CN114298086A

CN114298086A - STBC-OFDM signal blind identification method and device based on deep learning and fourth-order lag moment spectrum

Info

Publication number: CN114298086A
Application number: CN202111442156.1A
Authority: CN
Inventors: 闫文君; 张聿远; 凌青; 张立民; 王程昱
Original assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Current assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Priority date: 2021-11-30
Filing date: 2021-11-30
Publication date: 2022-04-08

Abstract

The application provides an STBC-OFDM signal blind identification method based on deep learning and a fourth-order lag moment spectrum, which relates to the technical field of signal blind identification and comprises the following steps: calculating a fourth-order lag moment of the received signal and generating a fourth-order lag moment vector; combining the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing; and constructing an attention-guided multi-scale expansion convolutional network, wherein the attention-guided multi-scale expansion convolutional network comprises an attention-guided multi-scale expansion convolutional module, a feature fusion layer and a residual layer, inputting a fourth-order lag moment spectrum into the attention-guided multi-scale expansion convolutional module, outputting multi-scale guide features, inputting the multi-scale guide features into the feature fusion layer and the residual layer, and outputting a recognition result through a full connection layer with softmax as an activation function. The identification performance is remarkably improved, the method and the device have good adaptability to a strong interference environment, do not need prior information such as channels and noise, and are more suitable for non-cooperative communication compared with the existing algorithm.

Description

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

Technical Field

The application relates to the technical field of signal blind identification, in particular to a STBC-OFDM signal blind identification method and device based on deep learning and fourth-order lag moment spectrum.

Background

Blind Signal Identification (BSI) has wide application in many military and civilian fields, including radio monitoring, communication reconnaissance, electromagnetic countermeasure, and spectrum sensing. For example, when a transmitter signal is detected, it is often necessary to identify the signal type from the received data under the condition that a priori information such as channel and noise is unknown, so as to further decode and recover the original information of the transmitting end. With the shortage of spectrum resources, the combination of Space-time Block Code (STBC) and Orthogonal Frequency Division Multiplexing (OFDM) technology has been more and more widely applied due to the advantages of high Frequency band utilization and strong anti-multipath interference capability. As an important channel coding mode, the blind identification of the STBC-OFDM signal is deeply researched under the non-cooperative condition, the application range of the STBC-OFDM signal in the field of radio communication is further expanded, and the method has wide and profound significance.

Most of the existing algorithms are traditional identification methods based on feature extraction, the performance of the existing algorithms depends on professional knowledge and experience, more prior information is often needed, the limiting conditions on channels and noise are more, and the existing algorithms are not suitable for signal identification under the non-cooperative communication condition. According to the correlation difference of the STBC-OFDM signal coding mode, the traditional algorithm further establishes a decision tree and performs hypothesis test item by acquiring statistical characteristic quantity capable of reflecting the essential characteristics of the signal to complete recognition. Based on the space-time redundancy of the signals, the cross-correlation function between different receiving antennas can be calculated and taken as the identification feature, so that the effective identification of the SM-OFDM (Special Multiplexing-Orthogonal Frequency Division Multiplexing) and AL-OFDM (Alamouti-Orthogonal Frequency Division Multiplexing) signals can be realized. Based on the second order signal cyclostationarity, the signal type may be confirmed by comparing the second order cyclic statistic of the received signal to a threshold. According to the correlation difference of different coding matrixes, the type of the transmitted STBC-OFDM signal can be identified by calculating the fourth-order lag moment of the received signal. The traditional algorithm has been studied more deeply, but because it needs to artificially extract features and set inspection threshold, the problems of difficult feature selection, complex parameter adjusting process, sensitivity to noise and the like still exist, and the adaptability to the uncooperative communication situation with insufficient prior information is poor.

In recent years, with the rapid development of deep learning technology in the field of Computer Vision (CV), due to the improvement of the parallel computing capability of the GPU, the deep learning model obtains stronger classification performance with its strong mapping capability on mass data. By combining the performance advantages of the technology, scholars at home and abroad gradually apply the technology to the field of BSI, particularly in the fields of modulation identification, radiation source individual identification and radar signal identification, and the deep learning method has been widely applied. Compared with visual and visual characteristics of a visual image, the inherent characteristics of signals are often more difficult to directly discover by a deep learning model, so that the received signals need to be preprocessed firstly by a plurality of algorithms, and after the characteristics of a constellation diagram, a cyclic spectrum, a time-frequency image and the like are obtained, one or more of the received signals are used as the input of the deep learning model and integrated learning is combined to realize training and recognition. The successful application of the methods explains the feasibility of introducing deep learning into STBC-OFDM identification, but the coding mode of the STBC-OFDM signal is complex, and the STBC-OFDM signal cannot be effectively identified only by adopting a simple transformation mode or transferring signal transformation methods (constellation diagrams, cyclic spectrums and the like) in other fields.

In addition, with the increasing density and complexity of electromagnetic environments and informatization technologies, the influence of channel conditions and noise interference on signal transmission is increasingly greater, and the traditional STBC-OFDM signal identification method based on feature extraction and threshold decision cannot meet the actual requirement of accurate and rapid identification in the actual communication environment, so that the adaptability of the algorithm to low signal-to-noise ratio is urgently needed to be further improved. In addition, the existing method needs to artificially extract features and set a detection threshold, the identification performance of the method is greatly influenced by parameter selection, the judgment rule needs to be reset after the channel and noise are changed, and the robustness to different communication environments is poor. Considering that in the actual communication process, signals often need to be transmitted in a complex electromagnetic environment with a low signal-to-noise ratio, and the signals are greatly affected by noise and channel attenuation, how to correctly identify the type of the transmitter signal from a strong interference environment needs to be solved.

Disclosure of Invention

The present application is directed to solving, at least to some extent, one of the technical problems in the related art.

Therefore, a first objective of the present application is to provide a STBC-OFDM signal blind identification method based on deep learning and a fourth-order lag moment spectrum, which solves the technical problems that the existing method needs to artificially extract features and set inspection thresholds, the identification performance of the existing method is greatly influenced by parameter selection, and the robustness to different communication environments is poor, and simultaneously solves the technical problems that signals cannot be accurately and quickly identified due to the fact that the existing method is greatly influenced by noise and channel attenuation when being transmitted in a complex electromagnetic environment with a low signal-to-noise ratio. In addition, the deep learning model can directly identify FOLMS (fourth-order lag moment spectrum) samples obtained by preprocessing, so that the method does not need prior information such as channels and noises, and is more suitable for uncooperative communication than the existing algorithm.

The second purpose of the application is to provide an STBC-OFDM signal blind identification device based on deep learning and fourth-order lag moment spectrum.

A third object of the present application is to propose a non-transitory computer-readable storage medium.

In order to achieve the above object, an embodiment of the first aspect of the present application provides a method for blind identification of STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum, including: calculating a fourth-order lag moment of the received signal and generating a fourth-order lag moment vector; combining the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing; constructing an attention-guided multi-scale expansion convolution network, taking a fourth-order lag matrix spectrum as input, and outputting an identification result, wherein the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual error layer, and the attention-guided multi-scale expansion convolution network takes the fourth-order lag matrix spectrum as input and outputs the identification result, and comprises the following steps: inputting the fourth-order lag moment spectrum into an attention-guided multi-scale expansion convolution module, outputting multi-scale guide features, inputting the multi-scale guide features into a feature fusion layer and a residual error layer, and outputting a recognition result through a full connection layer with softmax as an activation function.

