CN114139585A - Method, model and system for multi-component interactive feature signal processing of complex signal - Google Patents

Abstract

The invention discloses a complex signal multi-component interactive feature signal processing method, model and system, which can process the complex signal to obtain useful information of the signal and enhance the performance of implicit features. The frequency-domain in-phase component and the frequency-domain quadrature component of the input complex signal are interactively processed, so that the features of the real part (in-phase component) of the final output complex signal are combined with the frequency-domain in-phase component and the frequency-domain positive component of the input complex signal. At the same time, the characteristics of the imaginary part (quadrature component) of the output complex signal are also combined with the frequency domain in-phase component and frequency domain quadrature component of the input complex signal, fully realizing the sharing and interaction of multi-component information, and the final output The complex signal has more tacit knowledge and useful information between the components of the electromagnetic signal, which lays the foundation for mining the tacit knowledge and useful information hidden between the components of the electromagnetic signal. processing compatibility.

Description

Method, model and system for multi-component interactive characteristic signal processing of complex signal

technical field

The present invention relates to the field of communication technologies, and in particular, to a method, model and system for processing a complex signal multi-component interactive characteristic signal.

Background technique

As the wave of the world's new technological revolution hits the field of warfare, electronic countermeasures have become the forerunner of modern warfare, revolutionizing the weapon system, army structure, warfare methods, and means of command. The ownership of the right to control the sky and control the sky. As an important module in the field of electronic warfare, electromagnetic environment simulation technology is the basis for weapon testing, military training and other electromagnetic environmental effects research in complex electromagnetic environments. the part that can not be lost. By simulating the realistic electromagnetic environment of the battlefield, this technology can not only detect the adaptability of radar, communications and other equipment to the electromagnetic environment, but also detect the ability of troops to command, tactical maneuver, fire strike, and overall protection in a complex electromagnetic environment. With the rapid development of cognitive radio and self-organizing networking, the electromagnetic environment of modern battlefields is becoming more and more complex, mainly characterized by dense, interleaved, complex and changeable. Electromagnetic signals are complex in form, frequency is changeable, and different signals overlap each other in frequency and time domains; the number of radiation sources is large, the distribution density of electromagnetic signals is large, and the distribution range is wide; uncertain factors such as signal interception probability and measurement accuracy will cause the measurement data to be biased.

In order for the digital bandpass communication system to have anti-aliasing ability and reduce the sampling rate, it is usually necessary to down-convert the signal, and use orthogonal transform, Hilbert transform and other methods to construct a complex signal with zero intermediate frequency, which mainly includes the in-phase components I and Quadrature component Q.

However, in fact, a key problem of complex electromagnetic environment simulation technology is how to propose a multi-dimensional display that can adapt to complex electromagnetic signals. The traditional model-based method follows the statistical pattern recognition framework, which is mainly composed of multiple processing modules such as data preprocessing, feature extraction, feature selection, and classifier design, and has good performance and remarkable results. However, the above framework lacks the overall consideration of each processing module such as feature extraction, feature selection, and classifier design, resulting in the model not being close enough to the data distribution, thus restricting the improvement of the model's adaptive ability; in addition, most of the existing processing methods are The I/Q components are first fused, and then a feature modeling method is constructed to mine features, and the ability to process complex signals is limited. In recent years, machine learning methods such as deep learning technology can map data to identification by constructing multi-layer nonlinear transformations. The high-level semantic feature space with good performance can effectively represent the useful information of the data, and provide a new technical means for the multi-component identifiable feature analysis of electromagnetic signals. Most of these methods first extract electromagnetic signal features in advance, and then build deep learning models such as back-end convolutional neural networks and long-term and short-term memory networks to further mine semantic features.

However, these methods are still staged in nature, limited to the framework of traditional signal recognition, and no corresponding new theories and new ideas have been proposed. Even though complex-valued neural networks have been used in the field of signal processing, these complex signal processing methods have poor compatibility with real-valued data and visualization techniques such as constellation diagrams and eye diagrams, resulting in a lack of model interpretability.

Although domestic researches have been carried out on the modeling of relevant electromagnetic signal characteristics, the research on the new problem of multi-component identifiable characteristic analysis of electromagnetic signals is not systematic enough, which will definitely be an important research content of complex electromagnetic environment simulation technology in the future. Therefore, taking deep learning technology as the main research tool, combined with theories such as pattern recognition and signal recognition, systematically conducting in-depth research on multi-component identifiable feature analysis of electromagnetic signals not only has important academic value and broad application prospects, but also improves the electronic countermeasure system. The anti-missile, early warning and guidance capabilities have important practical significance. Therefore, it is necessary to carry out research on multi-component interactive feature modeling methods of complex signals to provide a basic model for the analysis of identifiable features of electromagnetic signals.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a complex signal multi-component interactive characteristic signal processing method, model and system, so as to solve the above-mentioned content existing in the prior art.

In a first aspect, an embodiment of the present invention provides a method for processing a multi-component interactive characteristic signal of a complex signal. The method includes:

Obtain the frequency domain in-phase component and the frequency domain quadrature component; wherein, the frequency domain in-phase component constitutes the real part of the complex signal, and the frequency domain quadrature component constitutes the imaginary part of the complex signal;

obtaining a first kernel function, a second kernel function and an interactive kernel function; the first kernel function is the real part of the interactive kernel function, and the second kernel function is the imaginary part of the interactive kernel function;

Input the product of the frequency domain in-phase component and the first kernel function into the activation function to obtain the first in-phase activation component; input the product of the frequency domain quadrature component and the second kernel function into the activation function to obtain the first quadrature activation component ;

Input the product of the frequency domain quadrature component and the first kernel function into the activation function to obtain the second quadrature activation component; input the product of the frequency domain in-phase component and the second kernel function into the activation function to obtain the second in-phase activation component;

Obtain a candidate in-phase component and a candidate quadrature component; the candidate in-phase component is equal to the second quadrature activation component minus the second in-phase activation component; the candidate quadrature component is equal to the first in-phase activation component minus the first quadrature activation component;

A candidate complex signal is obtained, the real part of the candidate complex signal is the candidate in-phase component, and the imaginary part of the candidate complex signal is the candidate quadrature component.

Optionally, after obtaining the candidate complex signal, the method further includes:

Through the attention mechanism, the candidate complex signal is adjusted based on the shared kernel function to obtain an adjusted candidate complex signal;

Residual connection is performed on the adjustment candidate complex signal and the complex signal to obtain an output complex signal.

Optionally, after performing residual connection on the adjustment candidate complex signal and the complex signal to obtain an output complex signal, the method further includes:

Obtain the real and imaginary parts of the output of the k-1th layer of the multi-component feature interaction unit; k is a positive integer greater than 1; when k=2, the output of the k-1th layer of the multi-component feature interaction unit is all said output complex signal;

update the first kernel function, the second kernel function and the interactive kernel function;

The product of the real part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated first in-phase activation component; the imaginary part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the first orthogonal activation component;

The product of the imaginary part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated second orthogonal activation component; the real part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the second in-phase activation component;

Obtain the candidate complex signal of the kth layer, the real part of the candidate complex signal of the kth layer is the candidate in-phase component of the kth layer, and the imaginary part of the candidate complex signal of the kth layer is the candidate quadrature of the kth layer component; the candidate in-phase component of the kth layer is equal to the update of the second quadrature activation component minus the update of the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first in-phase activation component AC active component;

Through the attention mechanism, the candidate complex signal of the kth layer is adjusted based on the new shared kernel function, and the candidate complex signal for adjusting the kth layer is obtained;

Residual connection is performed on the candidate complex signal of the adjusted k-th layer and the output of the k-1-th layer to obtain the output complex signal of the k-th layer.

