A method for extracting radio frequency features of radar radiation sources

Abstract

This invention discloses a method for extracting radio frequency (RF) features from radar radiation sources. The method involves acquiring and preprocessing the RF signal from which features are to be extracted, then inputting it into pre-trained networks 1 and 2 for feature extraction. Both networks 1 and 2 are stacked autoencoder networks and contain multi-layer encoders and decoders of the same number of layers. The last encoder layer of network 2 outputs the extracted RF features. The training process includes inputting an ideal RF signal dataset and a modeled RF signal dataset into networks 1 and 2 respectively for hierarchical pre-training and global training until training is complete, resulting in trained networks 1 and 2. This invention can separate the signal features and RF features of a radiation source signal under the constraint of signal features, resulting in more complete extracted RF features. The extracted RF features can be used in applications requiring high accuracy of RF features, such as individual radiation source identification and interference suppression.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar emitter signal processing, and relates to a radar emitter radio frequency feature extraction method, in particular to a radar emitter radio frequency feature extraction method based on a double network. BACKGROUND

[0002] Due to the influence of multiple factors such as phase noise, frequency offset of the local oscillator source of the radar emitter device, and nonlinearity of the power amplifier of the radio frequency link, the radio frequency feature, which is a physical layer inherent attribute from the hardware device, is generated. Even if the devices from the same manufacturer and the same model, due to the unique, non-replicable, stable and extractable characteristics of the radio frequency feature, the radar transmitter can still be individually distinguished by the radio frequency feature contained therein. Therefore, the radio frequency feature has broad application prospects in emitter individual identification, model identification interference suppression, and interference identification. The current common radio frequency feature extraction methods include traditional feature extraction methods and neural network feature extraction methods. The traditional feature extraction methods include variational mode decomposition method, empirical mode decomposition method, wavelet transform method, or extraction of short-time autocorrelation function and short-time power spectral density of signals, the former method has a complex feature extraction process, consumes a long time, and brings a large calculation load, and the latter method is only suitable for steady-state signals or non-stationary signals with small change speed, and the completeness and accuracy of the features designed by the traditional method are not high. The generalization ability of the features extracted by the traditional method is insufficient in specific scenarios. The current commonly used neural network for extracting radio frequency features mainly includes convolutional neural network, autoencoder network, and long short-term memory network. The above network cannot completely separate the radio frequency feature from the signal feature, and the extracted radio frequency feature may be mixed with other signal features. SUMMARY

[0003] In view of the above prior art, the technical problem to be solved by the present application is to provide a radar emitter radio frequency feature extraction method based on a double network, which solves the radio frequency (RF) feature extraction of the radio frequency signal in a complex electromagnetic environment.

[0004] To solve the above technical problem, the radar emitter radio frequency feature extraction method of the present application comprises:

[0005] The radio frequency signal to be extracted is obtained and preprocessed, and the preprocessed radio frequency signal is input into a pre-trained network 1 and network 2 for feature extraction, wherein the network 1 and the network 2 are both stack autoencoder networks and both contain multiple layers of the same number of encoders and decoders, and the last layer of the encoder of the network 2 outputs the extracted radio frequency feature;

[0006] The training process of the network 1 and the network 2 comprises:

[0007] Step 1, the modeling radio frequency signal containing radio frequency features generated by the radar radiation source model built through simulation is taken as the training set of network 2, and the training set of network 1 with the same composition as the training set of network 2 and not containing the radio frequency features is taken as the training set of network 1;

[0008] Step 2, the training set of network 1 is input to network 1 after preprocessing, first, the single-layer encoder-decoder is pre-trained layer by layer according to the set training times of each layer, after the layer-by-layer pre-training is completed, global training of network 1 is carried out, and after the set global training times are reached, the network 1 model is saved; wherein, the signal features output by each layer of encoder in the training process are taken as the input of the next layer of encoder, the loss function of the single-layer encoder-decoder is used to initialize the weight and bias parameters of each layer of network in the layer-by-layer pre-training, and the global loss function is used to update the weight and bias parameters of network 1 in the global training;

[0009] Step 3, the training set of network 2 is input to network 2 after preprocessing, first, the single-layer encoder-decoder is pre-trained layer by layer according to the set training times of each layer, after the layer-by-layer pre-training is completed, global training of network 2 is carried out, and after the set global training times are reached, the network 2 model is saved; wherein, the radio frequency features extracted by each layer of encoder of network 2 and the signal features extracted by the corresponding layer of encoder of network 1 are merged as the input of the next layer of encoder of network 2, the loss function of the single-layer encoder-decoder is used to initialize the weight and bias parameters of each layer of network in the layer-by-layer pre-training; and the global loss function is used to update the weight and bias parameters of network 2 in the global training.