Optionally, in an embodiment of the present application, fourth order lag moments of the received signal are calculated, and a fourth order lag moment vector is generated, wherein the fourth order lag moments are expressed as:

wherein for the ith receiving antenna, the number of the receiving antennas is increased by N_bReceived signal sequence composed of OFDM blocks

Definition y (q, τ) denotes the fourth order lag moment at a delay parameter of (0, τ,0, τ),

the fourth order lag moment vector is expressed as:

V＝[E[y(0,τ)],E[y(1,τ)],…,E[y(N_s-1,τ)]]

wherein, E [ y (N)_s-1,τ)]Representing the fourth order lag moment y (N)_s-1τ) of the two.

Optionally, in one embodiment of the present application, the fourth order lag moment spectrum, is represented as:

wherein,

representing the kth fourth order lag moment vector.

Optionally, in an embodiment of the present application, inputting the fourth-order lag moment spectrum into the attention-guided multi-scale dilation convolution module, and outputting the multi-scale guided feature, includes the following steps:

obtaining multi-scale characteristics after the input fourth-order lag moment spectrum is subjected to multi-scale expansion convolution, and obtaining characteristic representation with pertinence to the fourth-order lag moment spectrum of different STBC-OFDM signals;

further extracting deep features of the multi-scale feature map through a standard convolutional layer;

and finally, guiding the multi-scale deep features through a convolution block attention module to generate multi-scale guide features.

Optionally, in an embodiment of the present application, guiding the multi-scale deep features via a convolution block attention module to generate the multi-scale guiding features includes:

inputting the multi-scale deep features into a channel attention module to generate channel attention features;

and taking the product of the channel attention feature and the multi-scale deep feature as the input of the spatial attention module, and outputting the final multi-scale guide feature.

Optionally, in an embodiment of the present application, entering the multi-scale deep features into a channel attention module includes:

inputting a feature map, and generating two channel attention maps by global average pooling and global maximum pooling of channel dimensions;

and sending the generated two channel attention maps into a shared multilayer perceptron, then carrying out element-level-based addition operation on the characteristics output by the shared multilayer perceptron, and then carrying out sigmoid activation operation to generate final channel attention characteristics.

Optionally, in an embodiment of the present application, taking a product of the channel attention feature and the multi-scale deep feature as an input of the spatial attention module, generating a final multi-scale guiding feature includes:

respectively carrying out global average pooling and global maximum pooling on spatial dimensions on 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 carrying out splicing operation on the first feature map and the second feature map;

generating space attention characteristics by the spliced characteristic graph after standard convolution operation and sigmoid activation function;

and multiplying the space attention characteristic and the input characteristic of the module to obtain the final output characteristic.

In order to achieve the above object, a second aspect of the present application provides an apparatus for blind identification of STBC-OFDM signal based on deep learning and fourth-order lag moment spectrum, including: a vector generation module, a splicing module and a result generation module, wherein,

the vector generation module is used for calculating a fourth-order lag moment of the received signal and generating a fourth-order lag moment vector;

the splicing module is used for merging the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing;

and the result generation module is used for constructing the attention-guided multi-scale expansion convolution network, taking the fourth-order lag moment spectrum as input and outputting a recognition result, wherein,

the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual error layer, a fourth-order lag moment spectrum is used as input, and an identification result is output, and the method specifically comprises the following steps:

inputting the fourth-order lag moment spectrum into an attention-guided multi-scale expansion convolution module, outputting multi-scale guide features, inputting the multi-scale guide features into a feature fusion layer and a residual error layer, and outputting a recognition result through a full connection layer with softmax as an activation function.

In order to achieve the above object, a non-transitory computer-readable storage medium is provided in a third aspect of the present application, and when executed by a processor, the instructions in the storage medium can execute a STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum.

The STBC-OFDM signal blind identification method based on deep learning and the fourth-order lag moment spectrum, the STBC-OFDM signal blind identification device based on deep learning and the fourth-order lag moment spectrum and the non-transitory computer readable storage medium solve the technical problems that in the existing method, characteristics need to be artificially extracted and an inspection threshold needs to be set, the identification performance of the existing method is greatly influenced by parameter selection, and robustness to different communication environments is poor, and meanwhile, the technical problems that in the existing method, signals are greatly influenced by noise and channel attenuation when being transmitted in a complex electromagnetic environment with a low signal to noise ratio and cannot be accurately and quickly identified are solved. In addition, the FOLMS sample obtained by preprocessing can be directly identified by the deep learning model, so that the method does not need prior information such as channels, noise and the like, and is more suitable for non-cooperative communication compared with the conventional algorithm.

Additional aspects and advantages of the present application will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the present application.

Drawings

The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

fig. 1 is a flowchart of a STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

FIG. 2 is a block diagram of an STBC-OFDM transmitter of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

FIG. 3 is a schematic structural diagram of the FOLMS/AMDC-net method of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to the embodiment of the present application;

FIG. 4 is a schematic diagram of the fourth order delay moment spectrum of 4 STBC-OFDM signals of the STBC-OFDM signal blind identification method based on deep learning and the fourth order delay moment spectrum according to the embodiment of the present application;

FIG. 5 is a schematic diagram of an overall structure of AMDC-net of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

FIG. 6 is a schematic diagram of the multi-scale dilation convolution of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to the embodiment of the present application;

FIG. 7 is a schematic diagram of a convolution block attention module scheme of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

FIG. 8 is a schematic diagram of an attention-directed multi-scale dilation convolution module of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

FIG. 9 is a schematic diagram of the position of an attention-directed multi-scale dilation convolution module of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum in an attention-directed multi-scale dilation convolution network according to an embodiment of the present application;

FIG. 10 is a schematic structural diagram of a feature fusion and residual block portion of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application;

FIG. 11 is a graph showing a comparison of the performance of AMDC-net and ASC-net for multi-scale feature evaluation according to an embodiment of the present application;

FIG. 12 is a schematic diagram of AMDC-net and ASC-net confusion matrices under different environments according to an embodiment of the present application;

FIG. 13 is a diagram illustrating performance comparison of four networks for attention mechanism evaluation according to an embodiment of the present application;

FIG. 14 is a schematic diagram of confusion matrices at-14 dB for four networks according to an embodiment of the present application;

FIG. 15 is a schematic diagram of the recognition performance of two fusion methods according to the embodiment of the present application;

FIG. 16 is a graph illustrating a comparison of recognition performance of various methods according to embodiments of the present application;

fig. 17 is a schematic structural diagram of an STBC-OFDM signal blind identification apparatus based on deep learning and fourth-order lag moment spectrum according to the second embodiment of the present application.

Detailed Description

Reference will now be made in detail to the embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are exemplary and intended to be used for explaining the present application and should not be construed as limiting the present application.

