CN116016068B - Data-driven Inter-frequency Intelligent Intervention Signal Representation Method and System - Google Patents

Abstract

The invention discloses a data-driven mutual-frequency intelligent intervention signal representation method and system, which uses the output of one or more high-frequency units to perform intervention operations on the input of one or more low-frequency units, and through which one or more low-frequency The output of the unit intervenes on the input of one or more high-frequency units to realize the interactive processing of low-frequency components and high-frequency components by the high-low-frequency intervention model, that is, the intervention of high-frequency components on low-frequency components, and the use of low-frequency components on Intervention of high-frequency components, high-frequency components and low-frequency components intervene and drive each other, the obtained high-frequency output components integrate the hidden information of low-frequency components, and the low-frequency output components integrate the hidden information of high-frequency components, making high The frequency component and the low frequency component can represent the signal information more accurately, and the signal features obtained based on the high frequency output component and the low frequency output component can effectively represent the signal.

Description

Data-driven Inter-frequency Intelligent Intervention Signal Representation Method and System

technical field

The present invention relates to the field of communication technology, in particular, to a data-driven inter-frequency intelligent intervention signal representation method and system.

Background technique

In a non-cooperative communication system, only by obtaining the modulation mode of the signal can the signal be correctly demodulated. The key technologies involved in signal identification, target identification, etc. are necessary components of non-cooperative communication systems. They are an effective means of intercepting enemy signals and an important step in spectrum detection. They are used in drone swarm operations, electromagnetic spectrum perception, and radio communications. It has important theoretical significance and application value in military and civilian fields. With the rapid development of communication technology, electromagnetic signals present the characteristics of large quantity, high density, and complicated forms, which makes the communication environment more and more complex, and it is difficult to obtain effective signal representation, which limits the performance improvement of related intelligent systems, mainly reflected in the following three Aspects: First, the continuous increase in the coverage frequency of radiation source signals has led to more and more unknown signal types and quantities; second, with the improvement of communication technology, a large number of equipment with complex systems have begun to appear, and the generated electromagnetic signals are complex in form and frequency. The third is that the continuous widening of the operating frequency band of communication equipment and the increasingly complex working system will cause different radiation sources to overlap each other in the frequency band and time domain. Traditional signal representation methods usually follow the pattern recognition framework, mainly including multiple independent processing modules such as data preprocessing, feature extraction, feature selection, and classifier design. Specifically:

(1) Signal preprocessing. For example: data filtering and noise reduction, multipath signal discrimination, carrier frequency estimation, normalization and data alignment, etc.;

(2) Feature extraction and feature selection. For example: signal transient characteristics, steady state characteristics, transform domain characteristics, characteristic optimization, establishment and update of characteristic database, etc.;

⑶ classifier design method. For example: a variety of classifier designs suitable for engineering applications, etc.;

However, when the traditional method is used to extract the features of the signal first, and then use classic recognition algorithms such as SVM and ELM for recognition, the performance of the corresponding models is significantly different. The main reason is that the signal changes with time to hide the implicit information. Traditional methods cannot fully represent this implicit characteristic, resulting in insufficient signal feature mining, which makes traditional identification methods gradually lose their effectiveness.

In recent years, the computing power of computers has been greatly enhanced, and deep learning has also developed rapidly. Deep learning automatically learns and adjusts the weights and biases in the neural network according to the backpropagation algorithm, which is fundamentally different from the traditional model-based method in solving problems. to the abstract feature extraction process. In addition, deep learning technologies such as Transformer have strong generalization capabilities. The trained deep network model can adapt to complex communication environments and is a powerful tool for effectively representing signals.

However, the deep learning method itself does not have the ability to explicitly mine implicit knowledge from the signal, and the signal features obtained only by implicit mining are difficult to effectively represent the signal.

Contents of the invention

The object of the present invention is to provide a data-driven inter-frequency intelligent intervention signal representation method and system to solve the above-mentioned problems in the prior art.

In the first aspect, an embodiment of the present invention provides a data-driven inter-frequency intelligent intervention signal representation method, the method includes:

Obtain the time-frequency characteristics of the frequency conversion signal;

Extract low-frequency components and high-frequency components in time-frequency features;

Through the pre-trained high and low frequency intervention model, the low frequency component and the high frequency component are interactively processed to obtain the high frequency output component and the low frequency output component;

Among them, the high-low frequency intervention model includes high-frequency network and low-frequency network, the input of high-frequency network is high-frequency component, the input of low-frequency network is low-frequency component, the output of high-frequency network is high-frequency output component, and the output of low-frequency network is low-frequency output Component; the high-frequency network includes multiple high-frequency units, and the low-frequency network includes multiple low-frequency units. The output of one or more high-frequency units interferes with the input of one or more low-frequency units, and one or more low-frequency The output of the unit performs an intervention operation on the input of one or more high-frequency units, so as to realize the interactive processing of the low-frequency component and the high-frequency component by the high-low frequency intervention model.

Optionally, the high-frequency network includes 2 high-frequency units, and the low-frequency network includes 2 low-frequency units;

The two low-frequency units are respectively the first LSTM network and the third LSTM network; the two high-frequency units are respectively the second LSTM network and the fourth LSTM network;

The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component;

The input of the third LSTM network includes the first fusion component obtained after the output intervention operation of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network;

The input of the fourth LSTM network includes the second fusion component obtained after intervention operation on the output of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network.

Optionally, based on the output of the second LSTM network, the output intervention operation of the first LSTM network is used to obtain the first fusion component, including:

performing a connection operation on the output of the first LSTM network and the output of the second LSTM network to obtain a first connection component;

A convolution operation is performed on the first connected component to obtain a first fusion component.