[0010] Further, the modeling radio frequency signal includes unintentional modulation information and intentional modulation information.

[0011] Further, the preprocessing of the training set of network 1 and the training set of network 2 includes extracting in-phase / quadrature data from the radio frequency signal, concatenating the extracted data, and then performing normalization processing.

[0012] Further, the loss function of the single-layer encoder-decoder of network 1 and the global loss function are respectively:

[0013]

[0014]

[0015] Wherein, Loss1 is the loss function of the single-layer encoder-decoder, Loss 11 is the global loss function, x1 represents the input of each layer of encoder, is the reconstructed signal obtained by the output of each layer of decoder; y1 represents the input signal of network 1, is the reconstructed signal obtained by the output of the last layer of decoder.

[0016] Further, the loss function and the global loss function of the single-layer codec of the network 2 are respectively:

[0017]

[0018]

[0019] Wherein, Loss2 is the loss function of the single-layer codec, Loss 22 is the global loss function, x2 represents the input of each layer of the encoder, is the reconstruction signal obtained by the output of each layer of the decoder; represents the frequency domain input obtained by the Fourier transform of the time domain input x2, represents the time domain reconstruction data is the frequency domain reconstruction data obtained by the Fourier transform of the time domain reconstruction data, is the reconstruction signal obtained by the output of the last layer of the decoder, represents the frequency domain input obtained by the Fourier transform of the time domain input y2, represents the time domain reconstruction data is the frequency domain reconstruction data obtained by the Fourier transform of the time domain reconstruction data.

[0020] The beneficial effects of the present application: the present application mainly solves the problem of extracting the radio frequency (RF) features of the RF signal in the complex electromagnetic environment, and the extracted RF features can be used for subsequent radiation source individual identification or interference suppression. First, the radar radiation source is modeled by intentional and unintentional modulation; then two stacked auto-encoder (SAE) networks are designed to better automatically extract the RF features of different abstraction levels in the RF signal. Network 1 aims to extract the signal features of the ideal RF signal through an unsupervised manner; then, network 1 and network 2 are combined to learn the potential RF features that can stably represent the radar radiation source from the RF signal under the joint training mode of the time domain and the frequency domain. The double network can extract the RF features that are not sensitive to the changes of the signal parameters. The present application is different from other methods in the following aspects: the extracted RF features are shallow, mixed with other features of the signal, and the RF features are incomplete and have more feature loss due to the complex data preprocessing, etc. The advantages of the double network are that the signal features and the fingerprint features of the radiation source signal can be separated under the constraint of the signal features, the extracted RF features are more complete, and the extracted RF features can be used for radiation source individual identification, interference suppression and other application fields that require high accuracy of the RF features. The present application provides a new idea for communication radiation sources and time domain and frequency domain RF feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flow chart of the present application;

[0022] Figure 2 is a schematic diagram of the radiation source modeling;

[0023] Figure 3 is a schematic diagram of the loss function setting;

[0024] Figure 4 is a schematic diagram of the stack autoencoder;

[0025] Figure 5 is a specific flow of the dual network scheme;

[0026] Figure 6 is a time domain diagram of the original LFM signal;

[0027] Figure 7 is a time domain diagram of the recovered LFM signal;

[0028] Figure 8 is a frequency domain diagram of the original LFM signal;

[0029] Figure 9 is a frequency domain diagram of the recovered LFM signal. DETAILED DESCRIPTION

[0030] The present application will be further described below in conjunction with the accompanying drawings and examples.

[0031] In conjunction with Figure 1 , the present application comprises the following steps:

[0032] 1) Radar radiation source modeling:

[0033] According to the working principle of the radar, the present application uses a simulation platform and measured parameters to build two types of radar radiation source models by selecting different design parameters, and under the same type of model, to build four individual radar radiation source models by slightly adjusting the parameters of each module to add unintentional modulation information. On this basis, by adjusting the working parameters of the signal source model of different modulation methods, intentional modulation information is added to generate RF signals. Therefore, the model can flexibly adjust the working parameters of the radar to provide RF signals carrying radio frequency information. For example, Figure 2As shown, a signal source model is built by using a direct digital synthesizer (DDS), and two types of radar emitter models are built by combining the built signal source model and two different RF link models. In the modeling process, the DDS signal source model generates source signals of different modulation types, adds intentional modulation information, and adds unintentional modulation information due to the phase truncation information of each component and the non-linear conditions of the power amplifier, thereby completing the modeling of different types of radar emitters. Each radar emitter model can generate RF signals under different modulation modes and different operating parameters.