Optionally, after obtaining the output complex signal, the method further includes:

performing a global pooling process on the obtained complex signal;

Discrete Fourier transform is performed on the output complex signal after global pooling to obtain classified data;

Classify the classified data to obtain classified information;

The classification information is normalized to obtain the signal feature extraction result.

Optionally, the obtaining the in-phase component in the frequency domain and the quadrature component in the frequency domain includes:

respectively encode the in-phase and quadrature components of the signal;

The coded in-phase component and the coded quadrature component are respectively mapped into the frequency domain to obtain the frequency-domain in-phase component and the frequency-domain quadrature component, respectively.

In a second aspect, an embodiment of the present invention further provides a complex signal multi-component interaction feature signal processing model, the model includes a multi-component feature interaction unit; the number of layers of the multi-component feature interaction unit is greater than or equal to 1 layer;

When the number of layers of the multi-component feature interaction unit is 1, the multi-component feature interaction unit is configured to execute any of the methods described above.

Optionally, the multi-component feature interaction unit is also used for:

Through the attention mechanism, the candidate complex signal is adjusted based on the shared kernel function to obtain an adjusted candidate complex signal;

Residual connection is performed on the adjustment candidate complex signal and the complex signal to obtain an output complex signal.

Optionally, when the number of layers of the multi-component feature interaction unit is greater than 1, for the kth layer of the multi-component feature interaction unit, where k is a positive integer greater than 1, the following method is performed:

Obtain the real and imaginary parts of the output of the k-1th layer of the multi-component feature interaction unit; k is a positive integer greater than 1; when k=2, the output of the k-1th layer of the multi-component feature interaction unit is all said output complex signal;

update the first kernel function, the second kernel function and the interactive kernel function;

The product of the real part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated first in-phase activation component; the imaginary part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the first orthogonal activation component;

The product of the imaginary part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated second orthogonal activation component; the real part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the second in-phase activation component;

Obtain the candidate complex signal of the kth layer, the real part of the candidate complex signal of the kth layer is the candidate in-phase component of the kth layer, and the imaginary part of the candidate complex signal of the kth layer is the candidate quadrature of the kth layer component; the candidate in-phase component of the kth layer is equal to the update of the second quadrature activation component minus the update of the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first in-phase activation component AC active component;

Through the attention mechanism, the candidate complex signal of the kth layer is adjusted based on the new shared kernel function, and the candidate complex signal for adjusting the kth layer is obtained;

Residual connection is performed on the candidate complex signal of the adjusted k-th layer and the output of the k-1-th layer to obtain the output complex signal of the k-th layer.

Optionally, the model further includes a post-processing unit, and the post-processing unit is used for:

performing a global pooling process on the obtained complex signal;

Discrete Fourier transform is performed on the output complex signal after global pooling to obtain classified data;

Classify the classified data to obtain classified information;

The classification information is normalized to obtain the signal feature extraction result.

In a third aspect, an embodiment of the present invention further provides a complex signal multi-component interactive feature signal processing system, the system includes a memory, a processor, and a computer program stored in the memory and running on the processor, the processing When the computer executes the program, the steps of any one of the above-mentioned methods are implemented.

Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

Embodiments of the present invention provide a method, model, and system for processing a multi-component interactive characteristic signal of a complex signal. The method includes: obtaining a frequency-domain in-phase component and a frequency-domain quadrature component; wherein the frequency-domain in-phase component constitutes a complex signal The real part of the frequency domain quadrature component constitutes the imaginary part of the complex signal; the first kernel function, the second kernel function and the interaction kernel function are obtained; the first kernel function is the real part of the interaction kernel function, and the second kernel function The function is the imaginary part of the interactive kernel function; the product of the in-phase component in the frequency domain and the first kernel function is input into the activation function to obtain the first in-phase activation component; the product of the quadrature component in the frequency domain and the second kernel function is input into the activation function , obtain the first quadrature activation component; input the product of the frequency domain quadrature component and the first kernel function into the activation function to obtain the second quadrature activation component; input the product of the frequency domain in-phase component and the second kernel function into the activation function In the function, the second in-phase activation component is obtained; the candidate in-phase component and the candidate quadrature component are obtained; the candidate in-phase component is equal to the second quadrature activation component minus the second in-phase activation component; the candidate quadrature component is equal to the first in-phase activation The first quadrature activation component is subtracted from the component; a candidate complex signal is obtained, the real part of the candidate complex signal is the candidate in-phase component, and the imaginary part of the candidate complex signal is the candidate quadrature component.

By adopting the above scheme, first, the above scheme is to process the complex signal, and extract the signal features on the basis of processing the complex signal. Information and latent features contain complementary information of different components, which enhances the identifiability of features. Second, traditional methods usually consider the real and imaginary parts to be independent, and use real-valued neural networks such as convolutional neural networks or long-short-term memory networks to process the real and imaginary parts of the signal respectively, but in fact, the real and imaginary parts of the signal are There will be interrelated tacit knowledge and useful information. Therefore, the present application performs interactive processing on the frequency domain in-phase component and the frequency domain quadrature component of the input complex signal, so that the real part (in-phase component of the final output complex signal) ) feature fuses the frequency-domain in-phase component and frequency-domain quadrature component of the input complex signal, and at the same time, the feature of the imaginary part (quadrature component) of the output complex signal also fuses the frequency-domain in-phase component and The frequency domain quadrature component realizes the full realization of the sharing and interaction of multi-component information. Compared with the simple addition, concatenation and other fusion methods in the prior art, the final output complex signal has more electromagnetic signal components. It lays a foundation for mining the tacit knowledge and useful information hidden between the various components of the electromagnetic signal. Third, because the above method processes complex signals, and real-valued signals can be simply converted into complex signals by using existing methods, the above methods are compatible with the processing of complex signals and real-valued signals. At the same time, the above methods The design process considers keeping the input and output dimensions consistent. Therefore, it can be embedded in any system or model that processes real-valued signals to enhance the system or model's ability to model and process signal features, that is, the above method can be a A system or model that can only process real-valued signals is upgraded to a system that can process both real-valued and complex signals, thereby enhancing the system or model's ability to process signals.

Description of drawings

FIG. 1 is a schematic diagram of a complex signal multi-component interactive characteristic signal processing model provided by an embodiment of the present invention.

FIG. 2 is a schematic structural diagram of a complex signal multi-component interactive characteristic signal processing system provided by an embodiment of the present invention.

Marked in the figure: 500-bus; 501-receiver; 502-processor; 503-transmitter; 504-memory; 505-bus interface.