Optionally, the input of the third LSTM network includes the first fusion component based on the output of the second LSTM network and the output of the fourth LSTM network after the intervention operation on the output of the first LSTM network, including:

performing a connection operation on the output of the first LSTM network and the output of the second LSTM network to obtain a first connection component;

performing a connection operation on the output of the first LSTM network and the output of the fourth LSTM network to obtain a second connection component;

performing a connection operation on the first connected component and the second connected component to obtain a third connected component;

A convolution operation is performed on the third connected component to obtain a first fusion component; the dimension of the first fusion component is the same as the dimension of the low frequency component.

Optionally, the high-frequency network includes 3 high-frequency units, and the low-frequency network includes 3 low-frequency units;

The three low-frequency units are respectively the first LSTM network, the third LSTM network and the fifth LSTM network; the three high-frequency units are respectively the second LSTM network, the fourth LSTM network and the sixth LSTM network;

The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component;

The input of the third LSTM network includes the first fusion component obtained after the output intervention operation of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network;

The input of the fourth LSTM network includes the second fusion component obtained after the output intervention operation of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network;

The input of the fifth LSTM network includes the third fusion component based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network after the intervention operation on the output of the third LSTM network;

The input of the sixth LSTM network includes the fourth fusion component obtained by intervening the output of the fourth LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network.

Optionally, after the low-frequency component and the high-frequency component are interactively processed through the pre-trained high-low-frequency intervention model to obtain the high-frequency output component and the low-frequency output component, the data-driven mutual-frequency intelligent intervention signal represents Methods also include:

Splicing the high-frequency output component and the low-frequency output component to obtain a spliced signal;

performing a convolution operation on the spliced signal to obtain a restored time-frequency domain signal;

Wherein, the dimension of the recovered time-frequency domain signal is the same as the dimension of the time-frequency feature.

In the second aspect, the embodiment of the present invention also provides a data-driven inter-frequency intelligent intervention signal representation system, the system includes:

Obtaining module, used for obtaining the time-frequency characteristic of the frequency conversion signal;

An extraction module is used to extract low-frequency components and high-frequency components in the time-frequency feature;

The interaction module is used to interactively process the low-frequency component and the high-frequency component through a pre-trained high-low frequency intervention model to obtain a high-frequency output component and a low-frequency output component;

Among them, the high-low frequency intervention model includes high-frequency network and low-frequency network, the input of high-frequency network is high-frequency component, the input of low-frequency network is low-frequency component, the output of high-frequency network is high-frequency output component, and the output of low-frequency network is low-frequency output Component; the high-frequency network includes multiple high-frequency units, and the low-frequency network includes multiple low-frequency units. The output of one or more high-frequency units interferes with the input of one or more low-frequency units, and one or more low-frequency The output of the unit performs an intervention operation on the input of one or more high-frequency units, so as to realize the interactive processing of the low-frequency component and the high-frequency component by the high-low frequency intervention model.

Optionally, the high-frequency network includes 2 high-frequency units, and the low-frequency network includes 2 low-frequency units;

The two low-frequency units are respectively the first LSTM network and the third LSTM network; the two high-frequency units are respectively the second LSTM network and the fourth LSTM network;

The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component;

The input of the third LSTM network includes the first fusion component obtained after the output intervention operation of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network;

The input of the fourth LSTM network includes the second fusion component obtained after intervention operation on the output of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network.

Optionally, the system also includes:

The splicing module is used to splice the high-frequency output component and the low-frequency output component to obtain a spliced signal;

A restoration module, configured to perform a convolution operation on the spliced signal to obtain a restored time-frequency domain signal;

Wherein, the dimension of the recovered time-frequency domain signal is the same as the dimension of the time-frequency feature.

In a third aspect, an embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and operable on the processor, and the processor implements any of the above-mentioned programs when executing the program. steps of the method described in the item.

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

The embodiment of the present invention provides a data-driven inter-frequency intelligent intervention signal representation method and system for extracting signal features, and extracting low-frequency components and high-frequency components in time-frequency features by obtaining time-frequency features of frequency-variable signals , through the pre-trained high and low frequency intervention model, the low frequency component and the high frequency component are interactively processed to obtain the high frequency output component and the low frequency output component. Among them, the high-low frequency intervention model includes high-frequency network and low-frequency network, the input of high-frequency network is high-frequency component, the input of low-frequency network is low-frequency component, the output of high-frequency network is high-frequency output component, and the output of low-frequency network is low-frequency output Component; the high-frequency network includes multiple high-frequency units, and the low-frequency network includes multiple low-frequency units. The output of one or more high-frequency units interferes with the input of one or more low-frequency units, and one or more low-frequency The output of the unit performs an intervention operation on the input of one or more high-frequency units, so as to realize the interactive processing of the low-frequency component and the high-frequency component by the high-low frequency intervention model.

By adopting the above scheme, the output of one or more high-frequency units interferes with the input of one or more low-frequency units, and the output of one or more low-frequency units interferes with the input of one or more high-frequency units. Operation, in order to realize the interactive processing of low-frequency components and high-frequency components by the high-low-frequency intervention model, that is, the intervention of high-frequency components on low-frequency components, and the intervention of low-frequency components on high-frequency components, the interaction between high-frequency components and low-frequency components Intervention and mutual driving, the obtained high-frequency output component combines the hidden information of the low-frequency component, and the low-frequency output component combines the hidden information of the high-frequency component, so that the high-frequency component and low-frequency component can represent the signal information more accurately. Based on The signal features obtained by the high-frequency output component and the low-frequency output component can effectively characterize the signal.

Description of drawings

Fig. 1 is a flowchart of a data-driven inter-frequency intelligent intervention signal representation method provided by an embodiment of the present invention.

Fig. 2 is a schematic diagram of a frequency separation process provided by an embodiment of the present invention.

FIG. 3 is a structural diagram of a basic unit of an LSTM network provided by an embodiment of the present invention.

Fig. 4 is a schematic diagram of a high-frequency network interfering with a low-frequency network according to an embodiment of the present invention.

Fig. 5 is a schematic diagram of a low-frequency network intervening in a high-frequency network according to an embodiment of the present invention.