[0034] 2) Construct a dataset:

[0035] In the present application, RF signals of three modulation modes are simulated, namely continuous wave signal (CW), linear frequency modulation signal (LFM) and binary phase shift keying signal (BPSK). By setting the mixed ideal CW signal, LFM signal and BPSK signal as the training data set of network 1, i.e. the RF signal input of network 1 does not include radio frequency characteristics; and setting the mixed modeling CW signal, LFM signal and BPSK signal as the training data set of network 2, since the autoencoder network belongs to an unsupervised network, there is no need to design a data set representing the label. The input training set of network 2 is the RF signal generated by the modeled radar emitter model, so the RF signal input of network 2 includes both signal characteristics and radio frequency characteristics. (Taking the ideal RF signal as an example, 6000 groups of continuous wave signals, linear frequency modulation signals and BPSK signals are generated respectively and arranged in 1 to 6000 rows as continuous wave signals, 6001 rows to 12000 rows as linear frequency modulation signals, and 12001 rows to 18000 rows as BPSK signals mixed as a data set. The data set of RF signal containing radio frequency characteristics is set as above).

[0036] 3) Data preprocessing:

[0037] The in-phase / quadrature (I / Q) data are extracted from the complex RF signal and concatenated, and then normalized. (The parameter setting is a sampling rate of 100 MHz and a pulse width of 20 μs, so the sampling points of the complex RF signal are 2000. That is, the input nodes of the network are 2000).

[0038] 4) Design neural network:

[0039] The ideal RF signal is used as the input of network 1. Each layer of the encoder and the decoder contains three hidden layers. The number of output nodes of the encoder is set to 50, so the number of nodes of each layer of the encoder is 2000-600-150-50; the number of nodes of each layer of the decoder is 50-150-1200-2000.

[0040] The modeled RF signal is taken as the input of the network 2. Each layer of the encoder and the decoder is set to have three hidden layers. The number of output nodes of the encoder is set to 50, and thus the number of nodes of each layer of the encoder is 2000-600-150-50; since the output of each layer of the encoder of the network 1 is involved in the merging, the number of nodes of each layer of the decoder is 100-300-1200-2000.

[0041] The loss function of the network 1 is designed, and the network 1 is an SAE network, and since the input data has no label, the SAE network is an unsupervised learning algorithm.

[0042] The network 1 is set to have a three-layer encoder-decoder network, and the loss function of a single layer during the pre-training is:

[0043]

[0044] In the formula, x1 represents the input signal of each layer of the encoder, is the reconstructed signal obtained by the output of each layer of the decoder. After the pre-training, the network loss function set for the global training of the fine-tuning parameters is:

[0045]

[0046] In the formula, y1 represents the original input signal, is the reconstructed signal obtained by the output of the last layer of the decoder.

[0047] The loss function of the network 2 is designed, the spectrum of the RF signal of different radar radiation source models is analyzed, it is found that the spectrum carries different degrees and different types of stray information, which is a unique attribute of the unintended feature, and the stray information carried is more obvious in the frequency domain, and thus the loss function of the network 2 is jointly subjected to the mean square error constraint in the time domain and the frequency domain, as shown in the formula: Figure 3 so that the radio frequency features can be better extracted from the RF signal.

[0048] The network 2 is set to have a three-layer encoder-decoder network, and the loss function of a single layer of the neural network during the pre-training is:

[0049]

[0050] In the formula, the first term is the mean square error of the time domain input data x2 and the time domain reconstructed data of each layer; the second term is the mean square error of the frequency domain input data obtained by performing Fourier transform on the time domain input data x2 and the frequency domain reconstructed data obtained by performing Fourier transform on the time domain reconstructed data of each layer. After the pre-training, the network loss function set for the global training of the fine-tuning parameters is:

[0051]

[0052] In the formula, y2 represents the original input signal, and y2 is the reconstructed signal obtained by the final decoder output.