Detailed ways

Example

In the digital frequency band communication system, in order to improve the efficiency of signal transmission on the channel and achieve the purpose of long-distance signal transmission, the digital baseband signal needs to undergo a digital modulation process to move the spectrum to high frequencies to form a frequency band suitable for transmission in the channel. At the same time, in order to increase the frequency band utilization rate and reduce the sampling rate in the subsequent processing, the receiving end needs to digitally demodulate the signal to obtain the I/Q (frequency domain in-phase component/frequency domain quadrature component) complex with zero intermediate frequency. Signal. However, this brings new challenges to the feature mining and recognition of electromagnetic signals. Traditional methods usually consider the real and imaginary parts to be independent, and use real-valued neural networks such as convolutional neural networks or long-short-term memory networks to process the real part of the signal separately. and the imaginary part, the complementarity of the quadrature and in-phase components is underutilized.

Therefore, it is very meaningful to propose a multi-component interactive feature signal processing method for complex signals to effectively process complex signals.

The present invention will be described in detail below with reference to the accompanying drawings.

Example 1

An embodiment of the present invention provides a method for processing a multi-component interactive characteristic signal of a complex signal, and the method includes:

S101: Obtain a frequency-domain in-phase component and a frequency-domain quadrature component.

The in-phase component in the frequency domain constitutes the real part of the complex signal, and the quadrature component in the frequency domain constitutes the imaginary part of the complex signal.

S102: Obtain a first kernel function, a second kernel function, and an interactive kernel function.

The first kernel function is the real part of the interaction kernel function, and the second kernel function is the imaginary part of the interaction kernel function. The initialization parameters of the first kernel function and the second kernel function are random parameters.

S103: Input the product of the frequency-domain in-phase component and the first kernel function into the activation function to obtain the first in-phase activation component; input the product of the frequency-domain quadrature component and the second kernel function into the activation function to obtain the first quadrature Active component. The product of the frequency domain quadrature component and the first kernel function is input into the activation function to obtain the second quadrature activation component; the product of the frequency domain in-phase component and the second kernel function is input into the activation function to obtain the second in-phase activation component.

S104: Obtain candidate in-phase components and candidate quadrature components.

Wherein, the candidate in-phase component is equal to the second quadrature activation component minus the second in-phase activation component; the candidate quadrature component is equal to the first in-phase activation component minus the first quadrature activation component;

S105: Obtain a candidate complex signal.

The real part of the candidate complex signal is a candidate in-phase component, and the imaginary part of the candidate complex signal is a candidate quadrature component.

By adopting the above scheme, first, the above scheme is to process the complex signal, and extract the signal features on the basis of processing the complex signal. Performance enhancements for information, latent features. Second, traditional methods usually consider the real and imaginary parts to be independent, and use real-valued neural networks such as convolutional neural networks or long-short-term memory networks to process the real and imaginary parts of the signal respectively, but in fact, the real and imaginary parts of the signal are There will be interrelated tacit knowledge and useful information. Therefore, the present application performs interactive processing on the frequency domain in-phase component and the frequency domain quadrature component of the input complex signal, so that the real part (in-phase component of the final output complex signal) ) feature fuses the frequency-domain in-phase component and frequency-domain quadrature component of the input complex signal, and at the same time, the feature of the imaginary part (quadrature component) of the output complex signal also fuses the frequency-domain in-phase component and The frequency domain quadrature component realizes the full realization of the sharing and interaction of multi-component information. Compared with the simple addition, concatenation and other fusion methods in the prior art, the final output complex signal has more electromagnetic signal components. It lays a foundation for mining the tacit knowledge and useful information hidden between the various components of the electromagnetic signal. Third, because the above method processes complex signals, and real-valued signals can be converted into complex signals, the above-mentioned methods are compatible with the processing of complex signals and real-valued signals, and the above-mentioned methods can be embedded in any one to process real-valued signals. In order to enhance the signal processing capability of the system or model, that is, a system or model that can only process real-valued signals can be expanded into a system that can process real-valued signals and complex signals, and the system or model can be enhanced. The ability to handle signals.

In order to ensure the interaction between the real part and the imaginary part, this application intends to construct a shared kernel function for the real and imaginary parts, and introduces an attention mechanism and a residual connection module to ensure the interaction between the responses of the multi-component interaction module. , after obtaining the candidate complex signal, the method further includes: adjusting the candidate complex signal based on the common kernel function through an attention mechanism to obtain an adjustment candidate complex signal, and then performing residual adjustment on the adjustment candidate complex signal and the complex signal Differential connection to obtain the output complex signal.

Among them, the shared kernel function is randomly initialized.

Furthermore, in order to extract global features, the embodiment of the present invention intends to introduce a global pooling unit, and process the multi-component interaction unit based on the channel-level inverse discrete Fourier transform to realize the final classification, specifically: after obtaining the output complex signal Then, perform global pooling processing on the obtained complex signal, and then perform discrete Fourier transform on the output complex signal after performing the global pooling processing to obtain classification data, and then perform classification processing on the classification data to obtain classification information , and finally normalize the classification information to obtain the signal feature extraction result.

Wherein, the obtaining the in-phase component in the frequency domain and the quadrature component in the frequency domain includes: respectively mapping the encoded in-phase component and the encoded quadrature component into the frequency domain, and obtaining the in-phase component in the frequency domain and the quadrature component in the frequency domain respectively . According to the time domain convolution theorem, the convolution of two signals in the time domain is equivalent to the product in the frequency domain. Therefore, the present application combines the channel characteristics and spatial characteristics of the signal, constructs discrete Fourier transform, realizes frequency domain transformation, maps the coded in-phase component and the coded quadrature component to the frequency domain, and obtains the frequency domain in-phase component respectively. The component and the frequency domain quadrature component are specifically: the encoded in-phase component and the encoded quadrature component are respectively mapped to the frequency domain through discrete Fourier transform, and the frequency domain in-phase component and the frequency domain quadrature component are obtained respectively, namely The coded in-phase component is mapped into the frequency domain through discrete Fourier transform to obtain the in-phase component in the frequency domain, and the coded quadrature component is mapped into the frequency domain through the discrete Fourier transform to obtain the frequency domain quadrature component.

Optionally, before the respectively mapping the encoded in-phase component and the encoded quadrature component into the frequency domain, the method further includes: encoding the in-phase component and the quadrature component of the signal respectively.

As an optional implementation manner, when the input is a real-valued signal, before encoding the in-phase component and the quadrature component of the signal respectively, the method further includes converting the real-valued signal into a complex signal, and then obtaining The in-phase and quadrature components of the complex signal are then encoded respectively. Among them, the real-valued signal can be converted into a complex signal by Fourier transform.

By adopting the above scheme, the core problem of the present invention is to solve the multi-component feature interaction of non-stationary electromagnetic signals, and realize intelligent electromagnetic spectrum control and utilization, realize the processing of electromagnetic complex signals, and on this basis, the multi-component of complex signals. The features are interactively processed (the specific interactive processing method is as described in S101-S105), and the tacit knowledge and useful information hidden between the components of the electromagnetic signal are mined, which is mainly different from the existing simple addition and concatenation. and other fusion methods to fully realize the sharing and interaction of multiple components. The most important parameters are the interaction kernel function and the shared kernel function, and the existing mature attention mechanism and residual connection module are introduced into the multi-component feature interaction modeling framework to assist the multi-component feature interaction unit to function.