Fig. 6 is a schematic diagram of another high-frequency network intervening in a low-frequency network according to an embodiment of the present invention.

Fig. 7 is a schematic diagram of another low-frequency network intervening in a high-frequency network according to an embodiment of the present invention.

Fig. 8 is a schematic block diagram of an electronic device provided by an embodiment of the present invention.

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

Detailed ways

Example

In the prior art, the method of signal extraction is mainly to extract the signal features of the hermit based on the convolutional neural network. However, it is difficult to effectively represent the signal by only using the convolutional neural network or the long-short-term memory network to implicitly mine the signal features. The present invention considers that the evolution process of the signal can be expressed by the effective dimension of frequency, by explicitly extracting the low-frequency component and high-frequency component of the signal from the frequency domain, and constructing different branches to represent the low-frequency component and high-frequency component, and Realize high-low frequency explicit interaction in the process of model construction, and use high-frequency or low-frequency knowledge to drive the training process of deep learning model, so that the performance of the original data-driven deep learning method can be improved.

Below in conjunction with accompanying drawing, the present invention is described in detail.

Example 1

The embodiment of the present invention provides a data-driven inter-frequency intelligent intervention signal representation method, which is used to extract signal features. The extracted signal features can accurately represent the characteristics of the signal, and the data-driven inter-frequency intelligent intervention signal representation method is also It can be called a data-driven mutual-frequency intelligent intervention signal extraction method. As shown in Figure 1, the method includes:

S101: Obtain the time-frequency characteristics of the frequency conversion signal.

In the embodiment of the present invention, in the illustration of the frequency conversion signal, the axis of abscissa represents time, and the axis of ordinate represents amplitude in time domain. In the graph of time-frequency characteristics, the abscissa represents time, and the ordinate represents frequency. as shown in picture 2.

The specific way to obtain the time-frequency characteristics of the frequency conversion signal is:

The data s(m) is obtained after carrier frequency estimation, down-conversion, sampling and other processing of the frequency conversion signal, and the sliding window W(m) is constructed, and the finite number of points in the window is Fourier transformed to obtain the frequency conversion signal, namely

Among them, the frequency conversion signal is a non-stationary signal, S(f, k) represents the time-frequency characteristics; M represents the data length of the frequency conversion signal, M is a positive integer greater than 2, m represents the serial number of the data, and m is a non-stationary signal less than M-1 Negative integer, f is frequency, k is time (or window sliding step). j is the mathematical representation that constitutes an imaginary number, called the j operator, usually j=sqrt(-1), and the j operator means to rotate a complex number 90 degrees counterclockwise.

S102: Extract low-frequency components and high-frequency components in the time-frequency feature.

In the embodiment of the present invention, the low-frequency component refers to a component whose frequency is lower than or equal to a set threshold in the time-frequency feature, and the set threshold may be 10 Hz. The high-frequency component represents a component whose frequency is higher than a set threshold in the time-frequency feature. as shown in picture 2.

S103: Perform interactive processing on the low-frequency component and the high-frequency component by using a pre-trained high-low-frequency intervention model to obtain a high-frequency output component and a low-frequency output component.

In the embodiment of the present invention, the high-frequency component sequence including multiple high-frequency components and the low-frequency component sequence including multiple low-frequency components may be input to the high-low frequency intervention model. A plurality of high-frequency components are arranged in chronological order to form a high-frequency component sequence, and a plurality of low-frequency components are arranged in chronological order to form a low-frequency component sequence.

Among them, the high-low frequency intervention model includes high-frequency network and low-frequency network, the input of high-frequency network is high-frequency component, the input of low-frequency network is low-frequency component, the output of high-frequency network is high-frequency output component, and the output of low-frequency network is low-frequency output Component; the high-frequency network includes multiple high-frequency units, and the low-frequency network includes multiple low-frequency units. The output of one or more high-frequency units interferes with the input of one or more low-frequency units, and one or more low-frequency The output of the unit performs an intervention operation on the input of one or more high-frequency units, so as to realize the interactive processing of the low-frequency component and the high-frequency component by the high-low frequency intervention model.

By adopting the above scheme, the output of one or more high-frequency units interferes with the input of one or more low-frequency units, and the output of one or more low-frequency units interferes with the input of one or more high-frequency units. Operation, in order to realize the interactive processing of low-frequency components and high-frequency components by the high-low-frequency intervention model, that is, the intervention of high-frequency components on low-frequency components, and the intervention of low-frequency components on high-frequency components, the obtained high-frequency components are fused with low-frequency components The implicit information of the low-frequency component is fused with the implicit information of the high-frequency component, so that the high-frequency component and the low-frequency component can represent the signal information more accurately, and the signal features obtained based on the high-frequency component and the low-frequency component can effectively represent the signal.

In the embodiment of the present invention, the high-frequency component may be used to intervene on the low-frequency component first, or the low-frequency component may be used to intervene on the high-frequency component first.

As an optional implementation manner, after step S103, the data-driven inter-frequency intelligent intervention signal representation method further includes:

A splicing operation is performed on the high-frequency output component and the low-frequency output component to obtain a spliced signal.

Wherein, the splicing operation may use a splicing function concat (high frequency output component, low frequency output component).

A convolution operation is performed on the spliced signal to obtain a restored time-frequency domain signal.

It can be expressed as: recovering time-frequency domain signal=conv (splicing signal), that is, conv(concat(high-frequency output component, low-frequency output component)). Wherein, the dimension of restoring time-frequency domain signal is the same as the dimension of time-frequency feature.

By adopting the above solution, the obtained restored time-frequency domain signal can accurately represent the signal characteristics, and the ability and accuracy of representing the signal by the restored frequency-domain signal are improved.

As an optional implementation manner, the high-frequency network includes 2 high-frequency units, and the low-frequency network includes 2 low-frequency units;

The two low frequency units are respectively the first LSTM network and the third LSTM network. The two high-frequency units are respectively the second LSTM network and the fourth LSTM network.