[0053] After the signal features and radio frequency features are extracted by network 1 and network 2 respectively, the two networks respectively use the decoder to verify the effect of feature extraction on the recovered signal according to the extracted signal features and radio frequency features. Network 1 uses the mean square error in the time domain to obtain the loss function between the actual mixed ideal signal and the reconstructed mixed ideal signal, and network 2 uses the mean square error in the time domain and frequency domain to obtain the loss function between the actual mixed modeling signal and the recovered mixed modeling signal.

[0054] 5) Training network to extract radio frequency features

[0055] First, the ideal RF signal data set is normalized and input to network 1 to extract signal features through hierarchical pre-training and global training fine-tuning. In the training stage, the mean square error between the original input signal of network 1 and the recovered signal output by the decoder in the time domain is used as the loss function of the neural network to update the weight and bias parameters. After training, the network 1 model is saved.

[0056] The modeling RF signal data set is normalized and input to network 2 to extract radio frequency features through hierarchical pre-training and global training fine-tuning. However, the pre-training stage and the global training fine-tuning stage of each layer of network 2 are different. Because the modeling signal includes signal features and radio frequency features, in order to more accurately extract the radio frequency features, the signal features must be constrained, so the modeling signal data set needs to be input to the saved network 1 model for calling to form a double network training during the pre-training of each layer.

[0057] The two-layer codec is taken as an example to illustrate the pre-training of each layer of the dual network. The output of the modeling signal through the first-layer encoder of network 2 needs to be combined with the signal features of the output of the modeling signal through the first-layer encoder of network 1, and the spliced result is sent to the first-layer decoder of network 2. When the second-layer codec of network 2 is pre-trained, the input of the encoder is the combined output obtained through the previous-layer encoder. Since the output of the first-layer encoder of network 1 is called in the combined output, the signal features are already included in half. Therefore, the features obtained by compressing the combined output by the second-layer encoder are closer to the radio frequency features. As in the operation after the output of the first-layer encoder, the signal features of the output of the modeling signal through the second-layer encoder of network 1 need to be combined with the signal features of the output of the second-layer encoder, and then the combined result of the output of the second-layer encoder is sent to the decoder of the second layer. The pre-training of each subsequent layer is carried out in the same way. The final global training and fine-tuning operation of the dual network directly sends the combined encoding result of each layer to the next-layer encoder until the combined output of the last-layer encoder is obtained. The output of the last-layer encoder is sent to the corresponding decoder, and then the decoding is sequentially performed to obtain the recovered modeling signal.

[0058] Therefore, the extraction of the radio frequency features is mainly carried out in two steps. In the first step, the ideal RF signals (not containing radio frequency features) of different modulation modes are trained and the signal features are extracted through network 1. After the training is completed, the model of network 1 is saved. In the second step, when the radio frequency features of the modeling RF signal are extracted, the signal features extracted by network 1 need to be called to constrain the extraction of the radio frequency features by network 2. The features extracted after the output of each layer of the encoder of network 2 need to be combined with the signal features extracted by network 1, and the combined features include half of the radio frequency features and half of the signal features. Then, the combined features are sent to the next-layer encoder of network 2 to further extract more detailed radio frequency features. Therefore, the radio frequency features can be more conveniently and accurately extracted.

[0059] 6) Extracting radio frequency features from the to-be-tested data

[0060] After the training of the radio frequency features by network 2 in combination with network 1 is completed and the dual network model is saved, the dual network model is called to input the RF signal data whose radio frequency features are to be extracted into the dual network model and extract the radio frequency features thereof. The signal features extracted by network 1 from the RF signal data whose radio frequency features are to be extracted are used as auxiliary and optimization for the extraction of the radio frequency features by network 2. In this way, the radio frequency features extracted by network 2 after the multi-layer encoding are more complete.

[0061] The following embodiments are given in combination with specific parameters:

[0062] In combination with the above description, Figure 1 The present application comprises the following steps:

[0063] Step 1, data acquisition: by using simulation platform and measured parameters and selecting different design parameters, two types of radar radiation source models are built, and under the same type of model, four radar radiation source individual models are built by fine-tuning the parameters of each module to add unintentional modulation information. By adjusting the working parameters of the signal source model of different modulation modes, intentional modulation information is added to generate RF signals. The model can flexibly adjust the working parameters of the radar and provide RF signals carrying radio frequency information.