As an important feature extraction and transformation method, spectral analysis has very wide application in the field of signal processing and recognition. With the rise of deep learning in the field of Computer Vision (CV), the idea of converting signals into images by methods such as cyclic spectrum analysis, constellation analysis, time-frequency transformation and the like and further realizing recognition by using a neural network is widely applied. With the benefit of the rapid development of the two-dimensional signal essential characteristics and the deep learning exhibited by the spectrum analysis, the new method in the field of modulation recognition and radar signal recognition can significantly improve the performance of the existing algorithm and compress the recognition time of the signal sample (benefiting from the acceleration of the GPU parallel operation on the calculation process), thereby widely and deeply influencing the future research and development of the field.

The multi-scale convolution can fully extract feature information of different scales and fully utilize the advantage of feature complementarity, so that the multi-scale convolution is widely applied to various information processing fields such as image processing and the like. However, the performance is improved and the consumed computing resources are greatly increased by adopting the multi-scale convolution to extract the features, and considering that the expanded convolution can ensure that the size of the output feature map is kept unchanged when a larger Receptive Field (received Field) is obtained, which means that the features of the image in a larger area can be obtained under the condition that the parameters to be trained are as few as possible. Therefore, the multi-scale expansion convolution combining the advantages of the two occurs, and remarkable performance improvement is achieved in the problems of depth estimation, image correction, image classification and the like.

Inspired by human perception and visual systems, attention mechanism is widely applied in the field of computer vision, and acquired deep features are subjected to attention guidance by selectively putting computing resources into target areas with important attention. Zhou et al improve the U-shape Network by using a new Attention mechanism, and propose a Hierarchical U-shape Attention Network (HUAN) with strong robustness and high efficiency, which can greatly reduce memory consumption and significantly improve mask quality. To further improve the feature guidance performance of the Attention mechanism, Woo et al propose a lightweight and efficient Convolutional Block Attention Module (CBAM) to infer Attention in terms of channel and spatial dimensions, which can be embedded in any CNN architecture, with universality and broad application prospects. Based on CBAM, zhao et al propose a Complex Convolutional Block Attention Module (CCBAM) to construct more information features to improve the representation capability of complex convolutional layer. By adding the attention mechanism module at a proper position in the constructed deep learning framework, the extracted features can be effectively and correctly subjected to attention guidance, and the feature difference can be automatically highlighted and strengthened, so that the utilization rate and the identification performance of computing resources are further improved.

The method and the device for blind identification of STBC-OFDM signals based on deep learning and fourth-order lag moment spectrum according to the embodiment of the application are described below with reference to the accompanying drawings.

Fig. 1 is a flowchart of a STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

As shown in fig. 1, the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum includes the following steps:

step 101, calculating a fourth-order lag moment of a received signal and generating a fourth-order lag moment vector;

step 102, merging the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing;

and 103, constructing an attention-guided multi-scale expansion convolution network, taking a fourth-order lag moment spectrum as input, and outputting an identification result, wherein the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual error layer, the fourth-order lag moment spectrum is taken as input, and the identification result is output.

According to the STBC-OFDM signal blind identification method based on deep learning and the fourth-order lag moment spectrum, the fourth-order lag moment of a received signal is calculated, and a fourth-order lag moment vector is generated; combining the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing; the method comprises the steps of constructing an attention-guided multi-scale expansion convolution network, taking a fourth-order lag moment spectrum as input, and outputting a recognition result, wherein the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual layer, taking the fourth-order lag moment spectrum as input, and outputting the recognition result. Therefore, the technical problems that the recognition performance of the existing method needs to be greatly influenced by parameter selection and is poor in robustness to different communication environments due to the fact that the existing method needs to artificially extract features and set a detection threshold value can be solved, meanwhile, the technical problems that signals are greatly influenced by noise and channel attenuation when being transmitted in a complex electromagnetic environment with a low signal-to-noise ratio and cannot be accurately and quickly recognized in the existing method can be solved, the method has excellent comprehensive recognition performance, is short in recognition time of a single sample, remarkably improves recognition accuracy under the low signal-to-noise ratio, and has good adaptability to a strong interference environment. In addition, the deep learning model can directly identify FOLMS (fourth-order lag moment spectrum) samples obtained by preprocessing, so that the method does not need prior information such as channels and noises, and is more suitable for uncooperative communication than the existing algorithm.

The application introduces the field of STBC-OFDM signal identification in deep learning, and provides an identification scheme of a multi-scale expansion convolution network based on a fourth-order lag moment spectrum and attention guidance. Firstly, fourth-order lag moment of a received signal is calculated, fourth-order lag moment vectors are generated, and the generated vectors are combined into a fourth-order lag moment spectrum to serve as network input by further adopting two-dimensional vector splicing. Then, on the basis of fully extracting detail information of different scales of the image by utilizing multi-scale expansion convolution, a convolution block attention module is further introduced to construct an attention-guided multi-scale expansion convolution module, so that network resources are more concentrated on a target area needing important attention. And finally, fusing the multi-scale guide features to enhance the complementarity of the network features, adding a residual error layer to further enhance the utilization rate and the characterization capability of the deep fusion features, and outputting the recognition result through a full connection layer with softmax as an activation function.

In order to introduce Deep Learning (DL) into STBC-OFDM recognition, the method designs an identification characteristic which can fully reflect signal essential characteristics and is suitable for network training, and constructs a fourth-order lag moment spectrum as an input sample of a network so as to realize the conversion of a signal recognition problem to an image recognition problem. Meanwhile, in order to extract multi-scale guiding features, an attention-guiding multi-scale expansion convolution module is designed. Specifically, a multi-scale expansion convolution algorithm with different expansion rates is adopted to obtain a larger Receptive Field (received Field) under the condition that the size of a feature map is not changed, and meanwhile, detail information of images with different scales can be fully extracted, so that the problem of image information loss caused by extraction of a single feature is avoided. Furthermore, the convolution block attention module is adopted to guide the extracted multi-scale features to obtain the multi-scale guide features, so that the feature representation capability is stronger, and the essential characteristics of the input fourth-order lag moment spectrum can be more effectively reflected.

In order to fully utilize the complementarity of different types of features, the multi-scale guide features are processed to reserve the detail information of each channel image and improve the characterization capability of deep fusion features. The introduction of the residual error layer enables the multi-scale fusion features to realize cross-layer flow, and through the fusion of the detail information of the shallow network and the high-dimensional features of the deep network mapping, the structure of the network is more compact and easy to train, and the problem of performance degradation caused by the increase of the number of network layers is effectively relieved.

Consider having n_tA transmitting antenna and n_rIn an STBC-OFDM wireless communication system with receiving antennas, data symbols are first encoded by STBC (Space-time Block Code) to generate a multi-channel data stream, and then are subjected to Inverse Fast Fourier Transform (IFFT) and cyclic prefix (cyclic prefix) addition to generate OFDM (Orthogonal Frequency Division Multiplexing) modulation. Assuming that the length of an OFDM block is N, each OFDM block can be expressed as:

wherein,

the j data symbols of the U + U OFDM block representing the f antenna, where U is the length of the coding matrix (for example, U is 1 for SM code and 2 for AL code), b and U represent the U OFDM block in the b coding matrix block, and U is 0,1, …, and U-1.