Example 2

In the embodiment of the present invention, when multiple cycles are required, the multi-component feature interaction unit is required to have multiple layers, then after one cycle in the above-mentioned manner, that is, after the adjustment candidate complex signal and the complex signal are subjected to residual error After the output complex signal is obtained, the multi-component interaction characteristic signal processing method of the complex signal further includes:

Obtain the real and imaginary parts of the output of the k-1th layer of the multi-component feature interaction unit; k is a positive integer greater than 1; when k=2, the output of the k-1th layer of the multi-component feature interaction unit is all said output complex signal;

update the first kernel function, the second kernel function and the interactive kernel function;

The product of the real part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated first in-phase activation component; the imaginary part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the first orthogonal activation component;

The product of the imaginary part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated second orthogonal activation component; the real part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the second in-phase activation component;

Obtain the candidate complex signal of the kth layer, the real part of the candidate complex signal of the kth layer is the candidate in-phase component of the kth layer, and the imaginary part of the candidate complex signal of the kth layer is the candidate quadrature of the kth layer component; the candidate in-phase component of the kth layer is equal to the update of the second quadrature activation component minus the update of the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first in-phase activation component AC active component;

Through the attention mechanism, the candidate complex signal of the kth layer is adjusted based on the new shared kernel function, and the candidate complex signal for adjusting the kth layer is obtained;

Residual connection is performed on the candidate complex signal of the adjusted k-th layer and the output of the k-1-th layer to obtain the output complex signal of the k-th layer.

Then, after obtaining the output complex signal of the kth layer, the method further includes: performing global pooling processing on the output complex signal of the kth layer; performing discrete Fourier transformation on the output complex signal after the global pooling processing. Transform to obtain classified data; perform classification processing on the classified data to obtain classification information; perform normalization processing on the classified information to obtain a signal feature extraction result.

The method described in the embodiment of the present invention can be applied to an existing convolutional neural network for real-valued signals (existing real-valued convolutional neural network). Specifically, the method described in this application can be directly inserted into the i-th layer network in the existing real-valued convolutional neural network, and the i-th layer network can be constructed into a multi-component interaction layer (multi-component interaction module) to realize this A method for processing a complex signal multi-component interactive characteristic signal provided by an embodiment of the present invention. i=0, 1, 2, 3, ..., N, where N is any positive integer.

In order to explain the technical solution of the present invention more clearly, the technical solution of the present invention is further described by taking the example of inserting the technical solution of the present invention into the i-th layer network in the existing real-valued convolutional neural network. See Example 3 for details.

Example 3

S201: Obtain the frequency domain in-phase component Re ⁽ⁱ⁾ of the i-th layer network and the frequency domain quadrature component Im ⁽ⁱ⁾ of the i-th layer network.

The frequency domain in-phase component Re ⁽ⁱ⁺¹⁾ constitutes the real part of the i-th layer network complex signal, and the frequency domain quadrature component Im ⁽ⁱ⁾ constitutes the imaginary part of the i-th layer network complex signal. i=0, 1, 2, 3, ..., N, where N is any positive integer.

If the input is a real-valued signal, it is necessary to convert the real-valued signal into a complex signal first, then extract the in-phase component I and quadrature component Q of the complex signal, and perform nonlinear transformation on the in-phase component I and quadrature component Q to achieve The in-phase component I and the quadrature component Q of the signal, respectively, are encoded.

As an embodiment, the nonlinear transformation is performed on the in-phase component I and the quadrature component Q as shown in the following formula (1):

T=P(x) (1)

x由

和

两个部分组成x^I表示数据x的实部，x^Q表示数据x的虚部，

N为样本数目，L为数据采样长度，P(x)表示通过非线性变换对x进行非线性变换。Among them, x is any given data, which can be a real-valued signal, a complex-valued signal, or any other given data,

x by

and

The two parts are composed of x ^I representing the real part of the data x, x ^Q representing the imaginary part of the data x,

N is the number of samples, L is the data sampling length, and P(x) represents the nonlinear transformation of x through nonlinear transformation.

Optionally, the nonlinear transformation is performed on the in-phase component I and the quadrature component Q, that is, nonlinear transformation is performed on the in-phase component I and nonlinear transformation is performed on the quadrature component Q, that is, formula (1) can be split into T ^I =P (x ^I ), T ^Q =P(x ^Q ), ^TI represents the encoded in-phase component, and T ^Q represents the encoded quadrature component.

In the embodiment of the present invention, encoding the I/Q signal is to enrich the characteristic channels of the electromagnetic signal.

It should be noted that, in this embodiment of the present invention, after encoding the in-phase component I and the quadrature component Q of the signal, respectively, the encoded in-phase component and the encoded quadrature component are mapped into the frequency domain, respectively. Obtain the frequency domain in-phase component Re ⁽ⁱ⁾ of the i-th layer and the frequency-domain quadrature component Im ⁽ⁱ⁾ of the i-th layer, that is, obtain the frequency-domain in-phase component Re ⁽ⁱ⁾ of the i-th layer and the frequency-domain positive component of the i-th layer. The intersection component Im ⁽ⁱ⁾ includes: respectively mapping the encoded in-phase component and the encoded quadrature component to the frequency domain, to obtain the frequency domain in-phase component Re ⁽ⁱ⁾ of the i-th layer and the frequency domain of the i-th layer respectively Quadrature component Im ⁽ⁱ⁾ , where i is 0 or a positive integer.

Mapping the encoded in-phase component and the encoded quadrature component into the frequency domain can be implemented by discrete Fourier transform, as shown in the specific formula (2):

表示在频域中对T(C，L)进行离散傅里叶变换，ω表示是频域角频率。F(ω)表示在时域中对ω进行傅里叶变换。Among them, T(C, L) consists of T ^Q and T ^Q ,

represents the discrete Fourier transform of T(C, L) in the frequency domain, and ω represents the angular frequency in the frequency domain. F(ω) represents the Fourier transform of ω in the time domain.

According to the time domain convolution theorem, the convolution of two signals in the time domain is equivalent to the product in the frequency domain. Therefore, in the embodiment of the present application, the discrete Fourier transform is constructed by combining the channel characteristics and the spatial characteristics of the signal to realize the frequency domain transform, that is, the above formula (2).

S202: Obtain the first kernel function W ₁ ⁱ , the second kernel function W ₂ ⁱ of the i-th layer network, and the interaction kernel function Wi of the ⁱ -th layer network.

第i层网络的第一核函数W₁ ⁱ和第二核函数W₂ ⁱ是随机生成的。j是构成虚数的数学表示，称为j算子，通常j＝sqrt(-1)，j算子表示把一个复数逆时针旋转90度。Wherein, the first kernel function W ₁ ⁱ of the i-th layer network is the real part of the interactive kernel function Wi of the ⁱ -th layer network, and the second kernel function W ₂ ⁱ is the interactive kernel function of the i-th layer network. The ^imaginary part of Wi, as shown in the formula:

The first kernel function W ₁ ⁱ and the second kernel function W ₂ ⁱ of the i-th layer network are randomly generated. j is a mathematical representation of an imaginary number, called the j operator, usually j=sqrt(-1), and the j operator means to rotate a complex number 90 degrees counterclockwise.