[0064] Step 2, data set division: the radar radiation source individual model is used to simulate RF signals of three different modulation modes, namely continuous wave signal (CW), linear frequency modulation signal (LFM) and binary phase shift keying signal (BPSK). By setting the mixed ideal CW signal, LFM signal and BPSK signal as the training data set of network 1, i.e. the RF signal input of network 1 does not include radio frequency characteristics; and setting the mixed modeling CW signal, LFM signal and BPSK signal as the training data set of network 2.

[0065] Step 3, data preprocessing: since neural network does not support complex as input, the in-phase and quadrature components of the training data set of mixed ideal signal of network 1 need to be extracted before inputting into network 1, and the mixed modeling signal training data set of network 2 also needs to be processed as I / Q data before inputting into network.

[0066] Step 4, training of double network model based on stack autoencoder:

[0067] Figure 4 The model diagram of deep stack autoencoder is shown in the figure. The structure of stack autoencoder mainly consists of two parts: encoder and decoder. Figure 5 Taking a single-layer network as an example, the overall structure and process diagram of the double network model based on stack autoencoder is shown in the figure. First, the encoder and decoder are used to build network 1 model, and the ideal RF signal training set is input into network 1 after normalization to extract signal features through hierarchical pre-training and global training fine-tuning. The loss function of hierarchical pre-training of network 1 is the mean square error of the decoding output of each layer and the encoding input. The fine-tuning loss function of global training of network 1 is the mean square error of the time domain of the reconstructed signal of the final decoding layer output and the signal time domain of the original encoding input. After training, the network 1 model is saved.

[0068] The model of network 2 is built and trained again using the encoder and decoder, and the RF signal data set containing the modeling is normalized and input to network 2 to extract the RF features through hierarchical pre-training and global training fine-tuning. The loss function of the hierarchical pre-training of network 2 is set to the mean square error constraint in the time domain and the frequency domain. The time domain constraint is the mean square error between the time domain input data of each layer of encoding and the time domain reconstruction data of each layer of decoding output. The frequency domain constraint is the mean square error between the frequency domain input data obtained by Fourier transform of the time domain input data of each layer of encoding and the frequency domain reconstruction data obtained by Fourier transform of the time domain reconstruction data of each layer of decoding output. The loss function of the global training of network 2 is fine-tuned, which has two items. The first item is the mean square error between the time domain of the reconstructed signal output by the final decoding layer and the time domain of the original encoded input signal. The second item is the mean square error between the frequency domain data obtained by Fourier transform of the time domain of the reconstructed signal output by the final decoding layer and the original frequency domain input data obtained by Fourier transform of the time domain of the original encoded input signal.

[0069] Since the input modeling RF signal of network 2 contains both signal features and RF features, and in order to more accurately extract RF features, the signal features must be constrained, the modeling signal data set is input to the saved network 1 model for calling to form a double network training at the same time during hierarchical pre-training. Therefore, the hierarchical pre-training and global training of network 2 and network 1 are different. The difference lies in the need to call network 1 as a constraint to assist network 2 in extracting RF features.

[0070] The specific constraint operation is that the output of the modeling RF signal through the first layer encoder of network 2 needs to merge the signal features of the output of the modeling RF signal through the first layer encoder of network 1, and the result of splicing is sent to the first layer decoder of network 2. During the pre-training of the second layer encoder and decoder of network 2, the input of the encoder is the merged output obtained by the previous layer encoder. Since the first layer encoding output of network 1 is called in the merged output, it already contains half of the signal features, so the features obtained by compressing the merged output by the second layer encoder are closer to the RF features. Similarly, the input of the encoder during the second layer training of network 2 is the merged output of the first layer encoders of network 1 and network 2.

[0071] The model network parameters of the RF feature extraction based on the stack auto-encoder network are as follows:

[0072]

[0073] The model network 1 structure of the RF feature extraction based on the stack auto-encoder network is as follows:

[0074]

[0075] The model network 2 structure based on the stack auto-encoder network for radio frequency feature extraction is as follows:

[0076]

[0077] Step 5, stack auto-encoder based double network model test: the RF signal data to be extracted fingerprint features is input into the double network model built based on the stack auto-encoder. The saved double network model is called, and the RF signal data to be extracted fingerprint features is first passed through network 1. The different dimension features output by each layer of the encoder of network 1 are combined with the output of each layer of the encoder of network 2 and used as the input of the next layer of the encoder. Finally, the output of the third layer of the encoder of network 2 contains the required extracted radio frequency features.