Suppose d_Xb+xRepresenting the OFDM blocks transmitted in each STBC coding matrix C (X n for SM coding matrix_tWhere AL X is 2, STBC 3X is 3, and STBC 4X is 4), X is the number of OFDM blocks included in each STBC coding matrix C, X is the sequence number of OFDM blocks in each STBC coding matrix C, and X is 0,1, …, and X-1. d_Xb+xIs uncorrelated between the data symbols in (1), i.e.:

E[d_Xb+x(k)d_Xb+x(k′)]＝0

the present application considers the class 4 signals most widely used in STBC-OFDM wireless communication systems in most existing algorithms, namely SM-OFDM, AL-OFDM, STBC3-OFDM and STBC 4-OFDM.

Considering an SM-OFDM signal using 2 transmit antennas, its coding matrix can be represented as:

the coding matrix for an AL-OFDM signal may be represented as:

the coding matrix for the STBC3-OFDM signal may be represented as:

the coding matrix for the STBC4-OFDM signal may be represented as:

at the transmitting end, for each OFDM block

An N-point inverse fast fourier transform (N-IFFT) is performed to obtain OFDM blocks in the time domain:

still further, a pair

Adding a cyclic prefix, and assuming that the length of the cyclic prefix is v, the OFDM block after adding the cyclic prefix can be represented as:

wherein the signals satisfy:

therefore, the signal sequence transmitted on the f-th transmitting antenna can be represented as:

at the receiving end, let the k-th signal in () be s^(f)(k) The kth received signal of the ith receiving antenna is:

wherein L is_hFor the number of transmission paths, h_fi(l) Is a channel coefficient, w⁽ⁱ⁾(k) Represents a mean of 0 and a variance of

White Gaussian noise, i.e.

In the field of STBC-OFDM identification, most of existing algorithms usually adopt a circular cross-correlation function (CCF) and the like as identification features, compare the identification features with proper thresholds, and realize identification of STBC-OFDM signals by constructing a hypothesis testing method. However, it is not feasible to train the identification features of the conventional feature extraction algorithm directly as samples of the deep learning algorithm, so that the network training difficulty is mainly due to the following two reasons. One is to assume that the threshold set by the test is a parameter rather than a vector, which means that the final identification feature has only one value, and for the neural network, training using this identification feature is not possible from the viewpoint of data volume and input sample dimension. On the other hand, if the final features are not adopted as training samples, intermediate features of a certain step in the acquisition of the identification features are used as network input samples, the requirement that the neural network input is often a two-dimensional matrix needs to be met as much as possible, and after the two-dimensional identification features with enough parameters are acquired as the network input, the identification of the complex signal can be realized by constructing a suitable network corresponding to the two-dimensional identification features. In fact, in the two substeps of producing the foms, generating the fourth-order lag moment vector achieves the function of obtaining the intermediate features, and the two-dimensional vector stitching achieves the requirement of meeting the dimension of the input sample of the neural network.

Let the received signal on the ith receiving antenna be R⁽ⁱ⁾The combined formula () shows

Wherein N is_bFor the number of OFDM blocks received,

indicating the jth OFDM block received on the ith receiving antenna

Further, in the embodiment of the present application, fourth order lag moments of the received signals are calculated, and a fourth order lag moment vector is generated, wherein the fourth order lag moments are expressed as:

the fourth order lag moment vector is expressed as:

V＝[E[y(0,τ)],E[y(1,τ)],…,E[y(N_s-1,τ)]]

Taking AL-OFDM signal as an example, because STBC-OFDM signal adopts different coding matrixes when STBC coding is carried out, the same space-time block code vector

And

and its internal elements are correlated, while vectors and elements between different space-time block codes are uncorrelated, e.g.

And

this rule also exists for other STBC-OFDM signals. In addition, because the coding matrix lengths of different signals are different when performing STBC coding, the fourth-order lag moment vectors calculated from the class 4 STBC-OFDM signals have different peak distributions, and this distribution characteristic can be used as a key identification feature.

Generated length of N_sThe fourth-order lag moment vector has a regularly changing peak value, but because the peak value distribution of the one-dimensional vector is sparse, the training effect of directly inputting the fourth-order lag moment vector into the network is poor. Meanwhile, in the field of Computer Vision (CV), the image dimension of the input layer usually adopts a square structure with the same side length, and the side length usually takes the power of 2, so we will use N as the power_sN (in generating STBC-OFDM signal_sI.e. by default chosen to the power of 2) of length N_sThe fourth order lag moment vectors are used as a group and are merged into a fourth order lag moment spectrum by adopting two-dimensional vector splicing.

Further, in the present embodiment, the fourth order lag moment spectrum, is represented as:

wherein,

representing the kth fourth order lag moment vector.

The kth column vector

Can be expressed as:

the fusion of multi-scale information and the mining of relationship information between different pixel points in the image can effectively improve the utilization rate of detail information of input features and fully utilize the complementarity between different types of features. And the expanded convolution (scaled convolution) can ensure that the size of the output feature map is kept unchanged when a larger Receptive Field (received Field) is obtained, which means that the features of the image in a larger area can be obtained under the condition that the parameters to be trained are as few as possible. Therefore, the FOLMS detail information can be extracted by adopting multi-scale expansion convolution, and image information loss caused by extracting only a single feature is avoided.

Specifically, in selecting the expansion rate (ratio) of the multiscale expansion convolution portion, considering that the characteristics distribution of peak-valley-peak and peak-valley-peak (AL-OFDM signal and STBC3-OFDM signal) exists in the form of the input STBC-OFDM signal, the expansion convolution with expansion rates of 2 and 3 is adopted to better extract the effective characteristics of the form. Assuming that the expansion rate of the expansion convolution is d, the receptive field size of the nth layer convolution kernel at this expansion rate is:

wherein, RF_n-1Denotes the field size, k, of the n-1 th layer convolution kernel_nIs the n-th layerSize of convolution kernel, stride_iRepresenting the step size of the i-th layer convolution kernel. According to the formula, under the condition that the size and the step length of the convolution kernel are fixed, the expansion rate is increased, so that the receptive field scale can be effectively expanded, and the image characteristic information in a wider area range can be acquired. Furthermore, a standard convolution layer is added to further extract deep features of the multi-scale feature map, so that the features acquired by the network have stronger characterization capability.

Further, in the embodiment of the present application, inputting the fourth-order lag moment spectrum into the attention-guided multi-scale dilation convolution module, and outputting the multi-scale guiding features, includes the following steps:

Considering that the extracted fourth-order lag moment spectrum does not have peak features in all regions, for example, for STBC4-OFDM, the feature distribution of its fomms is valley-peak, with features concentrated in the peak region part. Therefore, in order to enable the network to put more attention resources into the target area needing important attention so as to acquire more detailed information of the target needing attention and suppress other useless information, the convolution block attention module is adopted to further improve the identification performance of the network. Compared to other attention mechanism modules, such as the pinch-actuated module, CBAM exhibits significantly superior performance in both class recognition (ImageNet dataset) and target detection problem (MS COCO dataset), and it can also be used in conjunction with any CNN structure with good adaptability and generalization. The CBAM includes two sub-modules, a channel attention module and a spatial attention module.