S203: Input the product of the frequency domain in-phase component Re ⁽ⁱ⁾ of the i-th layer network and the first kernel function W ₁ ⁱ of the i-th layer network into the activation function to obtain the first in-phase activation component, and the i-th layer network The product of the frequency domain quadrature component Im ⁽ⁱ⁾ and the second kernel function W ₂ ⁱ of the i-th layer network is input into the activation function to obtain the first quadrature activation component.

S204: Input the product of the frequency domain quadrature component Im ⁽ⁱ⁾ equal to the i-th layer network and the first kernel function W ₁ ⁱ of the i-th layer network into the activation function to obtain a second quadrature activation component, and add the i-th layer The product between the frequency domain in-phase component Re ⁽ⁱ⁾ of the network and the second kernel function W ₂ ⁱ of the i-th layer network is input into the activation function to obtain the second in-phase activation component.

In this embodiment of the present invention, the activation function may adopt a ReLU function.

其中，获得第i层网络的候选复信号

由获得第i层网络的候选同相分量

和第i层的候选正交分量

S205: Obtain the candidate complex signal of the i-th layer network

Among them, the candidate complex signal of the i-th layer network is obtained

The candidate in-phase components of the i-th layer network are obtained by

and the candidate orthogonal components of the i-th layer

和第i层的候选正交分量

构成第i层网络的候选复信号

的实部和虚部，如公式(3)所示：Specifically, the candidate in-phase components of the i-th layer network are obtained by

and the candidate orthogonal components of the i-th layer

Candidate complex signals that make up the i-th layer network

The real and imaginary parts of , as shown in formula (3):

CR ⁽¹⁾ =W ⁱ *F(ω)=Re ⁽ⁱ⁺¹⁾ +jIm ⁽ⁱ⁺¹⁾ (3)

的实部为第i层网络的候选同相分量，所述第i层网络的候选复信号

的虚部为第i层网络的候选正交分量

That is, the candidate complex signal of the i-th layer network

The real part of is the candidate in-phase component of the i-th layer network, the candidate complex signal of the i-th layer network

The imaginary part of is the candidate orthogonal component of the i-th layer network

等于第二正交激活分量减去第二同相激活分量，具体为公式(4)所示：Among them, the candidate in-phase component of the i-th layer network

is equal to the second quadrature activation component minus the second in-phase activation component, as shown in formula (4):

等于第一同相激活分量减去第一正交激活分量，具体为如公式(5)所示：Candidate Orthogonal Components of the i-th Layer Network

Equal to the first in-phase activation component minus the first quadrature activation component, specifically as shown in formula (5):

Among them, ReLU is an activation function, which can be any activation function that satisfies the above input and output.

并引入注意力机制和残差连接模块，以保证多分量交互模块各响应之间的交互性，具体的，在获得第i层网络的候选复信号

之后，所述方法还包括：In order to ensure the interaction between the real part and the imaginary part, this application constructs a shared kernel function for the real part and the imaginary part

And introduce the attention mechanism and the residual connection module to ensure the interactivity between the responses of the multi-component interaction module. Specifically, after obtaining the candidate complex signal of the i-th layer network

Afterwards, the method further includes:

获得第i层网络的调整候选复信号；对所述第i层网络的调整候选复信号和第i层网络的复信号进行残差连接，获得第i层网络的输出复信号CR⁽¹⁾。具体如公式(6)所示：Through the attention mechanism, the candidate complex signal of the i-th layer network is adjusted based on the shared kernel function

Obtain the adjustment candidate complex signal of the i-th layer network; perform residual connection on the adjustment candidate complex signal of the i-th layer network and the complex signal of the i-th layer network to obtain the output complex signal CR ⁽¹⁾ of the i-th layer network. Specifically, as shown in formula (6):

where Re ⁽ⁱ⁺¹⁾ represents the real part of the output complex signal CR ⁽¹⁾ of the i-th layer network, and Im ⁽ⁱ⁺¹⁾ represents the imaginary part of the output complex signal CR ⁽¹⁾ of the i-th layer network.

In the embodiment of the present invention, if the depth of the multi-component interaction unit is k, that is, the multi-component interaction unit has k layers, and k is a positive integer greater than 1 or equal to 1.

When k=1, the output complex signal of the i-th layer network is CR ⁽¹⁾ , and CR ⁽¹⁾ also represents the output of the first layer of the multi-component interaction unit, that is, the output obtained by performing the above steps S201-S205 once .

When k is greater than 1, the output complex signal of the i-th layer network is CR ^(k) . In this case, the output complex signal CR ^(k) of the i-th layer network is obtained as follows:

Take k=2 as an example:

The output CR ⁽¹⁾ of the first layer of the multi-component interaction unit is used as the input of step S203, that is, the method as described in formula (2A) is performed:

F(ω)=CR ⁽¹⁾ =Re ⁽ⁱ⁺¹⁾ +Im ⁽ⁱ⁺¹⁾ (2A)

但是需要强调的是，第一核函数、第二核函数以及交互核函数的取值已更新。Then randomly generate new first kernel function, second kernel function and interactive kernel function, the representation is still

However, it should be emphasized that the values of the first kernel function, the second kernel function and the interactive kernel function have been updated.

Subsequently, referring to the solution described in steps S203-S205, process CR ⁽¹⁾ , and specifically refer to the following execution formula (3A), formula (4A) and formula (5A):

表示多分量交互单元有2层时第i层网络的候选复信号，

为多分量交互单元有2层时第i层网络的候选同相分量

表示多分量交互单元有2层时第i层网络的候选正交分量。in,

represents the candidate complex signal of the i-th layer network when the multi-component interaction unit has 2 layers,

is the candidate in-phase component of the i-th layer network when the multi-component interaction unit has 2 layers

Indicates the candidate orthogonal components of the i-th layer network when the multi-component interaction unit has 2 layers.

引入注意力机制和残差连接模块，以保证多分量交互模块各响应之间的交互性，即通过注意力机制，基于共用核函数调整所述第i层网络的候选复信号

获得第i层网络的调整候选复信号；对所述第i层网络的调整候选复信号和第i层网络的复信号进行残差连接，获得第i层网络的输出复信号CR⁽¹⁺¹⁾，具体参照公式(6A)所示的方式得到：The same shared kernel function that regenerates the real and imaginary parts

An attention mechanism and a residual connection module are introduced to ensure the interactivity between the responses of the multi-component interaction modules, that is, through the attention mechanism, the candidate complex signals of the i-th layer network are adjusted based on the shared kernel function.