Wherein, the input of the first LSTM network is a low frequency component, and the input of the second LSTM network is a high frequency component.

The input of the third LSTM network includes the first fusion component obtained by performing an intervention operation on the output of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network. The input of the fourth LSTM network includes the second fusion component obtained by performing an intervention operation on the output of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network.

In the embodiment of the present invention, the first LSTM network, the second LSTM network, the third LSTM network and the fourth LSTM network all adopt a Long Short-Term Memory (LSTM) structure. That is, each low-frequency unit and each high-frequency unit is an LSTM network.

The LSTM network structure includes three gating mechanisms, namely the input gate _it , the output gate o _t and the forgetting gate f _t . The input gate is used to control the input at the current moment and how much information needs to be saved in the state output at the previous moment, the output gate is used to determine how much information the internal state at the current moment needs to be output at the current moment, and the forget gate is used to control the previous state How much information needs to be forgotten about the internal state of the moment. Its specific network structure is shown in Figure 3.

The calculation formulas of the three gates of the LSTM network structure are as follows: (1), (2) and (3):

i _t = σ(W _ih h _t-1 +W _ix x _t +b _i ) (1)

o _t ＝σ(W _oh h _t-1 +W _ox x _t +b _o ) (2)

f _t ＝σ(W _fh h _t-1 +W _fx x _t +b _f ) (3)

Among them, σ(·) represents the Logistic function, t represents the current moment, h _t-1 is the state output at the previous moment, x _t is the input at the current moment, W represents the weight matrix corresponding to h _t-1 and x _t , specifically of: W _ih represents the weight matrix of h _t-1 in the input gate, W _ix represents the weight matrix of x _t in the input gate, and _bi represents the bias vector of the input gate. W _oh represents the weight matrix of h _t-1 in the output gate, W _ox represents the weight matrix of x _t in the output gate, and b _o represents the bias vector of the output gate. W _fh represents the weight matrix of h _t-1 in the forget gate, W _fx represents the weight matrix of x _t in the forget gate, and b _f represents the bias vector of the forget gate.

In the LSTM network, the candidate state _{c_int} is introduced, that is, the input is converted, and the calculation formula is shown in formula (4):

c_in _t ＝tanh(W _ch h _t-1 +W _cx x _t +b _c ) (4)

In the interim, c_in _t represents the candidate state of the LSTM network at time t, W _ch represents the weight matrix of h _t-1 in the candidate state, W _cx represents the weight matrix of x _t in the candidate state, and b _c represents the bias vector in the candidate state.

The candidate state is obtained by a nonlinear activation function.

The internal state c _t at the current moment is shown in formula (5):

c _t = f _t ⊙c _t-1 +i _t ⊙c_in _t (5)

Among them, c _t represents the internal state at the current moment, ⊙ represents the product of vector elements, and c _t-1 is the internal state at the previous moment. c _t transfers information to the external output of the unit while controlling the internal information circulation of the unit, which contains the information accumulated in the previous state.

The output state h _t of the LSTM network is:

h _t ＝o _t ⊙tanh(c _t ) (6)

The output values of the three gates of the LSTM network are all between 0 and 1, and the information transmission is controlled by a certain ratio. For the forgetting gate f _t , the closer its value is to 0, the more the state information of the previous moment is forgotten, but the _{c_int} controlled by the input _gate it will still affect the current state.

To sum up, the operation mode of the basic unit of an LSTM network is as follows: firstly, three kinds of gates are calculated by inputting x _t and the output state h _t-1 at the previous moment; then, the state c_in _t is calculated by the values of two kinds of gates, And update the internal state c _t at the current moment; finally, calculate the output state h _t at the current moment through the output gate and the nonlinear function.

Since the memory unit in LSTM selectively forgets the information at the previous moment, the time span for storing information is longer than the short-term memory for state rewriting at each moment, and shorter than the long-term memory for updating parameters from the entire training data set, so it is called called the long short-term memory network. Compared to the recursive calculation of the system state established by the RNN, the three gates establish a self-loop for the internal state of the LSTM unit. This gating mechanism of LSTM can store long-term representation information and establish long-distance temporal dependencies, thereby realizing time-series feature learning.

After the above-mentioned introduction to the memory unit in LSTM, it can be seen that the input of the LSTM structural unit includes the state output h _t-1 of the previous moment (last moment) and the input of x _t at the current moment, and the third LSTM network is an LSTM structure. Its operation mode has been explained clearly above. In the embodiment of the present invention, the input of the third LSTM network includes the first fusion component obtained after the output of the first LSTM network is intervened based on the output of the second LSTM network and/or the output of the fourth LSTM network, then the following The input and output of the third LSTM network will be introduced in detail with reference to FIG. 4 .

As an optional implementation, the low-frequency component and the high-frequency component are interactively processed through the pre-trained high-low frequency intervention model to obtain the high-frequency output component and the low-frequency output component, including:

Based on the output of the second LSTM network and/or the output of the fourth LSTM network, an intervention operation is performed on the output of the first LSTM network to obtain the first fusion component. As shown in Figure 4.

The first fusion component is used as the input of the third LSTM network.

Based on the output of the first LSTM network and/or the output of the third LSTM network, an intervention operation is performed on the output of the second LSTM network to obtain a second fusion component. As shown in Figure 5.

The second fusion component is used as the input of the fourth LSTM network.

Optionally, the output of the third LSTM network is used as the final output low-frequency output component. The output of the fourth LSTM network is used as the high-frequency output component of the final output

As another optional implementation manner, based on the output of the second LSTM network and/or the output of the fourth LSTM network, an intervention operation is performed on the output of the third LSTM network to obtain the final output low-frequency output component. Based on the output of the first LSTM network and/or the output of the third LSTM network, an intervention operation is performed on the output of the fourth LSTM network to obtain the final output high-frequency component.