Further, in an embodiment of the present application, guiding the multi-scale deep features via a convolution block attention module to generate the multi-scale guiding features includes:

Further, in the embodiment of the present application, the inputting the multi-scale deep features into the channel attention module specifically includes:

For channel attention modules, input feature maps

First, two channel attention maps are generated through global average-pooling (global average-pooling) and global maximum-pooling (global max-pooling) of channel dimensions (channel axes)

And

then, note channel M_c1And M_c2Feeding in a shared multi-layer perceptron, while, in order to reduce the parameters, the size of the shared network is set to

(r is a reduction ratio). Finally, the feature output by the shared multi-layer perceptron (shared MLP) is subjected to element-level-based addition operation, and then is subjected to sigmoid activation operation to generate a final channel attention feature, namely M_c. In summary, the calculation formula can be expressed as：

Wherein,

and

representing the shared network weights in MLP,

and

mean pooling and maximum pooling characteristics are represented, σ (-) represents sigmoid activation function.

Further, in the embodiment of the present application, taking the product of the channel attention feature and the multi-scale deep feature as an input of the spatial attention module, generating a final multi-scale guiding feature includes:

For the space attention module, the input of the module is the output of the channel attention module

And CBAM input feature map

Expressed as:

performing global average pooling and global maximum pooling on the spatial dimension (spatial axis) on the feature map F' to obtain a feature map

And

and will be characterized by

And

a splicing operation is performed, represented as:

then, the spliced feature map F_concatGenerating a space attention feature after a standard convolution operation with 7 multiplied by 7 dimensions and a sigmoid activation function, wherein the space attention feature is expressed as:

finally, the space is noted to be characterized

Multiplying the input characteristic diagram F of the module to obtain the final output characteristic expressed as:

different fusion methods are adopted to generate certain influence on the fusion effect of the multi-scale guide features, the output features of the convolution block attention module are fused by adopting a splicing fusion method, although the space complexity of the method is higher than that of the method adopting addition (add) fusion, the overall performance of the splicing fusion method is still better under the condition of comprehensively considering the identification performance and the time complexity. After the splicing fusion characteristics are obtained, a residual block is further adopted to weaken the problem of performance saturation degradation caused by the increase of the number of network layers, and meanwhile, the utilization rate of the splicing fusion characteristics is improved. The output characteristics of the residual block are:

y_RB＝MaxPool(f(Concat(F″_d1+F″_d2+F″_d3)+F(Concat(F″_d1+F″_d2+F″_d3),W_RB)))

wherein, F ″)_d1、F″_d2And F ″)_d3For the multi-scale guiding characteristics of 3 branches output when the expansion rates of the network are 1, 2 and 3 respectively, Concat (·) represents splicing fusion, F (·) is a residual mapping function, and F (·) is a ReLU (rectified Linear Unit) activation function. The multi-scale guide features can be more fully utilized through feature fusion and a residual block, cross-layer flow of feature information can be achieved, and the structure of the network is more compact and easy to train through fusion of detail information of a shallow network and high-dimensional features mapped by a deep network.

Fig. 2 is a block diagram of an STBC-OFDM transmitter of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

As shown in FIG. 2, the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum considers the signal with n_tA transmitting antenna and n_rIn the STBC-OFDM wireless communication system with the receiving antennas, data symbols are subjected to STBC coding to generate a plurality of paths of data streams, and then OFDM modulation is generated through two steps of Inverse Fast Fourier Transform (IFFT) and cyclic prefix (cyclic prefix) addition.

Fig. 3 is a schematic structural diagram of the foms/AMDC-net method of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to the embodiment of the present application.

As shown in FIG. 3, the FOLMS/AMDC-net (attention-directed multi-scale extended convolution network) scheme of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum comprises two main parts, namely, a fourth-order delay moment spectrum for feature extraction and an attention-directed multi-scale extended convolution network for identifying STBC-OFDM type. In particular, FOLMS involves generating a fourth order lag moment vector and a two-dimensional vector concatenation. The AMDC-net comprises an attention-guided multi-scale expansion convolution module, a feature fusion module, a residual block and a full connection layer with Softmax as an activation function.

Fig. 4 is a schematic diagram of fourth-order delay moment spectra of 4 STBC-OFDM signals in the STBC-OFDM signal blind identification method based on deep learning and fourth-order delay moment spectra according to the embodiment of the present application.

As shown in fig. 4, the number of received OFDM blocks N_bTaking 8000 as an example, a three-dimensional fourth-order lag moment spectrum with a signal-to-noise ratio of 10dB is obtained, and it is obvious from the figure that different STBC-OFDM signals have large differences, and the peaks and valleys of their fomms are regularly arranged at equal intervals, for example, the distribution of the fourth-order lag moment spectrum of the AL-OFDM signal in any vector sequence is peak-valley-peak, and the distribution of the vector sequence of the STBC3-OFDM signal is peak-valley-peak), which are of great significance for the deep learning algorithm to identify the signal type.

Fig. 5 is a schematic diagram of an overall structure of AMDC-net of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

As shown in fig. 5, the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum introduces deep learning into the STBC-OFDM signal identification field, and provides an identification scheme of a multi-scale expansion convolutional network based on fourth-order lag moment spectrum and attention guidance. Firstly, fourth-order lag moment of a received signal is calculated, fourth-order lag moment vectors are generated, and the generated vectors are combined into a fourth-order lag moment spectrum to serve as network input by further adopting two-dimensional vector splicing. Then, on the basis of fully extracting detail information of different scales of the image by utilizing multi-scale expansion convolution, a convolution block attention module is further introduced to construct an attention-guided multi-scale expansion convolution module, so that network resources are more concentrated on a target area needing important attention. And finally, fusing the multi-scale guide features to enhance the complementarity of the network features, adding a residual error layer to further enhance the utilization rate and the characterization capability of the deep fusion features, and outputting the recognition result through a full connection layer with softmax as an activation function.

Fig. 6 is a schematic diagram of the multi-scale dilation convolution of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to the embodiment of the present application.

As shown in fig. 6, when the expansion rate (scaling rate) of the multi-scale expansion convolution part is selected, considering that the feature distribution (AL-OFDM signal and STBC3-OFDM signal) of the peak-valley-peak and peak-valley-peak exists in the foms of the input STBC-OFDM signal, the expansion convolution with the expansion rates of 2 and 3 is adopted to better extract the effective feature of the foms, and the obtained multi-scale expansion convolution structure is shown in fig. 6.

Fig. 7 is a schematic diagram of a convolution block attention module scheme of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

As shown in fig. 7, in the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum, the identification performance of the network is further improved by adopting a convolution block attention module. Compared to other attention mechanism modules, such as the pinch-actuated module, CBAM exhibits significantly superior performance in both class recognition (ImageNet dataset) and target detection problem (MS COCO dataset), and it can also be used in conjunction with any CNN structure with good adaptability and generalization. The CBAM includes two sub-modules, a channel attention module and a spatial attention module.

