFDM signals have great differences, and their FOLMS peaks and troughs are arranged in an evenly spaced regular pattern. For example, the fourth-order lag moment spectrum of the AL-OFDM signal is distributed in any vector sequence as peak-trough-peak, while the vector sequence of the STBC3-OFDM signal is distributed as peak-trough-trough-trough-peak). These two points are of great significance for the deep learning algorithm to identify its signal type.

[0157] Figure 5 Schematic diagram of the overall structure of AMDC-net of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

[0158] like Figure 5 As shown in the figure, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum introduces deep learning into the field of STBC-OFDM signal recognition, and proposes a recognition scheme based on fourth-order lagged moment spectrum and attention-guided multi-scale dilated convolutional network. First, the fourth-order lagged moment of the received signal is calculated and the fourth-order lagged moment vector is generated. The generated vector is further merged into the fourth-order lagged moment spectrum as the network input by two-dimensional vector concatenation. Then, on the basis of fully extracting the detailed information of different scales of the image by using multi-scale dilated convolution, the convolution block attention module is further introduced to construct the attention-guided multi-scale dilated convolution module, so that the network resources are more concentrated on the target area that needs to be focused. Finally, the multi-scale guided features are fused to enhance the complementarity of the network features, and the residual layer is added to further enhance the utilization and representation ability of the deep fusion features, and the recognition results are output through the fully connected layer with softmax as the activation function.

[0159] Figure 6 Schematic diagram of multi-scale dilated convolution of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum in an embodiment of the present application.

[0160] like Figure 6 As shown in FIG, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum, when selecting the dilation rate of the multi-scale dilated convolution part, considering that the FOLMS of the input STBC-OFDM signal has the characteristic distribution of peak-trough-peak and peak-trough-trough-trough-peak (AL-OFDM signal and STBC3-OFDM signal), the dilated convolution with dilation rates of 2 and 3 is used to better extract the effective features of FOLMS. The obtained multi-scale dilated convolution structure is shown in FIG. Figure 6 shown.

[0161] Figure 7 Schematic diagram of the convolutional block attention module scheme of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application.

[0162] like Figure 7As shown in the figure, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum uses a convolutional block attention module to further improve the recognition performance of the network. Compared with other attention mechanism modules, such as the squeeze excitation module, CBAM has shown significant excellent performance in both classification recognition (ImageNet dataset) and target detection problems (MS COCO dataset), and it can also be used in combination with any CNN structure, with good adaptability and generalization. CBAM contains two submodules: channel attention module and spatial attention module.

[0163] Figure 8 Schematic diagram of the attention-guided multi-scale dilated convolution module of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application.

[0164] like Figure 8 As shown in the figure, this STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum uses a convolutional block attention module to guide the multi-scale features in order to focus more attention resources of the extracted multi-scale features on the target area that needs to be focused on and further improve the recognition performance of the network. An attention-guided multi-scale dilated convolution module is designed to further obtain multi-scale guided features with stronger feature mapping capabilities, which can better reflect the true characteristics of the fourth-order lagged moment spectrum features of the input STBC-OFDM signal. The structure of the attention-guided multi-scale dilated convolution module is shown in the figure. Figure 8 As shown in the figure, the input fourth-order lag moment spectrum features are multi-scaled after multi-scale dilated convolution to obtain multi-scale features, and obtain the fourth-order lag moment spectrum of different STBC-OFDM signals with peak difference and targeted feature representation. Then a standard convolution layer is used to further extract the deep features of the multi-scale feature map, so that the features obtained by the network have stronger representation capabilities. Finally, the multi-scale deep features are guided by the convolution block attention module to generate multi-scale guided features.

[0165] Fig. 9 This is a schematic diagram of the position of the attention-guided multi-scale dilated convolution module in the attention-guided multi-scale dilated convolution network of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application.