Combined with the introduction of the above-mentioned LSTM network structure, in the embodiment of the present invention, for the third LSTM network, its input includes the low-frequency component and the first fusion component at the current moment, that is, the first fusion component is used as the above-mentioned formula (1)～ For h _t-1 in (6), the low-frequency component at the current moment is taken as x _t , then h _t output in formula (6) is the low-frequency output component extracted in the embodiment of the present application.

Similarly, for the fourth LSTM network, its input includes the high-frequency component and the second fusion component at the current moment, that is, the second fusion component is used as h _t-1 in the above formulas (1)~(6), and the current moment The high-frequency component of is used as x _t , then h _t output in formula (6) is the high-frequency output component extracted in the embodiment of the present application.

As an optional implementation manner, based on the output of the second LSTM network, an intervention operation is performed on the output of the first LSTM network to obtain the first fusion component, including:

The output of the first LSTM network and the output of the second LSTM network are connected to obtain the first connection component; the convolution operation is performed on the first connection component to obtain the first fusion component. For details, please refer to formula (7):

l1=conv(concat(H1,L1)) (7)

Among them, l1 represents the first fusion component, H1 represents the output of the second LSTM network, L1 represents the output of the first LSTM network, concat() represents the connection operation between the output of the first LSTM network and the output of the second LSTM network, conv () indicates a convolution operation, the main purpose of which is to reduce the dimension of the first connected component, so that the dimension of the first connected component is the same as the output of the first LSTM network.

Perform an intervention operation on the output of the first LSTM network based on the output of the fourth LSTM network to obtain the first fusion component. The above-mentioned intervention operation on the output of the first LSTM network based on the output of the second LSTM network obtains the first fusion component. Similarly, it is only necessary to represent H1 as the output of the fourth LSTM network.

Based on the output of the second LSTM network and the output of the fourth LSTM network, an intervention operation is performed on the output of the first LSTM network to obtain the first fusion component, including:

performing a connection operation on the output of the first LSTM network and the output of the second LSTM network to obtain a first connection component;

performing a connection operation on the output of the first LSTM network and the output of the fourth LSTM network to obtain a second connection component;

performing a connection operation on the first connected component and the second connected component to obtain a third connected component;

A convolution operation is performed on the third connected component to obtain a first fusion component; the dimension of the first fusion component is the same as the dimension of the low frequency component.

Specifically, based on the output of the second LSTM network and the output of the fourth LSTM network, the output of the first LSTM network is intervened to obtain the operation mode of the first fusion component as shown in formula (8):

l1=conv(concat(concat(H1,L1),concat(H2,L1))) (8)

Wherein, H2 represents the output of the fourth LSTM network.

As an optional implementation, the output of the first LSTM network is intervened based on the output of the second LSTM network and the output of the fourth LSTM network to obtain the first fusion component. The operation mode is shown in formula (9):

l1=conv(concat(conv(concat(H1,L1)),conv(concat(H2,L1)))) (9)

That is, based on the output of the second LSTM network and the output of the fourth LSTM network, an intervention operation is performed on the output of the first LSTM network to obtain the first fusion component, including: connecting the output of the first LSTM network and the output of the second LSTM network operation to obtain the first connection component; the convolution operation is performed on the first connection component to obtain the first dimension reduction component, and the dimension of the first dimension reduction component is the same as the dimension of the output of the first LSTM network; the output of the first LSTM network and The output of the fourth LSTM network is connected to obtain a second connection component; the second connection component is convoluted to obtain a second dimension reduction component, and the dimension of the second dimension reduction component is the same as the dimension of the output of the first LSTM network; performing a connection operation on the first dimensionality reduction component and the second dimensionality reduction component to obtain a third connection component; performing a convolution operation on the third connection component to obtain a first fusion component; the dimensions of the first fusion component and the first fusion component The dimensions of the output of an LSTM network are the same.

In the embodiment of the present invention, for the operation of obtaining the first fused component, if the dimensions are the same, concatenate and then convolve; if the dimensions are different, convolute and then concatenate and convolute.

In the embodiment of the present invention, based on the output of the first LSTM network and/or the output of the third LSTM network, an intervention operation is performed on the output of the second LSTM network to obtain the second fusion component. The output and\or the output of the fourth LSTM network intervene in the output of the first LSTM network to obtain the first fusion component.

Through the above solution, the intervention of the low frequency component on the high frequency component and the intervention of the low frequency component through the high frequency component can be realized, and the extracted high frequency output component and low frequency output component can accurately represent the characteristics of the signal, improving the effectiveness of signal extraction.

As yet another optional embodiment, the high-frequency network includes 3 high-frequency units, and the low-frequency network includes 3 low-frequency units;

The three low-frequency units are respectively the first LSTM network, the third LSTM network and the fifth LSTM network; the three high-frequency units are respectively the second LSTM network, the fourth LSTM network and the sixth LSTM network;

The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component.

The input of the third LSTM network includes the first fusion component obtained after the output of the first LSTM network is intervened based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network; The input of the fifth LSTM network includes the third fusion component obtained after the output of the third LSTM network is intervened based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network;

The input of the fourth LSTM network includes the second fusion component obtained after the output of the second LSTM network is intervened based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network;

The input of the sixth LSTM network includes the fourth fusion component obtained by intervening the output of the fourth LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network.

That is, through the output of the second LSTM network in the high-frequency network and\or the output of the fourth LSTM network and\or the output of the sixth LSTM network to the output of the first LSTM network in the low-frequency network and\or the output of the third LSTM network The output and\or the output of the fifth LSTM network are intervened, as shown in Figure 6. Through the output of the first LSTM network in the low frequency network and\or the output of the third LSTM network and\or the output of the fifth LSTM network to the output of the second LSTM network in the high frequency network and\or the output of the fourth LSTM network and \or the output of the sixth LSTM network for intervention operations. As shown in Figure 7.

In the embodiment of the present invention, the low-frequency component and the high-frequency component are interactively processed through the pre-trained high-low frequency intervention model to obtain the high-frequency output component and the low-frequency output component, which may also include:

Based on the output of the second LSTM network and\or the output of the fourth LSTM network and\or the output of the sixth LSTM network, an intervention operation is performed on the output of the first LSTM network to obtain the first fusion component.