Obtain the adjustment candidate complex signal of the i-th layer network; perform residual connection on the adjustment candidate complex signal of the i-th layer network and the complex signal of the i-th layer network, and obtain the output complex signal of the i-th layer network CR ^{(1+1 )} , specifically referring to the method shown in formula (6A) to obtain:

From the above results of k=1 and k=2, it can be deduced that when k is greater than 1, the expression of the output complex signal CR ^(k) of the i-th layer network is shown in formula (7):

CR ^(k) = CR ^(k-1) (...CR ⁽¹⁾ )

=Re ^(i+k) +jIm ^(i+k)

(7)

In this embodiment of the present invention, no matter the value of k is any positive integer, after obtaining CR ^(k) in the manner shown above, the method for the multi-component interactive characteristic signal of a complex signal further includes: for the obtained k-th The complex signal CR ^(k) of the layer is subjected to global pooling processing; the complex signal CR ^(k) of the kth layer after the global pooling processing is subjected to the inverse discrete Fourier transform F ^-1 (G(CR ^(k) ) ) to obtain classification data; classify the data to obtain classification information; perform normalization processing on the classification information to obtain the signal feature extraction result. Specifically, it can be obtained in the manner represented by formula (8):

Among them, P represents the feature extraction result of the classification information signal, G is the global pooling process, Classifier is the classifier, Softmax is the normalized exponential function, and F ^-1 (G(CR ^(k) )) represents the global redization process to obtain The data G(CR ^(k) ) is inverse discrete Fourier transform.

Example 4

Based on the above-mentioned embodiments, the present application uses formula symbols to describe the first step of Embodiment 2. When the multi-component feature interaction unit has multiple layers, the method includes:

S301: Obtain the real part and the imaginary part of the output of the k-1th layer of the multi-component feature interaction unit.

k is a positive integer greater than 1; when k=2, the output of the k-1th layer of the multi-component feature interaction unit is the output complex signal.

S302: Update the first kernel function, the second kernel function, and the interaction kernel function.

S303: Input the product of the real part of the output of the k-1th layer and the updated first kernel function into the activation function, and obtain the updated first in-phase activation component; After the product of the second kernel function is input into the activation function, the updated first quadrature activation component is obtained.

S304: Input the product of the imaginary part of the output of the k-1th layer and the updated first kernel function into the activation function, and obtain the updated second orthogonal activation component; add the real part of the output of the k-1th layer to the updated After the product of the second kernel function is input into the activation function, the updated second in-phase activation component is obtained.

S305: Obtain a candidate complex signal of the kth layer, where the real part of the candidate complex signal of the kth layer is the candidate in-phase component of the kth layer, and the imaginary part of the candidate complex signal of the kth layer is the candidate of the kth layer Orthogonal component; the candidate in-phase component of the kth layer is equal to the update of the second quadrature activation component minus the update of the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first in-phase activation component A quadrature activation component.

For the methods described in the above steps S301-S305, please understand in conjunction with formula (3B), formula (4B) and formula (5B):

Among them, M is the total number of layers of multi-component feature interaction units.

表示多分量交互单元在第k层时第i层网络的候选复信号，

为多分量交互单元在第k层时第i层网络的候选同相分量，

表示多分量交互单元在第k层时第i层网络的候选正交分量。in,

represents the candidate complex signal of the i-th layer network when the multi-component interaction unit is in the k-th layer,

is the candidate in-phase component of the i-th layer network when the multi-component interaction unit is in the k-th layer,

Represents the candidate orthogonal components of the i-th layer network when the multi-component interaction unit is at the k-th layer.

引入注意力机制和残差连接模块，以保证多分量交互模块各响应之间的交互性，即通过注意力机制，基于共用核函数调整所述第i层网络的候选复信号

(第k层的候选复信号)，获得第i层网络的调整候选复信号(调整第k层的候选复信号)；对所述第i层网络的调整候选复信号和上一层的输出进行残差连接，获得第i层网络的输出复信号CR^(k)(调整第k层的候选复信号)，即通过注意力机制，基于跟新的共用核函数调整所述第k层的候选复信号，获得调整第k层的候选复信号；对所述调整第k层的候选复信号和第k-1层的输出进行残差连接，获得第k层的输出复信号。具体参照公式(6B)所示的方式得到：The same shared kernel function that regenerates the real and imaginary parts

An attention mechanism and a residual connection module are introduced to ensure the interactivity between the responses of the multi-component interaction module, that is, through the attention mechanism, the candidate complex signal of the i-th layer network is adjusted based on the shared kernel function

(candidate complex signal of the k-th layer), obtain the adjustment candidate complex signal of the i-th layer network (adjust the candidate complex signal of the k-th layer); The residual connection is used to obtain the output complex signal CR ^(k) of the i-th layer network (adjust the candidate complex signal of the k-th layer), that is, through the attention mechanism, the candidate complex signal of the k-th layer is adjusted based on the new shared kernel function. signal to obtain a candidate complex signal for adjusting the kth layer; perform residual connection on the candidate complex signal for adjusting the kth layer and the output of the k-1th layer to obtain the output complex signal of the kth layer. Specifically referring to the method shown in formula (6B), it is obtained:

When k is greater than 1, the expression of the output complex signal CR ^(k) of the i-th layer network is shown in formula (7B):

Then, after obtaining the output complex signal of the kth layer, the method further includes: performing global pooling processing on the output complex signal of the kth layer; performing discrete Fourier transformation on the output complex signal after the global pooling processing. Transform to obtain classified data; classify the classified data to obtain classification information; normalize the classified information to obtain the signal feature extraction result.

The method described in the embodiment of the present invention can be applied to an existing convolutional neural network for real-valued signals (existing real-valued convolutional neural network). Specifically, the method described in this application can be directly inserted into the i-th layer network in the existing real-valued convolutional neural network, and the i-th layer network can be constructed into a multi-component interaction layer (multi-component interaction module) to realize this A method for processing a complex signal multi-component interactive characteristic signal provided by an embodiment of the present invention. i=0, 1, 2, 3, ..., N, where N is any positive integer.

Based on the above-mentioned method for processing complex signal multi-component interactive feature signals, an embodiment of the present invention proposes a complex signal multi-component interactive feature signal processing model. Please refer to FIG. 1 , the model includes a preprocessing unit, a multi-component feature interaction unit and a post-processing unit.

The preprocessing unit includes a nonlinear transform subunit and a discrete Fourier transform subunit, which are used to encode the in-phase component and the quadrature component of the signal respectively; the encoded in-phase component and the encoded quadrature The components are mapped into the frequency domain, and the frequency domain in-phase component and the frequency domain quadrature component are obtained respectively.

The multi-component feature interaction unit is used to realize the information handover processing of the real part and the imaginary part of the complex signal. The multi-component feature interaction unit includes one or more layers, and the multi-component feature interaction unit is specifically used for: obtaining the frequency domain in-phase component and the frequency domain quadrature component; wherein, the frequency domain in-phase component constitutes the real part of the complex signal, and the frequency domain quadrature component constitutes the imaginary part of the complex signal; obtain the first kernel function, the second kernel function and the interaction kernel function; The first kernel function is the real part of the interactive kernel function, and the second kernel function is the imaginary part of the interactive kernel function; the product of the frequency domain in-phase component and the first kernel function is input into the activation function, and the first in-phase activation is obtained. input the product of the frequency domain quadrature component and the second kernel function into the activation function to obtain the first quadrature activation component; input the product of the frequency domain quadrature component and the first kernel function into the activation function to obtain the second positive cross activation component; input the product of the frequency domain in-phase component and the second kernel function into the activation function to obtain the second in-phase activation component; obtain the candidate in-phase component and the candidate quadrature component; the candidate in-phase component is equal to the second quadrature activation component Subtract the second in-phase activation component; the candidate quadrature component is equal to the first in-phase activation component minus the first quadrature activation component; obtain a candidate complex signal, the real part of the candidate complex signal is the candidate in-phase component, the candidate complex signal The imaginary part of the signal is the candidate quadrature component. Through the attention mechanism, the candidate complex signal is adjusted based on the shared kernel function to obtain an adjustment candidate complex signal; the adjustment candidate complex signal and the complex signal are residually connected to obtain an output complex signal.