[0166] like Fig. 9As shown, the attention-guided multi-scale dilated convolution network of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer. The fourth-order lagged moment spectrum is input into the attention-guided multi-scale dilated convolution module, and multi-scale guided features are output, including the following steps: the input fourth-order lagged moment spectrum is subjected to multi-scale dilated convolution to obtain multi-scale features, and a targeted feature representation of the fourth-order lagged moment spectrum of different STBC-OFDM signals is obtained; a standard convolution layer is used to further extract the deep features of the multi-scale feature map; finally, the multi-scale deep features are guided by the convolution block attention module to generate multi-scale guided features, and then the multi-scale guided features are input into the feature fusion layer and the residual layer, and the recognition result is output through a fully connected layer with softmax as the activation function.

[0167] Fig.10 This is a structural schematic diagram of the feature fusion and residual block parts of the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum in an embodiment of the present application.

[0168] like Fig.10 As shown in the figure, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum uses the splicing fusion method to fuse the output features of the convolution block attention module. Although the spatial complexity of this method is higher than that of the addition fusion method, the overall performance of the splicing fusion method is still better when considering the recognition performance and time complexity. After obtaining the splicing fusion features, the residual block is further used to weaken the performance saturation degradation problem caused by the increase in the number of network layers, and at the same time, the utilization rate of the splicing fusion features can be improved.

[0169] Four types of STBC-OFDM signals are studied, namely SM-OFDM, AL-OFDM, STBC3-OFDM and STBC4-OFDM. The average recognition probability of the four types of signals is used as the measure of recognition performance.

[0170]

[0171] Wherein, C is the STBC-OFDM signal type, that is, C = {SM-OFDM, AL-OFDM, STBC3-OFDM, STBC4-OFDM}, Indicates that the signal sent is C i The recognition result is also C i The probability of a signal, P(C i ) is the sending signal C i Here, because the number of samples generated by the four types of signals is the same, the emission probabilities of the four types of signals all satisfy P(C i )=1 / 4.

[0172] This application uses matlab to generate STBC-OFDM signals and extract fourth-order lag moment spectrum features. The construction and training of the AMDC-net network rely on Keras based on the tensorflow backend, and the training process is accelerated by GPU. When generating STBC-OFDM signals, the number of subcarriers of the OFDM block is selected as N = 128, the cyclic prefix length is ν = 16, the modulation mode is quadrature phase shift keying quadrature PSK (Quadrature Phase Shift Keying, QPSK), and the frequency selective fading channel is used for simulation. The number of receiving antennas is N r = 1. The received signal is b = 8000 OFDM blocks are used as a group to calculate the fourth-order lag moment vector, and the length of the fourth-order lag moment vector is selected as N s =16. The signal-to-noise ratio range of the generated samples is -16dB to 16dB, with a step size of 2dB. A single STBC-OFDM signal generates 100 fourth-order lag moment spectrum input samples at each signal-to-noise ratio. The total sample capacity is 6800, and the ratio of training samples to verification samples is 8:2. In the network optimization process, this application uses the Adam optimizer to solve the optimal solution, and selects cross entropy as the loss function.

[0173] In order to verify the effectiveness of using multi-scale features in the model to identify STBC-OFDM signals, the AMDC-net model proposed in this application is compared with the attention-guided single-branch convolutional network (ASC-net) model. The parameter settings and basic structure of each layer of the ASC-net model used for performance comparison are consistent with those of the AMDC-net. The difference between the two is that the ASC-net only retains Figure 5 Only the upper branch of the standard convolution is used in , so there is no feature fusion part in this network.

[0174] Fig.11 Schematic diagram of performance comparison between AMDC-net and ASC-net according to an embodiment of the present application.

[0175] like Fig.11As shown in the figure, the performance comparison between the AMDC-net model and the ASC-net (attention-guided single-branch convolutional network) model shows that the AMDC-net model using multi-scale fusion features has significantly improved the performance of the ASC-net model at low signal-to-noise ratio, and the average recognition probability at -16dB has increased by 7.9% compared with the ASC-net. This shows that the multi-scale fusion features extracted by multi-scale dilated convolution effectively enrich the feature types and enhance the complementarity between different features, making the AMDC-net model have stronger representation capabilities, thereby verifying the effectiveness of using multi-scale features.

[0176] In order to further analyze the performance difference between AMDC-net and ASC-net under various code types, the confusion matrices of the two networks at -16dB, -14dB and -12dB for 4 types of STBC-OFDM signals are given.