Fig. 8 is a schematic diagram of an attention-directed multi-scale dilation convolution module of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to an embodiment of the present application.

As shown in fig. 8, in order to focus more attention resources of extracted multi-scale features on a target region needing important attention and further improve the identification performance of a network, the blind identification method for the STBC-OFDM signal based on deep learning and the fourth-order lag moment spectrum is adopted, a convolution block attention module is adopted to guide the multi-scale features, an attention-guided multi-scale expansion convolution module is designed to further obtain the multi-scale guide features with stronger feature mapping capability, and the real characteristics of the fourth-order lag moment spectrum features of the input STBC-OFDM signal can be better reflected. The structure of the attention-directed multi-scale expansion convolution module is shown in fig. 8, and the input fourth-order lag moment spectrum features are subjected to multi-scale expansion convolution to obtain multi-scale features, so that feature representations with peak difference and pertinence on the fourth-order lag moment spectra of different STBC-OFDM signals are obtained. And then, further extracting deep features of the multi-scale feature map through a standard convolutional layer, so that the features acquired by the network have stronger characterization capability. And finally, guiding the multi-scale deep features through a convolution block attention module to generate multi-scale guide features.

Fig. 9 is a schematic position diagram of an attention-guided multi-scale dilation convolution module of an STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum in an attention-guided multi-scale dilation convolution network according to an embodiment of the present application.

As shown in fig. 9, the attention-directed multi-scale extended convolution network of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum includes an attention-directed multi-scale extended convolution module, a feature fusion layer, and a residual layer, inputs the fourth-order lag moment spectrum into the attention-directed multi-scale extended convolution module, and outputs a multi-scale guided feature, including the following steps: obtaining multi-scale characteristics after the input fourth-order lag moment spectrum is subjected to multi-scale expansion convolution, and obtaining characteristic representation with pertinence to the fourth-order lag moment spectrum of different STBC-OFDM signals; further extracting deep features of the multi-scale feature map through a standard convolutional layer; and finally, guiding the multi-scale deep features through a convolution block attention module to generate multi-scale guiding features, inputting the multi-scale guiding features into a feature fusion layer and a residual error layer, and outputting a recognition result through a full connection layer with softmax as an activation function.

Fig. 10 is a schematic structural diagram of a feature fusion and residual block part of the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum according to the embodiment of the present application.

As shown in fig. 10, in the STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum, the output features of the convolution block attention module are fused by using a splicing fusion method, and although the spatial complexity of the method is higher than that of the method using addition (add) fusion, the overall performance of the splicing fusion method is still better under the condition of comprehensively considering the identification performance and the time complexity. After the splicing fusion characteristics are obtained, a residual block is further adopted to weaken the problem of performance saturation degradation caused by the increase of the number of network layers, and meanwhile, the utilization rate of the splicing fusion characteristics can be improved.

Studies were conducted on class 4 STBC-OFDM signals, namely SM-OFDM, AL-OFDM, STBC3-OFDM, and STBC 4-OFDM. Using the average recognition probability of class 4 signals as a measure of recognition performance

Where C is of STBC-OFDM signal type, i.e., C ═ { SM-OFDM, AL-OFDM, STBC3-OFDM, STBC4-OFDM },

indicates that the transmission signal is C_iThe recognition result of time is also C_iProbability of signal, P (C)_i) To transmit a signal of C_iBecause the number of generated samples of the 4-type signal is the same, the transmission probabilities of the 4-type signal all satisfy P (C)_i)＝1/4。

According to the method, the matlab is adopted to generate the STBC-OFDM signal and perform the four-order lag moment spectrum characteristic extraction, the establishment and training of the AMDC-net network are completed by relying on Keras based on the tensierflow rear end, and the training process is accelerated by the GPU. Selecting OFDM when generating STBC-OFDM signalThe number of sub-carriers of the block is 128, the length of the cyclic prefix is 16, the modulation mode is Quadrature Phase Shift Keying (QPSK), frequency selective fading channel is adopted for simulation, and the number of receiving antennas is N _r1. Every N of received signals_b8000 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 to be N _s16. The range of the signal-to-noise ratio of the generated sample is-16 dB, the step length is 2dB, 100 fourth-order lag moment spectrum input samples are generated by the single-class STBC-OFDM signal under each signal-to-noise ratio, the total capacity of the sample is 6800, and the proportion of the training sample to the verification sample is 8: 2. In the network optimization process, the Adam optimizer is adopted to solve the optimal solution, and the cross entropy is selected as a loss function.

In order to verify the effectiveness of identifying the STBC-OFDM signal by adopting multi-scale features in the model, the AMDC-net model and an attention-guided single-branch convolutional network (ASC-net) model which are proposed by the application are compared. The parameter settings and basic structure of each layer of the ASC-net model for performance comparison are consistent with those of AMDC-net, and the difference between the ASC-net and the AMDC-net is that only the upper branch of FIG. 5 which only adopts standard convolution is reserved, so that the network has no feature fusion part.

FIG. 11 is a graph showing a comparison of the performance of AMDC-net and ASC-net according to the embodiment of the present application.

As shown in fig. 11, comparing the performances of the AMDC-net model and the ASC-net (attention-directed single-branch convolutional network) model, the AMDC-net model using the multi-scale fusion feature is significantly improved in performance at low signal-to-noise ratio compared to the ASC-net under the same condition of attention-direction, and the average recognition probability at-16 dB is increased by 7.9% compared to the ASC-net. The multi-scale fusion features extracted by adopting the multi-scale expansion convolution effectively enrich feature types, enhance the complementarity among different features, and enable the AMDC-net model to have stronger representation capability, thereby verifying the effectiveness of adopting the multi-scale features.

For further analysis of the performance difference of AMDC-net and ASC-net under various code types, confusion matrixes of 4 types of STBC-OFDM signals under-16 dB, -14dB and-12 dB of two networks are provided,

FIG. 12 is a schematic diagram of AMDC-net and ASC-net confusion matrices according to an embodiment of the present application.

As shown in fig. 12, the main reason for poor deep learning model identification performance at low signal-to-noise ratio is confusion between the SM-OFDM signal and the AL-OFDM signal, which is the case for AMDC-net and ASC-net, and the reason for this phenomenon may be that the peak value of the foms of the AL-OFDM signal is low, so that the peak characteristics thereof are buried in interference and noise under the harsh environment condition of low signal-to-noise ratio. From the observation and comparison in the longitudinal direction, it can be seen that under the same signal-to-noise ratio condition, the coincidence degree of the True label (True label) and the Predicted label (Predicted label) of the AMDC-net is higher, and the misjudgment rate is obviously lower than that of the ASC-net. The performance of the two networks is steadily increased along with the increase of the signal-to-noise ratio by observing and comparing from a transverse angle, particularly, the average recognition probability of the two networks to 4 types of STBC-OFDM signals under-12 dB is more than 94.5%, the diagonal distribution of a confusion matrix is obvious, the AMDC-net and the ASC-net have stronger adaptability to low signal-to-noise ratio, and meanwhile, the rationality of the idea that a deep learning model is introduced into STBC-OFDM signal recognition by constructing a fourth-order hysteresis moment spectrum is verified.