When the multi-component feature interaction unit includes multiple layers, the multi-component feature interaction unit is further used to: obtain the real part and the imaginary part of the output of the k-1th layer of the multi-component feature interaction unit; k is a positive integer greater than 1 When k=2, the output of the k-1th layer of the multi-component feature interaction unit is the described output complex signal; Update the first kernel function, the second kernel function and the interactive kernel function; The output of the k-1th layer The product of the real part and the updated first kernel function is input into the activation function, and the updated first in-phase activation component is obtained; the product of the imaginary part of the output of the k-1th layer and the updated second kernel function is input into the activation function In the function, the first orthogonal activation component is obtained and updated; the product of the imaginary part of the output of the k-1th layer and the updated first kernel function is input into the activation function, and the second orthogonal activation component is obtained and updated; The product of the real part of the output of the -1 layer and the updated second kernel function is input into the activation function, and the updated second in-phase activation component is obtained; the candidate complex signal of the kth layer is obtained, and the candidate complex signal of the kth layer is obtained. The real part is the candidate in-phase component of the k-th layer, and the imaginary part of the candidate complex signal of the k-th layer is the candidate quadrature component of the k-th layer; the candidate in-phase component of the k-th layer is equal to updating the second quadrature activation component Subtract and update the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first quadrature activation component; through the attention mechanism, based on the new shared kernel function to adjust the said The candidate complex signal of the kth layer is obtained to obtain a candidate complex signal for adjusting the kth layer; the residual connection is performed on the candidate complex signal of the adjustment kth layer and the output of the k-1th layer to obtain the output complex signal of the kth layer .

The post-processing unit is used to perform global pooling processing on the output complex signal; perform discrete Fourier transform on the output complex signal after the global pooling processing to obtain classification data; perform classification processing on the classification data to obtain classification information; normalize the classification information to obtain the signal feature extraction result.

The post-processing unit includes an inverse discrete Fourier transform 1 subunit, a discrete Fourier transform 2 subunit, and a classifier.

To sum up, based on the real-valued convolution module, this application proposes a complex-valued module for processing real and complex signals to be compatible with real-valued neural networks. On this basis, a multi-component interactive neural network is constructed to effectively process complex signals. Signal, the specific framework is shown in Figure 1.

As far as the invention itself is concerned, the multi-component identifiable feature analysis of electromagnetic signals in complex electromagnetic environments is one of the most cutting-edge research directions in the field of electronic countermeasures. Theories such as deep neural network and signal recognition show good feature modeling and representation capabilities in signal mining, and provide corresponding theoretical basis and technical support for the analysis of identifiable features of electromagnetic signals. The application of signals in military and civilian fields The needs also provide a realistic needs background for the research of this application. For example, radio monitoring and spectrum management, radio interference analysis, solve the illegal use of spectrum resources, forcible occupation, intentional interference, secret theft and other problems; in addition, real-time monitoring and analysis of new threat targets such as drones and illegal radio stations can be carried out. It can be seen that the key issues involved in the invention are current applied mathematics, especially the hot spot in the field of harmonic analysis, and also the frontier direction in the field of signal processing. The modules of the present invention are not completely independent, and their theoretical foundations and the problems to be solved all exist in certain overlapping places. The core problem is to solve the multi-component feature interaction of non-stationary electromagnetic signals, and to realize intelligent electromagnetic spectrum control and utilization. Therefore, this project is not only forward-looking and innovative in theory, but also has practical significance and broad application prospects in terms of actual needs.

The invention mainly processes the electromagnetic complex signal by constructing a multi-component feature interaction unit, and on this basis, constructs a complex signal multi-component interaction feature modeling framework, and mines the tacit knowledge and useful information hidden between the components of the electromagnetic signal , the main difference is that it is different from the existing simple addition, concatenation and other fusion methods, and fully realizes the sharing and interaction of multiple components. The most important parameters are the interaction kernel function and the shared kernel function. The existing mature attention mechanism and residual connection module are introduced into the multi-component feature interaction modeling framework to assist the multi-component feature interaction unit to function.

Regarding the models in the above-mentioned embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments of the method, and will not be described in detail here.

An embodiment of the present invention also provides a complex signal multi-component interactive feature signal processing system, as shown in FIG. 2 , comprising a memory 504, a processor 502, and a computer program stored in the memory 504 and running on the processor 502, When the processor 502 executes the program, the steps of any one of the foregoing methods for processing complex signals with multi-component interaction characteristics are implemented.

2, the bus architecture (represented by bus 500), bus 500 may include any number of interconnected buses and bridges, bus 500 will include one or more processors represented by processor 502 and memory 504. The various circuits of the memory are linked together. The bus 500 may also link together various other circuits, such as peripherals, voltage regulators and power management circuits, etc., which are well known in the art and therefore will not be described further herein. Bus interface 505 provides an interface between bus 500 and receiver 501 and transmitter 503 . Receiver 501 and transmitter 503 may be one and the same element, a transceiver, providing a means for communicating with various other devices over a transmission medium. The processor 502 is responsible for managing the bus 500 and general processing, while the memory 504 may be used to store data used by the processor 502 in performing operations.

Embodiments of the present invention also provide a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the steps of any of the foregoing methods for processing complex signals with multi-component interaction characteristics.

The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used with teaching based on this. The structure required to construct such a system is apparent from the above description. Furthermore, the present invention is not directed to any particular programming language. It is to be understood that various programming languages may be used to implement the inventions described herein, and that the descriptions of specific languages above are intended to disclose the best mode for carrying out the invention.

In the description provided herein, numerous specific details are set forth. It will be understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Similarly, it is to be understood that in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together into a single embodiment, figure, or its description. This disclosure, however, should not be construed as reflecting an intention that the invention as claimed requires more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the Detailed Description are hereby expressly incorporated into this Detailed Description, with each claim standing on its own as a separate embodiment of this invention.

Those skilled in the art will understand that the modules in the device in the embodiment can be adaptively changed and arranged in one or more devices different from the embodiment. The modules or units or components in the embodiments may be combined into one module or unit or component, and further they may be divided into multiple sub-modules or sub-units or sub-assemblies. All features disclosed in this specification (including accompanying claims, abstract and drawings) and any method so disclosed may be employed in any combination, unless at least some of such features and/or procedures or elements are mutually exclusive. All processes or units of equipment are combined. Each feature disclosed in this specification (including accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

Furthermore, it will be understood by those skilled in the art that although some of the embodiments herein include certain features, but not others, included in other embodiments, that combinations of features of the different embodiments are intended to be within the scope of the present invention And form different embodiments. For example, in the following claims, any of the claimed embodiments may be used in any combination.