[0177] Fig.12 Schematic diagram of the confusion matrix of AMDC-net and ASC-net according to an embodiment of the present application.

[0178] like Fig.12 As shown in the figure, from an overall observation, the main reason for the poor recognition performance of the deep learning model under low SNR is the confusion between SM-OFDM signals and AL-OFDM signals, which is true for both AMDC-net and ASC-net. The reason for this phenomenon may be that the peak value of the FOLMS of the AL-OFDM signal is low, which makes its peak feature buried in interference and noise under the harsh environment of low SNR. From a vertical perspective, it can be seen that under the same SNR condition, the true label (True label) of AMDC-net is more consistent with the predicted label (Predicted label), and its misjudgment rate is significantly lower than that of ASC-net. Observing and comparing from a horizontal perspective, the performance of the two networks increases steadily with the increase of the signal-to-noise ratio, especially at -12dB, the average recognition probability of the two networks for four types of STBC-OFDM signals reaches more than 94.5%, and the diagonal distribution of the confusion matrix is ​​obvious, indicating that AMDC-net and ASC-net have strong adaptability to low signal-to-noise ratio. It also verifies the rationality of the idea of ​​constructing a fourth-order lag moment spectrum in this application to introduce a deep learning model into STBC-OFDM signal recognition.

[0179] The effectiveness of the attention mechanism is verified by comparing four networks that contain components of the AMDC-net model: MDC-net+CBAM represents the original AMDC-net model; MDC-net represents the original attention-guided multi-scale dilated convolutional network without the convolutional block attention module, i.e., the multi-scale dilated convolutional network; SC-net+CBAM represents a single-branch convolutional network with a convolutional block attention module, which has the same structure as ASC-net; SC-net represents a single-branch convolutional network or CBAM without multi-scale features. The four networks are simulated under the same data set, sample ratio, batch size, and hardware environment, which can more accurately evaluate the effectiveness of the attention mechanism in the model.

[0180] Fig.13 Schematic diagram of performance comparison of four networks used for attention mechanism evaluation according to an embodiment of the present application.

[0181] like Fig.13 As shown, thanks to CBAM, the attention resources of the features are more concentrated on the target area that needs to be focused on. The performance of both SC-net and MDC-net after adding the attention mechanism has been improved to a certain extent, especially the average recognition probability of MDC-net has increased by 5.2% after adding CBAM, which shows that the use of the CBAM model to properly guide the extracted preliminary features can improve the network's ability to represent FOLMS features, thereby verifying the effectiveness of using the attention mechanism in AMDC-net. Furthermore, by observing the enlarged image of the average recognition probability of the four types of networks at -14dB, it is found that compared with the SC-net model that only uses a single branch to extract features, the MDC-net model that integrates multi-scale features has a more significant performance improvement after adding CBAM, that is, the introduction of fusion features further stimulates the effectiveness of the attention mechanism. This illustrates the rationality of combining the attention mechanism with multi-scale features in this application from the side. CBAM guides the attention of features at different scales to further improve the recognizability of the fusion features. The two promote each other and complement each other.

[0182] Fig.14 Schematic diagram of the confusion matrix of the four networks of the embodiment of the present application at -14 decibels.

[0183] like Fig.14As shown in the figure, in order to further analyze the difference in recognition accuracy of the four networks for different STBC-OFDM signals, the confusion matrix of the four networks at -14dB is given. By comparing the MDC-net+CBAM and MDC-net, it can be seen that the recognition accuracy of the MDC-net model for AL-OFDM signals has been significantly improved after adding CBAM, which has brought about a significant improvement in the overall recognition performance. This shows that the use of the attention mechanism for feature guidance can effectively alleviate the problem of AL-OFDM signal peak annihilation caused by noise interference. For the SC-net+CBAM model and the SC-net model that do not use multi-scale features, the recognition rates of SM-OFDM, STBC3-OFDM and STBC4-OFDM are all above 92.5%, but the misjudgment rate of AL-OFDM signals is high, and nearly 60% of AL-OFDM are mistakenly predicted as SM-OFDM. Therefore, further improving the recognition performance of the single-branch convolutional network for AL-OFDM signals will effectively improve the overall performance of the STBC-OFDM signal recognition method based on deep learning under low signal-to-noise ratio.