Verifying the effectiveness of adopting an attention mechanism by comparing 4 networks comprising the AMDC-net model components, wherein MDC-net + CBAM represents an original AMDC-net model; the MDC-net represents an original attention-guided multi-scale expansion convolution network without a convolution block attention module part, namely the multi-scale expansion convolution network; SC-net + CBAM represents a single-branch convolutional network with a convolutional block attention module, and the structure of the network is the same as that of ASC-net; SC-net represents a single-branch convolutional network or CBAM without multi-scale features. The four networks were subjected to simulation experiments in the same dataset, sample scale, batch size and hardware environment, which enabled a more accurate assessment of the effectiveness of using the attention mechanism in the model.

Fig. 13 is a schematic diagram illustrating performance comparison of four networks for attention mechanism evaluation according to an embodiment of the present application.

As shown in fig. 13, by virtue of CBAM, the attention resources of features are more focused on the target area needing important attention, and the performance of SC-net or MDC-net after the attention mechanism is added is improved to some extent, especially the average recognition probability of MDC-net after the CBAM is added is improved by 5.2%, which indicates that the characterization capability of the network on the FOLMS feature can be improved by properly guiding the extracted preliminary features by using the CBAM model, so that the effectiveness of the attention mechanism used in AMDC-net is verified. Furthermore, by observing the enlarged image of the average recognition probability of the 4 types of networks under-14 dB, the performance improvement of the MDC-net model fusing the multi-scale features after the CBAM is added is more remarkable compared with the SC-net model only adopting a single branch path to extract the features, namely, the introduction of the fusion features further stimulates the efficiency of the attention mechanism. The reasonability of the application in combination of the attention mechanism and the multi-scale features is laterally explained, the CBAM guides the attention of the features under different scales to further improve the identifiability of the fused features, and the CBAM and the fused features mutually promote and supplement each other.

Fig. 14 is a schematic diagram of confusion matrices at-14 db for four networks according to an embodiment of the present application.

As shown in FIG. 14, in order to further analyze the difference of the identification accuracy of the 4 networks on different STBC-OFDM signals, a confusion matrix of the 4 networks under-14 dB is given. By comparing the MDC-net + CBAM with the MDC-net, the identification accuracy of the AL-OFDM signal by the MDC-net model after the CBAM is added is obviously improved, and the integral identification performance is obviously improved, which shows that the problem of annihilation of AL-OFDM signal peak values caused by noise interference can be effectively relieved by adopting an attention mechanism to conduct feature guidance. For an SC-net + CBAM model and an SC-net model which do not adopt multi-scale features, the recognition rate of the model to SM-OFDM, STBC3-OFDM and STBC4-OFDM is more than 92.5%, but the misjudgment rate of AL-OFDM signals is high, and nearly 60% of AL-OFDM signals are wrongly predicted to be SM-OFDM. Therefore, the performance of the single-branch convolutional network for identifying the AL-OFDM signal is further improved, and the overall performance of the STBC-OFDM signal identification method based on deep learning under low signal-to-noise ratio is effectively improved.

The method mainly considers two modes of splicing and adding which are widely applied in a deep learning framework when a multi-scale guide feature fusion (fusion) method is selected, and the two fusion methods are compared from two angles of identification performance and calculation complexity for analyzing the overall comprehensive performance of different methods.

Fig. 15 is a schematic diagram of the recognition performance of two fusion methods according to the embodiment of the present application.

As shown in FIG. 15, overall, the performance of the splice fusion at low SNR is slightly better than that of the splice fusion at-16 dB and-14 dB, the average recognition probability is improved by about 2.6%, and the performance gain is gradually weakened as the SNR is increased. In fact, this recognition advantage of stitching fusion comes from its retention of multi-scale guided feature components. Compared with additive fusion, the splicing fusion is completed by stacking the feature maps instead of overlapping, so that the channel dimension of each multi-scale guide feature is reserved instead of concentrating all guide features in a certain feature map due to the additive fusion. Therefore, the adoption of splicing fusion can enable the FOLMS deep detail information extracted by each attention-guiding branch to be completely utilized by a subsequent residual block, thereby further improving the identification performance of the AMDC-net.

In order to more comprehensively and deeply compare the two fusion methods, the calculation complexity of splicing fusion and addition fusion is analyzed, and the trainable parameters P of the network are mainly used_trainAverage training time per generation T_trainAnd average test time T of individual samples_testAnd considering the three aspects. It is worth noting that in order to more fairly compare the performance of the join and add, the simulation experiments were performed with the same sample dimensions, data set size, training/test ratio, and training batch size. As shown in table one, although the trainable parameters of the stitching fusion are twice of those of the additive fusion, the training and testing time difference of the two fusion methods is very small due to the improvement of the parallel operation capability of the GPU, the identification time of a single sample is less than 0.2ms, and the method has strong real-time performance. Therefore, the comprehensive performance of the recognition performance and the calculation complexity are comprehensively compared, and the comprehensive performance of the recognition accuracy and the recognition efficiency is better by adopting splicing and fusing in the AMDC-net model.

Watch 1

Since the existing STBC-OFDM identification algorithms are few and mostly conventional algorithms based on feature extraction, to further illustrate the advantages of the deep learning based FOLMS/AMDC-net identification scheme for identifying STBC-OFDM signals, the AMDC-net is compared with CCF (Cross-Correlation Function), SOCS (Second-Order cyclic Statistics), and FOLP (Fourth Order Lag Product), and specifically, the CCF identifies the STBC-OFDM signal type by using the Cross-Correlation features between the received signals from different antennas. SOCS relies on second order signal cyclostationarity to identify the signal type by comparing the second order cyclic statistic of the received signal to a threshold. The FOLP identifies the type of the transmitted STBC-OFDM signal according to the fourth-order lag moments of the received signal according to the correlation difference of different space-time block code coding matrixes. The algorithms are typical traditional algorithms which are based on hypothesis test and statistical characteristic quantity and utilize decision trees and artificial threshold values to judge the signal types, and have certain representativeness in the field of STBC-OFDM signal identification.

Fig. 16 is a diagram illustrating comparison of recognition performances of different methods according to an embodiment of the present application.

As shown in fig. 16, compared with the recognition performances of 4 types of methods such as AMDC-net, CCF, SOCS, and FOLP, the recognition performance of AMDC-net under low signal-to-noise ratio is obviously superior to that of other traditional algorithms, and the extracted deep features have stronger recognition capability. The effective characteristics more suitable for distinguishing the STBC-OFDM signals can be automatically learned by the deep learning model, the identification performance of the traditional algorithm depends on the threshold value set by people to a great extent, the adaptability of the parameters set by people to the complex environment is poor in the process of hypothesis testing, and the extracted accumulated characteristic quantity may have large fluctuation. According to the method, the STBC-OFDM signal identification is introduced into a deep learning method, so that the problem that the manual extraction features in the traditional algorithm are greatly influenced by noise interference is solved, and the identification performance of the signal under the low signal-to-noise ratio is further remarkably improved.