[0078] After the radio frequency features are extracted, the features output by the last layer of the encoder of network 1 and network 2 can be combined and input into the decoder of network 2 for layer-by-layer decoding and recovery. The effect of the radio frequency feature extraction can be verified by comparing the time domain and frequency domain of the reconstructed double network with the time domain and frequency domain of the original signal. Figure 6 and Figure 7 are the time domain graphs of the original LFM signal and the recovered LFM signal respectively, Figure 8 and Figure 9 are the frequency domain graphs of the original LFM signal and the recovered LFM signal respectively. Figure 6 , 7 , 8 and 9 are the horizontal coordinates representing the number of points, and the vertical coordinates representing the normalized amplitude.

Claims

1. A method of extracting radio frequency features of a radar emitter, the method comprising: The method comprises the following steps: acquiring a radio frequency signal to be extracted and pre-processing the radio frequency signal, and inputting the pre-processed radio frequency signal into a network 1 and a network 2 trained in advance for feature extraction, wherein the network 1 and the network 2 are both stack auto-encoder networks and both comprise a plurality of layers of the same layer number of encoders and decoders, and the network 2 outputs extracted radio frequency features from a last layer of the encoders; the training process of the network 1 and the network 2 comprises the following steps: step 1, generating a modeling radio frequency signal containing radio frequency features by a radar radiation source model built through simulation as a training set of the network 2, and generating an ideal radio frequency signal not containing radio frequency features as a training set of the network 1, wherein the training set of the network 2 and the training set of the network 1 are of the same composition; step 2, inputting the training set of the network 1 into the network 1 after pre-processing, first performing layer-by-layer pre-training on a single layer of the encoders and decoders according to a set number of training times of each layer, then performing global training on the network 1 after the layer-by-layer pre-training is completed, and saving a network 1 model after a set number of global training times is reached; wherein a signal feature output by each layer of the encoders during the training process is used as an input of a next layer of the encoders, a loss function of the single layer of the encoders and decoders is used to initialize weights and bias parameters of each layer of the network during the layer-by-layer pre-training, and a global loss function is used to update the weights and the bias parameters of the network 1 during the global training; step 3, inputting the training set of the network 2 into the network 2 after pre-processing, first performing layer-by-layer pre-training on a single layer of the encoders and decoders according to a set number of training times of each layer, then performing global training on the network 2 after the layer-by-layer pre-training is completed, and saving a network 2 model after a set number of global training times is reached; wherein a radio frequency feature extracted by each layer of the encoders of the network 2 and a signal feature extracted by a corresponding layer of the encoders of the network 1 are combined as an input of a next layer of the encoders of the network 2 during the training process, a loss function of the single layer of the encoders and decoders is used to initialize weights and bias parameters of each layer of the network during the layer-by-layer pre-training, and a global loss function is used to update the weights and the bias parameters of the network 2 during the global training.

2. The method of claim 1, wherein: The modeling radio frequency signal comprises unintentional modulation information and intentional modulation information.

3. The method of claim 1, wherein: The pre-processing of the training set of the network 1 and the training set of the network 2 comprises extracting in-phase / quadrature data from the radio frequency signal, concatenating the extracted data, and then performing normalization processing.

4. The method of claim 1, wherein: The loss function of the single layer of the encoders and decoders of the network 1 and the global loss function are respectively: wherein Loss1 is a loss function of a single-layer codec, Loss 11 is a global loss function, x1 represents an input of each layer of the encoder, is a reconstructed signal obtained by output of each layer of the decoder; y1 represents an input signal of the network 1, is a reconstructed signal obtained by output of the last layer of the decoder.

5. The method of claim 1, wherein: The loss function of the single layer of the encoders and decoders of the network 2 and the global loss function are respectively: wherein Loss2 is a loss function of a single layer codec, Loss 22 is a global loss function, x2 represents an input of each layer of the encoder, is a reconstructed signal obtained by output of each layer of the decoder; represents a frequency domain input obtained by Fourier transform of the time domain input x2, represents time domain reconstructed data obtained by Fourier transform, y2 represents an input signal of the network 2, is a reconstructed signal obtained by output of the last layer of the decoder, represents a frequency domain input obtained by Fourier transform of the time domain input y2, represents time domain reconstructed data obtained by Fourier transform.

Images