[0184] This application mainly considers the two methods of splicing and addition that are widely used in deep learning frameworks when selecting the multi-scale guided feature fusion method. In order to analyze the overall comprehensive performance of different methods, the two fusion methods are compared from the perspectives of recognition performance and computational complexity.

[0185] Fig.15 Schematic diagram of the recognition performance of the two fusion methods of the embodiments of the present application.

[0186] As shown in 15, overall, the performance of splicing fusion at low SNR is slightly better than addition, and the average recognition probability at -16dB and -14dB is increased by about 2.6%. The performance gain is gradually weakened as the SNR increases. In fact, this recognition advantage of splicing fusion comes from its retention of multi-scale guided feature components. Compared with additive fusion, splicing fusion completes the stacking of feature maps rather than superposition, which allows the channel dimensions of each multi-scale guided feature to be retained, rather than all guided features being concentrated in a certain feature map due to additive fusion. Therefore, the use of splicing fusion can make the deep detail information of FOLMS extracted by each attention-guided branch be fully utilized by the subsequent residual block, thereby further improving the recognition performance of AMDC-net.

[0187] In order to make a more comprehensive and in-depth comparison between the two fusion methods, the computational complexity of splicing fusion and additive fusion is analyzed, mainly from the perspective of the network's trainable parameters P train , the average training time per generation T train and the average test time T for a single sample testThree aspects are considered. It is worth noting that in order to compare the performance of connection and addition more fairly, the simulation experiment was completed under the same sample dimension, data set size, training / test ratio and training batch size. As shown in Table 1, although the trainable parameters of splicing fusion are twice that of additive fusion, thanks to the improvement of GPU parallel computing power, the training and testing time of the two fusion methods are very close, and the recognition time for a single sample is less than 0.2ms, which has strong real-time performance. Therefore, comprehensively comparing the recognition performance and computational complexity, the splicing fusion in the AMDC-net model has better comprehensive performance in recognition accuracy and efficiency.

[0188]

[0189] Table 1

[0190] Since there are few existing STBC-OFDM recognition algorithms, and most of them are traditional algorithms based on feature extraction, in order to further illustrate the advantages of the FOLMS / AMDC-net recognition scheme based on deep learning in identifying STBC-OFDM signals, AMDC-net is compared with CCF (Cross-Correlation Function), SOCS (Second-Order Cyclostationarity Statistics) and FOLP (Fourth Order Lag Product). Specifically, CCF uses the cross-correlation characteristics between the received signals from different antennas to identify the type of STBC-OFDM signals. SOCS relies on the cyclostationarity of the second-order signal and confirms the signal type by comparing the second-order cyclic statistics of the received signal with the threshold. FOLP identifies the type of transmitted STBC-OFDM signals based on the fourth-order lag moment of the received signal according to the correlation difference of different space-time block code coding matrices. The above algorithms are typical traditional algorithms based on hypothesis testing and statistical features, using decision trees and artificial thresholds to judge the signal type, and are representative in the field of STBC-OFDM signal recognition.

[0191] Fig.16 Schematic diagram for comparing the recognition performance of different methods in the embodiments of the present application.

[0192] like Fig.16As shown, the recognition performance of four types of methods, AMDC-net, CCF, SOCS and FOLP, is compared. The recognition performance of AMDC-net under low signal-to-noise ratio is significantly better than other traditional algorithms, and the extracted deep features have stronger recognition capabilities. This is because the deep learning model can automatically learn effective features that are more suitable for distinguishing STBC-OFDM signals, while the recognition performance of traditional algorithms depends largely on manually set thresholds. In the process of hypothesis testing, manually set parameters have poor adaptability to complex environments, and the extracted cumulative feature quantities may fluctuate greatly. This application avoids the problem that manually extracted features in traditional algorithms are greatly affected by noise interference by introducing deep learning methods into STBC-OFDM signal recognition, and further significantly improves the recognition performance of signals under low signal-to-noise ratios.