Various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in the apparatus according to the embodiments of the present invention. The present invention can also be implemented as apparatus or apparatus programs (eg, computer programs and computer program products) for performing part or all of the methods described herein. Such a program implementing the present invention may be stored on a computer-readable medium, or may be in the form of one or more signals. Such signals may be downloaded from Internet sites, or provided on carrier signals, or in any other form.

Claims (10)

1. A complex signal multi-component interactive feature signal processing method, wherein the method comprises: Obtain the frequency domain in-phase component and the frequency domain quadrature component; wherein, the frequency domain in-phase component constitutes the real part of the complex signal, and the frequency domain quadrature component constitutes the imaginary part of the complex signal; obtaining a first kernel function, a second kernel function and an interactive kernel function; the first kernel function is the real part of the interactive kernel function, and the second kernel function is the imaginary part of the interactive kernel function; Input the product of the frequency domain in-phase component and the first kernel function into the activation function to obtain the first in-phase activation component; input the product of the frequency domain quadrature component and the second kernel function into the activation function to obtain the first quadrature activation component ; Input the product of the frequency domain quadrature component and the first kernel function into the activation function to obtain the second quadrature activation component; input the product of the frequency domain in-phase component and the second kernel function into the activation function to obtain the second in-phase activation component; Obtain a candidate in-phase component and a candidate quadrature component; the candidate in-phase component is equal to the second quadrature activation component minus the second in-phase activation component; the candidate quadrature component is equal to the first in-phase activation component minus the first quadrature activation component; A candidate complex signal is obtained, the real part of the candidate complex signal is the candidate in-phase component, and the imaginary part of the candidate complex signal is the candidate quadrature component. 2. The method for multi-component interactive feature signal processing of complex signals according to claim 1, wherein after obtaining the candidate complex signals, the method further comprises: Through the attention mechanism, the candidate complex signal is adjusted based on the shared kernel function to obtain an adjusted candidate complex signal; Residual connection is performed on the adjustment candidate complex signal and the complex signal to obtain an output complex signal. 3 . The method for multi-component interactive feature signal processing of complex signals according to claim 2 , wherein after residual connection is performed on the adjustment candidate complex signal and the complex signal to obtain an output complex signal, the method further comprises: 4 . : Obtain the real and imaginary parts of the output of the k-1th layer of the multi-component feature interaction unit; k is a positive integer greater than 1; when k=2, the output of the k-1th layer of the multi-component feature interaction unit is all said output complex signal; update the first kernel function, the second kernel function and the interactive kernel function; The product of the real part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated first in-phase activation component; the imaginary part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the first orthogonal activation component; The product of the imaginary part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated second orthogonal activation component; the real part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the second in-phase activation component; Obtain the candidate complex signal of the kth layer, the real part of the candidate complex signal of the kth layer is the candidate in-phase component of the kth layer, and the imaginary part of the candidate complex signal of the kth layer is the candidate quadrature of the kth layer component; the candidate in-phase component of the kth layer is equal to the update of the second quadrature activation component minus the update of the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first in-phase activation component AC active component; Through the attention mechanism, the candidate complex signal of the kth layer is adjusted based on the new shared kernel function, and the candidate complex signal for adjusting the kth layer is obtained; Residual connection is performed on the candidate complex signal of the adjusted k-th layer and the output of the k-1-th layer to obtain the output complex signal of the k-th layer. 4. The complex signal multi-component interactive feature signal processing method according to claim 2, wherein after obtaining the output complex signal, the method further comprises: performing a global pooling process on the obtained complex signal; Discrete Fourier transform is performed on the output complex signal after global pooling to obtain classified data; Classify the classified data to obtain classified information; The classification information is normalized to obtain the signal feature extraction result. 5. The multi-component interactive feature signal processing method for complex signals according to claim 1, wherein the obtaining the frequency-domain in-phase component and the frequency-domain quadrature component comprises: respectively encode the in-phase and quadrature components of the signal; The coded in-phase component and the coded quadrature component are respectively mapped into the frequency domain to obtain the frequency-domain in-phase component and the frequency-domain quadrature component, respectively. 6. A complex signal multi-component interaction feature signal processing model, characterized in that the model comprises a multi-component feature interaction unit; the number of layers of the multi-component feature interaction unit is greater than or equal to 1 layer; When the number of layers of the multi-component feature interaction unit is 1, the multi-component feature interaction unit is used to execute the method of claim 1 . 7. The complex signal multi-component interaction feature signal processing model according to claim 6, wherein the multi-component feature interaction unit is further used for: Through the attention mechanism, the candidate complex signal is adjusted based on the shared kernel function to obtain an adjusted candidate complex signal; Residual connection is performed on the adjustment candidate complex signal and the complex signal to obtain an output complex signal. 8. The complex signal multi-component interaction feature signal processing model according to claim 6, wherein when the number of layers of the multi-component feature interaction unit is greater than 1, for the kth layer of the multi-component feature interaction unit, k is a positive integer greater than 1, execute the following method: Obtain the real and imaginary parts of the output of the k-1th layer of the multi-component feature interaction unit; k is a positive integer greater than 1; when k=2, the output of the k-1th layer of the multi-component feature interaction unit is all said output complex signal; update the first kernel function, the second kernel function and the interactive kernel function; The product of the real part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated first in-phase activation component; the imaginary part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the first orthogonal activation component; The product of the imaginary part of the output of the k-1th layer and the updated first kernel function is input into the activation function to obtain the updated second orthogonal activation component; the real part of the output of the k-1th layer and the updated The product of the second kernel function is input into the activation function to obtain and update the second in-phase activation component; Obtain the candidate complex signal of the kth layer, the real part of the candidate complex signal of the kth layer is the candidate in-phase component of the kth layer, and the imaginary part of the candidate complex signal of the kth layer is the candidate quadrature of the kth layer component; the candidate in-phase component of the kth layer is equal to the update of the second quadrature activation component minus the update of the second in-phase activation component; the candidate quadrature component of the kth layer is equal to the update of the first in-phase activation component minus the update of the first in-phase activation component AC active component; Through the attention mechanism, the candidate complex signal of the kth layer is adjusted based on the new shared kernel function, and the candidate complex signal for adjusting the kth layer is obtained; Residual connection is performed on the candidate complex signal of the adjusted k-th layer and the output of the k-1-th layer to obtain the output complex signal of the k-th layer. 9. The complex signal multi-component interactive feature signal processing model according to claim 7, wherein the model further comprises a post-processing unit, and the post-processing unit is used for: performing a global pooling process on the obtained complex signal; Discrete Fourier transform is performed on the output complex signal after global pooling to obtain classified data; Classify the classified data to obtain classified information; The classification information is normalized to obtain the signal feature extraction result. 10. A complex signal multi-component interactive feature signal processing system, characterized in that the system comprises a memory, a processor and a computer program stored in the memory and executable on the processor, the processor executing the program while implementing the steps of the method of any one of claims 1-5.

Images