According to the above formulas (1)-(6), the first fusion component is used as h _t-1 in the input of the third LSTM network. In addition to referring to any of the methods shown in the above formulas (7) to (9) for the specific method of the intervention operation, similarly, the following methods can also be referred to:

The first fusion component can be obtained according to any one of the formulas (10), (11):

l1=conv(concat(concat(H1,L1),concat(H2,L1,concat(H3,L1))) (10)

l1=conv(concat(conv(concat(H1,L1)),conv(concat(H2,L1)),conv(concat(H3,L1))))(11)

Among them, H3 represents the output of the sixth LSTM network.

Based on the output of the second LSTM network and\or the output of the fourth LSTM network and\or the output of the sixth LSTM network, an intervention operation is performed on the output of the third LSTM network to obtain a third fusion component.

According to the above formulas (1)-(6), the third fusion component is used as h _t-1 in the input of the fifth LSTM network.

Based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network, an intervention operation is performed on the output of the second LSTM network to obtain the second fusion component.

According to the above formulas (1)-(6), the second fusion component is used as h _t-1 in the input of the fourth LSTM network.

Based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network, an intervention operation is performed on the output of the fourth LSTM network to obtain a fourth fusion component.

According to the above formulas (1)-(6), the fourth fusion component is used as h _t-1 in the input of the sixth LSTM network.

Optionally, the output of the fifth LSTM network is used as the final output low-frequency output component, and the output of the sixth LSTM network is used as the final output high-frequency output component.

In the embodiment of the present invention, the manner of obtaining the second fusion component, the third fusion component, and the fourth fusion component is similar to the manner of obtaining the first fusion component, which will not be repeated here.

As another optional implementation, based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network, the output of the fifth LSTM network is intervened to obtain the final output The low-frequency output component is based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network to obtain the final high-frequency output component after the output of the sixth LSTM network is intervened.

In summary, in the data-driven inter-frequency intelligent intervention signal representation method provided by the embodiment of the present invention, the high-low frequency intervention model includes a high-frequency network and a low-frequency network, the input of the high-frequency network is a high-frequency component, and the input of the low-frequency network is the low-frequency component, the output of the high-frequency network is the high-frequency output component, and the output of the low-frequency network is the low-frequency output component;

The high-frequency network includes N high-frequency LSTM networks, and the low-frequency network includes N low-frequency LSTM networks; N high-frequency LSTM networks are connected end-to-end, and N low-frequency LSTM networks are connected end-to-end; N is a positive integer greater than or equal to 2;

In the low-frequency network, the input of the first low-frequency LSTM network is the low-frequency component at the first moment, and the input of the n-th low-frequency LSTM network includes: the n-th high-frequency intervention low-frequency fusion component and the low-frequency component at the n-th moment; The intervening low-frequency fusion component is the component obtained after performing an intervention operation on the output of the n-1th low-frequency LSTM network based on the output of all high-frequency LSTM networks in the high-frequency network;

In the high-frequency network, the input of the first high-frequency LSTM network is the high-frequency component at the first moment, and the input of the n-th high-frequency LSTM network includes: the nth low-frequency intervention high-frequency fusion component and the high-frequency component at the nth moment ; The n-th low-frequency intervention high-frequency fusion component is the component obtained after performing an intervention operation on the output of the n-1 high-frequency LSTM network based on the output of all low-frequency LSTM networks in the low-frequency network; n is greater than 1 and less than or equal to N positive integer of .

Both the high-frequency LSTM network and the low-frequency LSTM network are LSTM network structures as shown in Figure 2.

Based on the output of all high-frequency LSTM networks in the high-frequency network, the output of the n-1th low-frequency LSTM network is intervened, including:

Connect the output of the n-1th low-frequency LSTM network to the output of each high-frequency LSTM network to obtain N first connection components;

Perform convolution dimension reduction operation on the first connection component to obtain the first convolution integral; the dimension of the first convolution integral is the same as the output dimension of the n-1th low-frequency LSTM network; N first connection components correspond to N A volume of the first volume;

performing a connection operation on the N first convolutional components to obtain a convolutional connection component;

Perform convolution dimensionality reduction operation on the convolutional connection components to obtain the nth low-frequency intervention high-frequency fusion component; the dimension of the n-th low-frequency intervention high-frequency fusion component is the same as the output dimension of the n-1th low-frequency LSTM network.

The above-mentioned first LSTM network, third LSTM network, and fifth LSTM network are low-frequency LSTM networks; the second LSTM network, fourth LSTM network, and sixth LSTM network are high-frequency LSTM networks.

With reference to Fig. 6 and Fig. 7, the mutual intervention operation and interactive operation between the high-frequency signal and the low-frequency signal in the high-low frequency intervention model are described.

For low frequency networks:

For the t-1th moment, its input includes x _t-1 , where x _t-1 represents the low frequency component at the t-1th moment.

For the tth moment, t is a positive integer greater than 2, then the input of the nth low-frequency LSTM network corresponding to the tth moment is the low-frequency component x _t of the tth moment and the output of the high-frequency LSTM network of the high-frequency network to the n-1th The output h _t-1 of the low-frequency LSTM network is the result of the intervention operation l _t , that is, replace h _t-1 in formulas (1) to (6) with l _t . The way of obtaining l _t is the way of obtaining l1 shown in any one of formulas (7) to (9).

For the t+1th moment, the input of the n+1th low-frequency LSTM network corresponding to the t+1th moment is the low-frequency component x _t+1 of the t+1th moment and the output of the high-frequency LSTM network of the high-frequency network to the nth The output h _t of the low-frequency LSTM network is the result of the intervention operation l _t+1 , that is, replace h _t-1 in formulas (1) to (6) with l t+ ₁ . l _t+1 is obtained in the same manner as l1 shown in any one of formulas (7) to (11).