[0193] Simulation experiments show that the recognition performance of the FOLMS / AMDC-net (network) of this application has been significantly improved compared with the existing algorithms, and the performance gain is obvious under low signal-to-noise ratio. In addition, in order to verify the effectiveness of using multi-scale dilated convolution and convolution block attention modules, the proposed AMDC-net is compared with MDC-net, single-branch convolution network (SC-net) + CBAM and SC-net. Although the recognition performance of SC-net is the worst among the four, its recognition rate for SM-OFDM, STBC3-OFDM and STBC4-OFDM still reaches more than 92.5%. Therefore, the anti-interference ability of SC-net can be significantly improved by improving its performance for AL-OFDM.

[0194] Fig.17 A schematic diagram of the structure of a STBC-OFDM signal blind recognition device based on deep learning and fourth-order lag moment spectrum provided in Example 2 of the present application.

[0195] like Fig.17 As shown, the STBC-OFDM signal blind recognition device based on deep learning and fourth-order lag moment spectrum includes: a vector generation module, a splicing module, and a result generation module, wherein:

[0196] A vector generation module 10, used to calculate the fourth-order lag moment of the received signal and generate a fourth-order lag moment vector;

[0197] A splicing module 20, used for merging the fourth-order lag moment vector into a fourth-order lag moment spectrum by using two-dimensional vector splicing;

[0198] The result generation module 30 is used to construct an attention-guided multi-scale dilated convolutional network, which takes the fourth-order lagged matrix as input and outputs the recognition result. The attention-guided multi-scale dilated convolutional network includes an attention-guided multi-scale dilated convolutional module, a feature fusion layer, and a residual layer. It takes the fourth-order lagged matrix as input and outputs the recognition result, including: inputting the fourth-order lagged matrix into the attention-guided multi-scale dilated convolutional module, outputting the multi-scale guided feature, and then inputting the multi-scale guided feature into the feature fusion layer and the residual layer, and outputting the recognition result via a fully connected layer with softmax as the activation function.

[0199] The STBC-OFDM signal blind recognition device based on deep learning and fourth-order lagged moment spectrum of the embodiment of the present application includes: a vector generation module, a splicing module, and a result generation module, wherein the vector generation module is used to calculate the fourth-order lagged moment of the received signal and generate a fourth-order lagged moment vector; the splicing module is used to merge the fourth-order lagged moment vector into a fourth-order lagged moment spectrum by using two-dimensional vector splicing; the result generation module is used to construct an attention-guided multi-scale dilated convolution network, take the fourth-order lagged moment spectrum as input, and output the recognition result, wherein the attention-guided multi-scale dilated convolution network includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer, takes the fourth-order lagged moment spectrum as input, and outputs the recognition result, including: inputting the fourth-order lagged moment spectrum into the attention-guided multi-scale dilated convolution module, outputting multi-scale guided features, then inputting the multi-scale guided features into the feature fusion layer and the residual layer, and outputting the recognition result via a fully connected layer with softmax as the activation function. Therefore, it can solve the technical problems that the existing methods need to manually extract features and set test thresholds, their recognition performance is greatly affected by parameter selection, and their robustness to different communication environments is poor. At the same time, it can solve the technical problems that the existing methods cannot accurately and quickly identify signals when they are transmitted in a complex electromagnetic environment with a low signal-to-noise ratio and are greatly affected by noise and channel attenuation. It has excellent comprehensive recognition performance, a short recognition time for a single sample, and significantly improves the recognition accuracy under low signal-to-noise ratio, and has good adaptability to strong interference environments. In addition, the deep learning model can directly identify the FOLMS samples obtained by preprocessing, so this method does not require prior information such as channels and noise, and is more suitable for non-cooperative communication than existing algorithms.

[0200] In order to implement the above-mentioned embodiments, the present application also proposes a non-temporary computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lagged moment spectrum of the above-mentioned embodiment is implemented.