Simulation experiments show that the identification performance of the FOLMS/AMDC-net (network) is remarkably improved compared with that of the existing algorithm, and the performance gain is obvious under the condition of low signal-to-noise ratio. In addition, in order to verify the effectiveness of using the multiscale dilation convolution and volume block attention modules, the proposed AMDC-net was also compared to MDC-net, single-branch convolution network (SC-net) + CBAM and SC-net. Although the identification performance of SC-net is the worst of the four types, the identification rates of the SC-net to SM-OFDM, STBC3-OFDM and STBC4-OFDM still reach more than 92.5 percent, so that the anti-interference capability of the SC-net can be obviously improved by improving the performance of the SC-net to AL-OFDM.

As shown in fig. 17, the STBC-OFDM signal blind identification apparatus based on deep learning and fourth-order lag moment spectrum includes: a vector generation module, a splicing module and a result generation module, wherein,

a vector generation module 10, configured to calculate a fourth-order lag moment of the received signal and generate a fourth-order lag moment vector;

a stitching module 20, configured to merge the fourth-order lag moment vectors into a fourth-order lag moment spectrum by using two-dimensional vector stitching;

the result generating module 30 is configured to construct an attention-directed multi-scale expansion convolutional network, take the fourth-order lag moment spectrum as an input, and output a recognition result, where the attention-directed multi-scale expansion convolutional network includes an attention-directed multi-scale expansion convolutional module, a feature fusion layer, and a residual layer, and takes the fourth-order lag moment spectrum as an input, and outputs a recognition result, and the recognition result includes: inputting the fourth-order lag moment spectrum into an attention-guided multi-scale expansion convolution module, outputting multi-scale guide features, inputting the multi-scale guide features into a feature fusion layer and a residual error layer, and outputting a recognition result through a full connection layer with softmax as an activation function.

The STBC-OFDM signal blind identification device based on deep learning and fourth-order lag moment spectrum in the embodiment of the application comprises: the device comprises a vector generation module, a splicing module and a result generation module, wherein the vector generation module is used for calculating a fourth-order lag moment of a received signal and generating a fourth-order lag moment vector; the splicing module is used for merging the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing; the result generation module is used for constructing the attention-guided multi-scale expansion convolution network, taking the fourth-order lag moment spectrum as input and outputting the identification result, wherein the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual error layer, and the identification result is output by taking the fourth-order lag moment spectrum as input and comprises the following steps: inputting the fourth-order lag moment spectrum into an attention-guided multi-scale expansion convolution module, outputting multi-scale guide features, inputting the multi-scale guide features into a feature fusion layer and a residual error layer, and outputting a recognition result through a full connection layer with softmax as an activation function. Therefore, the technical problems that the recognition performance of the existing method needs to be greatly influenced by parameter selection and is poor in robustness to different communication environments due to the fact that the existing method needs to artificially extract features and set a detection threshold value can be solved, meanwhile, the technical problems that signals are greatly influenced by noise and channel attenuation when being transmitted in a complex electromagnetic environment with a low signal-to-noise ratio and cannot be accurately and quickly recognized in the existing method can be solved, the method has excellent comprehensive recognition performance, is short in recognition time of a single sample, remarkably improves recognition accuracy under the low signal-to-noise ratio, and has good adaptability to a strong interference environment. In addition, the FOLMS sample obtained by preprocessing can be directly identified by the deep learning model, so that the method does not need prior information such as channels, noise and the like, and is more suitable for non-cooperative communication compared with the conventional algorithm.

In order to implement the above embodiments, the present application also proposes a non-transitory computer-readable storage medium on which a computer program is stored, which, when executed by a processor, implements the STBC-OFDM signal blind recognition method based on deep learning and fourth-order lag moment spectrum of the above embodiments.

In the description herein, reference to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present application, "plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.

Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing steps of a custom logic function or process, and alternate implementations are included within the scope of the preferred embodiment of the present application in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present application.

The logic and/or steps represented in the flowcharts or otherwise described herein, e.g., an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.

It should be understood that portions of the present application may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.

It will be understood by those skilled in the art that all or part of the steps carried by the method for implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and when the program is executed, the program includes one or a combination of the steps of the method embodiments.

In addition, functional units in the embodiments of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a stand-alone product, may also be stored in a computer readable storage medium.

The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc. Although embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application, and that variations, modifications, substitutions and alterations may be made to the above embodiments by those of ordinary skill in the art within the scope of the present application.

Claims

1. A STBC-OFDM signal blind identification method based on deep learning and fourth-order lag moment spectrum is characterized by comprising the following steps:

calculating a fourth-order lag moment of the received signal and generating a fourth-order lag moment vector;

combining the fourth-order lag moment vectors into a fourth-order lag moment spectrum by adopting two-dimensional vector splicing;

constructing an attention-guided multi-scale expansion convolution network, taking the fourth-order lag moment spectrum as an input, and outputting a recognition result, wherein,

the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual error layer, wherein the fourth-order lag moment spectrum is used as an input to output an identification result, and the identification result comprises the following steps:

inputting the fourth-order lag moment spectrum into an attention-guided multi-scale expansion convolution module, outputting multi-scale guide features, inputting the multi-scale guide features into a feature fusion layer and a residual layer, and outputting a recognition result through a full connection layer with softmax as an activation function.

2. The method of claim 1, wherein fourth order lag moments of the received signal are calculated and a fourth order lag moment vector is generated, wherein the fourth order lag moments are represented as:

the fourth order lag moment vector is represented as:

V＝[E[y(0,τ)],E[y(1,τ)],…,E[y(N_s-1,τ)]]

3. The method of claim 1, wherein the fourth order lag moment spectrum is represented as:

wherein,

representing the kth fourth order lag moment vector.

4. The method of claim 1, wherein inputting the fourth order lag moment spectrum into an attention-directed multi-scale dilation convolution module and outputting multi-scale directed features comprises the steps of:

5. The method of claim 4, wherein the guiding the multi-scale deep features via a convolution block attention module to generate multi-scale guiding features comprises:

taking the product of the channel attention feature and the multi-scale deep feature as an input of a spatial attention module, and outputting a final multi-scale guide feature.

6. The method of claim 5, wherein inputting the multi-scale deep features into the channel attention module comprises:

7. The method of claim 5, wherein taking the product of the channel attention feature and the multi-scale deep feature as an input to a spatial attention module, generating a final multi-scale guide feature, comprises:

8. The STBC-OFDM signal blind identification device based on deep learning and fourth-order lag moment spectrum is characterized by comprising a vector generation module, a splicing module and a result generation module, wherein,

the result generation module is used for constructing an attention-guided multi-scale expansion convolution network, taking the fourth-order lag moment spectrum as input and outputting an identification result, wherein,

the attention-guided multi-scale expansion convolution network comprises an attention-guided multi-scale expansion convolution module, a feature fusion layer and a residual error layer, wherein the fourth-order lag moment spectrum is used as an input to output an identification result, and the identification result specifically comprises the following steps:

9. A non-transitory computer-readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the method of any one of claims 1-7.