For high frequency networks:

For the t-1th moment, its input includes y _t-1 , and y _t-1 represents the high-frequency component at the t-1th moment.

For the tth moment, t is a positive integer greater than 2, then the input of the nth high-frequency LSTM network corresponding to the tth moment is the high-frequency component y _t of the tth moment and the output of the low-frequency LSTM network of the low-frequency network to the n-1th The output z _t-1 of the high-frequency LSTM network is the result of the intervention operation g _t , that is, replace h _t-1 in formulas (1) to (6) with g _t . The way to obtain g _t is the way to obtain l1 shown in any one of formulas (7) to (9).

For the t+1th moment, the input of the n+1th high-frequency LSTM network corresponding to the t+1th moment is the high-frequency component y _t+1 of the t+1th moment and the output of the low-frequency LSTM network of the low-frequency network to the nth The output z _t of the high-frequency LSTM network is the result of the intervention operation g _t , that is, replace h _t-1 in formulas (1) to (6) with g _t . The way to obtain g _t is the way to obtain l1 shown in any one of formulas (7) to (11).

In summary, the technical solutions shown in Figure 6 and Figure 7 fully demonstrate the interaction between high-frequency signals and low-frequency signals, and the high- and low-frequency intervention models are trained through the interaction of high-frequency components and low-frequency components to realize high-frequency or low-frequency knowledge. Driving the training process of the deep learning model improves the performance of the original data-driven deep learning method, and the obtained high-frequency output components and low-frequency output components can accurately represent the signal characteristics.

In the embodiment of the present invention, the high-frequency network and the low-frequency network are trained at the same time, that is, the interaction between the high-frequency component and the low-frequency component, the high-frequency network and the low-frequency network are carried out simultaneously until the output converges, then the network training Finish.

Example 2

Based on the data-driven inter-frequency intelligent intervention signal representation method provided by the above-mentioned embodiments, an embodiment of the present invention provides a data-driven inter-frequency intelligent intervention signal representation system for performing the above method, and the system includes:

The obtaining module is used to obtain the time-frequency characteristics of the frequency conversion signal.

The extraction module is used to extract low-frequency components and high-frequency components in the time-frequency feature.

The interaction module is used to interactively process the low-frequency component and the high-frequency component through a pre-trained high-low-frequency intervention model to obtain a high-frequency output component and a low-frequency output component.

The splicing module is configured to splice the high-frequency output component and the low-frequency output component to obtain a spliced signal.

The restoration module is configured to perform a convolution operation on the spliced signal to obtain a restored time-frequency domain signal.

Wherein, the dimension of the recovered time-frequency domain signal is the same as the dimension of the time-frequency feature. The high-low frequency intervention model includes a high-frequency network and a low-frequency network. The input of the high-frequency network is the high-frequency component, the input of the low-frequency network is the low-frequency component, the output of the high-frequency network is the high-frequency output component, and the output of the low-frequency network is the low-frequency output component. The high-frequency network includes multiple high-frequency units, and the low-frequency network includes multiple low-frequency units. The output of one or more high-frequency units intervenes with the input of one or more low-frequency units, and the output of one or more low-frequency units The output performs an intervention operation on the input of one or more high-frequency units, so as to realize the interactive processing of the low-frequency component and the high-frequency component by the high-low frequency intervention model.

Regarding the model in the above embodiment, the specific manner in which each module executes operations has been described in detail in the embodiment of the method, and will not be described in detail here.

The embodiment of the present invention also provides a complex signal multi-component interactive characteristic signal processing system, as shown in FIG. 8 , including a memory 504, a processor 502, and a computer program stored in the memory 504 and operable on the processor 502, When the processor 502 executes the program, steps in any method of the data-driven inter-frequency intelligent intervention signal representation method described above are implemented.

Wherein, in FIG. 8, the bus architecture (represented by bus 500), bus 500 may include any number of interconnected buses and bridges, and 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 thus will not be further described herein. The bus interface 505 provides an interface between the bus 500 and the receiver 501 and the transmitter 503 . Receiver 501 and transmitter 503 may be the same element, a transceiver, providing means for communicating with various other devices over a transmission medium. Processor 502 is responsible for managing bus 500 and general processing, while memory 504 may be used to store data used by processor 502 in performing operations.

An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of any method of the data-driven inter-frequency intelligent intervention signal representation method described above are implemented. .

The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other device. Various generic systems can also be used with the teachings based on this. The structure required to construct such a system is apparent from the above description. Furthermore, the present invention is not specific to any particular programming language. It should be understood that various programming languages can be used to implement the content of the present invention described herein, and the above description of specific languages is for disclosing the best mode of the present invention.

In the description provided herein, numerous specific details are set forth. However, it is understood 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 the understanding of this description.

Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the invention, in order to streamline this disclosure and to facilitate an understanding of one or more of the various inventive aspects, various features of the invention are sometimes grouped together in a single embodiment, figure, or its description. This method of disclosure, however, is not to be interpreted as reflecting an intention that the claimed invention 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 can 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. Modules or units or components in the embodiments may be combined into one module or unit or component, and furthermore may be divided into a plurality of sub-modules or sub-units or sub-assemblies. All features disclosed in this specification (including accompanying claims, abstract and drawings) and any method or method so disclosed may be used in any combination, except that at least some of such features and/or processes or units 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, those skilled in the art will appreciate that although some embodiments herein include some features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of the invention. And form different embodiments. For example, in the following claims, any of the claimed embodiments may be used in any combination.