[0201] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0202] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0203] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0204] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0205] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0206] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0207] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0208] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A blind recognition method for STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum, characterized in that: The following steps are involved: Calculate the fourth-order lag moment of the received signal and generate a fourth-order lag moment vector; Using two-dimensional vector concatenation to merge the fourth-order lag moment vectors into a fourth-order lag moment spectrum; Construct an attention-guided multi-scale dilated convolutional network, take the fourth-order lag moment spectrum as input, and output the recognition result, where: The attention-guided multi-scale dilated convolution network includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer. The fourth-order lag moment spectrum is used as input and the recognition result is output, including: Input the fourth-order lagged moment spectrum into the attention-guided multi-scale dilated convolution module, output the multi-scale guided features, then input the multi-scale guided features into the feature fusion layer and the residual layer, and output the recognition result through the fully connected layer with softmax as the activation function; The fourth-order lag moment of the received signal is calculated and a fourth-order lag moment vector is generated, wherein the fourth-order lag moment is expressed as: Among them, for the The receiving antenna OFDM blocks ,definition Indicates that when the delay parameter is The fourth-order lag moment at The fourth-order lag moment vector is expressed as: in, represents the fourth-order lag moment The mean of The fourth-order lag moment spectrum is expressed as: in, Indicates A fourth-order lag moment vector.

2. The method according to claim 1, characterized in that Inputting the fourth-order lag moment spectrum into the attention-guided multi-scale dilated convolution module and outputting the multi-scale guided features comprises the following steps: The input fourth-order lag moment spectrum is subjected to multi-scale dilated convolution to obtain multi-scale features, and obtains targeted feature representations of the fourth-order lag moment spectra of different STBC-OFDM signals; A standard convolutional layer is used to further extract deep features of the multi-scale feature map; Finally, the multi-scale deep features are guided by the convolutional block attention module to generate multi-scale guided features.

3. The method according to claim 2, characterized in that The step of guiding the multi-scale deep features via the convolutional block attention module to generate multi-scale guided features includes: Inputting the multi-scale deep features into a channel attention module to generate channel attention features; The product of the channel attention feature and the multi-scale deep feature is used as the input of the spatial attention module, and the final multi-scale guided feature is output.

4. The method according to claim 3, characterized in that Inputting the multi-scale deep features into the channel attention module comprises: The input feature map first undergoes global average pooling and global maximum pooling on the channel dimension to produce two channel attention maps; The two generated channel attention maps are fed into a shared multilayer perceptron, and then the features output by the shared multilayer perceptron are summed up element-wise and then activated by a sigmoid operation to generate the final channel attention features.

5. The method according to claim 3, characterized in that The product of the channel attention feature and the multi-scale deep feature is used as the input of the spatial attention module to generate the final multi-scale guided feature, including: Performing global average pooling and global maximum pooling on the spatial dimension of the product of the channel attention feature and the multi-scale deep feature to obtain a first feature map and a second feature map, and performing a splicing operation on the first feature map and the second feature map; The concatenated feature map is subjected to standard convolution operation and sigmoid activation function to generate spatial attention features; Multiply the spatial attention feature and the input feature of the module to obtain the final output feature.

6. A blind recognition device for STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum, characterized in that: It includes vector generation module, splicing module and result generation module, among which, The vector generation module is used to calculate the fourth-order lag moment of the received signal and generate a fourth-order lag moment vector; The splicing module is used to merge the fourth-order lag moment vectors into a fourth-order lag moment spectrum by using two-dimensional vector splicing; The result generation module is used to construct an attention-guided multi-scale dilated convolutional network, take the fourth-order lag moment spectrum as input, and output the recognition result, wherein: The attention-guided multi-scale dilated convolution network includes an attention-guided multi-scale dilated convolution module, a feature fusion layer, and a residual layer. The fourth-order lag moment spectrum is used as input and the recognition result is output, which is specifically: Input the fourth-order lagged moment spectrum into the attention-guided multi-scale dilated convolution module, output the multi-scale guided features, then input the multi-scale guided features into the feature fusion layer and the residual layer, and output the recognition result through the fully connected layer with softmax as the activation function; The fourth-order lag moment of the received signal is calculated and a fourth-order lag moment vector is generated, wherein the fourth-order lag moment is expressed as: Among them, for the The receiving antenna OFDM blocks ,definition Indicates that when the delay parameter is The fourth-order lag moment at The fourth-order lag moment vector is expressed as: in, represents the fourth-order lag moment The mean of The fourth-order lag moment spectrum is expressed as: in, Indicates A fourth-order lag moment vector.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images