Claims (10)

1. A data-driven mutual frequency intelligent intervention signal representation method, characterized in that the method comprises: Obtain the time-frequency characteristics of the frequency conversion signal; Extract low-frequency components and high-frequency components in time-frequency features; Through the pre-trained high and low frequency intervention model, the low frequency component and the high frequency component are interactively processed to obtain the high frequency output component and the low frequency output component; Among them, the high-low frequency intervention model includes high-frequency network and low-frequency network, the input of high-frequency network is high-frequency component, the input of low-frequency network is low-frequency component, the output of high-frequency network is high-frequency output component, and the output of low-frequency network is low-frequency output Component; the high-frequency network includes multiple high-frequency units, the low-frequency network includes multiple low-frequency units, and the input of one or more low-frequency units is intervened through the output of one or more high-frequency units, and through one or more low-frequency units The output of the input of one or more high-frequency units is intervened to realize the interactive processing of low-frequency components and high-frequency components by the high-low-frequency intervention model. 2. The data-driven inter-frequency intelligent intervention signal representation method according to claim 1, wherein the high-frequency network comprises 2 high-frequency units, and the low-frequency network comprises 2 low-frequency units; The two low-frequency units are respectively the first LSTM network and the third LSTM network; the two high-frequency units are respectively the second LSTM network and the fourth LSTM network; The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component; The input of the third LSTM network includes the first fusion component obtained after the output intervention operation of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network; The input of the fourth LSTM network includes the second fusion component obtained after intervention operation on the output of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network. 3. The data-driven mutual-frequency intelligent intervention signal representation method based on claim 2, wherein the output intervention operation of the first LSTM network based on the output of the second LSTM network obtains the first fusion component, including: performing a connection operation on the output of the first LSTM network and the output of the second LSTM network to obtain a first connection component; A convolution operation is performed on the first connected component to obtain a first fusion component. 4. The data-driven inter-frequency intelligent intervention signal representation method according to claim 2, wherein the input of the third LSTM network comprises an output based on the output of the second LSTM network and the output of the fourth LSTM network to the first LSTM The output of the network is the first fusion component after the intervention operation, including: performing a connection operation on the output of the first LSTM network and the output of the second LSTM network to obtain a first connection component; performing a connection operation on the output of the first LSTM network and the output of the fourth LSTM network to obtain a second connection component; performing a connection operation on the first connected component and the second connected component to obtain a third connected component; A convolution operation is performed on the third connected component to obtain a first fusion component; the dimension of the first fusion component is the same as the dimension of the low frequency component. 5. The data-driven inter-frequency intelligent intervention signal representation method according to claim 1, wherein the high-frequency network comprises 3 high-frequency units, and the low-frequency network comprises 3 low-frequency units; The three low-frequency units are respectively the first LSTM network, the third LSTM network and the fifth LSTM network; the three high-frequency units are respectively the second LSTM network, the fourth LSTM network and the sixth LSTM network; The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component; The input of the third LSTM network includes the first fusion component obtained after the output intervention operation of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network; The input of the fourth LSTM network includes the second fusion component obtained after the output intervention operation of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network; The input of the fifth LSTM network includes the third fusion component based on the output of the second LSTM network and/or the output of the fourth LSTM network and/or the output of the sixth LSTM network after the intervention operation on the output of the third LSTM network; The input of the sixth LSTM network includes the fourth fusion component obtained by intervening the output of the fourth LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network and/or the output of the fifth LSTM network. 6. The data-driven mutual-frequency intelligent intervention signal representation method according to claim 1, wherein the low-frequency component and the high-frequency component are interactively processed by the pre-trained high-low frequency intervention model to obtain high After the high-frequency output component and the low-frequency output component, the data-driven inter-frequency intelligent intervention signal representation method also includes: Splicing the high-frequency output component and the low-frequency output component to obtain a spliced signal; performing a convolution operation on the spliced signal to obtain a restored time-frequency domain signal; Wherein, the dimension of the recovered time-frequency domain signal is the same as the dimension of the time-frequency feature. 7. A data-driven inter-frequency intelligent intervention signal representation system, characterized in that the system includes: Obtaining module, used for obtaining the time-frequency characteristic of the frequency conversion signal; An extraction module is used to extract low-frequency components and high-frequency components in the time-frequency feature; The interaction module is used to interactively process the low-frequency component and the high-frequency component through a pre-trained high-low frequency intervention model to obtain a high-frequency output component and a low-frequency output component; Among them, the high-low frequency intervention model includes high-frequency network and low-frequency network, the input of high-frequency network is high-frequency component, the input of low-frequency network is low-frequency component, the output of high-frequency network is high-frequency output component, and the output of low-frequency network is low-frequency output Component; the high-frequency network includes multiple high-frequency units, the low-frequency network includes multiple low-frequency units, and the input of one or more low-frequency units is intervened through the output of one or more high-frequency units, and through one or more low-frequency units The output of the input of one or more high-frequency units is intervened to realize the interactive processing of low-frequency components and high-frequency components by the high-low-frequency intervention model. 8. The data-driven inter-frequency intelligent intervention signal representation system according to claim 7, wherein the high-frequency network includes 2 high-frequency units, and the low-frequency network includes 2 low-frequency units; The two low-frequency units are respectively the first LSTM network and the third LSTM network; the two high-frequency units are respectively the second LSTM network and the fourth LSTM network; The input of the first LSTM network is the low frequency component, and the input of the second LSTM network is the high frequency component; The input of the third LSTM network includes the first fusion component obtained after the output intervention operation of the first LSTM network based on the output of the second LSTM network and/or the output of the fourth LSTM network; The input of the fourth LSTM network includes the second fusion component obtained after intervention operation on the output of the second LSTM network based on the output of the first LSTM network and/or the output of the third LSTM network. 9. The data-driven inter-frequency intelligent intervention signal representation system according to claim 8, wherein the system further comprises: The splicing module is used to splice the high-frequency output component and the low-frequency output component to obtain a spliced signal; A restoration module, configured to perform a convolution operation on the spliced signal to obtain a restored time-frequency domain signal; Wherein, the dimension of the recovered time-frequency domain signal is the same as the dimension of the time-frequency feature. 10. An electronic device, characterized in that the electronic device comprises a memory, a processor, and a computer program stored on the memory and operable on the processor, and the processor implements claim 1- when executing the program. 6. The steps of any one of the methods.