Signal spectrum enhancement method and system based on deep learning

Abstract

The present invention provides a signal spectrum enhancement method and system based on deep learning, the method comprising: obtaining a signal spectrum to be processed; inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model; wherein the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated. The signal spectrum enhancement method and system based on deep learning provided by the present invention realize spectrum enhancement through a pre-constructed spectrum enhancement model. Since the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated, it can reduce or even eliminate the influence of spectrum leakage to a great extent, thereby making the frequency resolution of the signal spectrum higher. Compared with the traditional windowing operation, the spectrum leakage elimination effect is more ideal.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to a signal spectrum enhancement method and system based on deep learning. Background Art

[0002] In the process of signal processing, it is often necessary to perform spectrum analysis on the signal, that is, to use the time-frequency domain analysis algorithm to transform the time domain signal into the frequency domain for subsequent processing. Taking the commonly used Fourier transform algorithm as an example, when the time domain signal is transformed using the above algorithm, only time domain data of limited length can be transformed, so the original signal needs to be truncated, and the truncation operation will cause the signal spectrum to be distorted and produce spectrum leakage, which will affect the subsequent signal processing process.

[0003] At present, spectrum leakage is often reduced by windowing the truncated signal. After windowing, the side lobes in the obtained spectrum are greatly reduced and the leakage phenomenon is alleviated, but at the cost of widening the main lobe. When there are multiple frequency components in the signal, the widening of the main lobe will cause adjacent frequency components to interfere with each other, and even cause weaker frequency components to be submerged in stronger frequency components.

[0004] It can be seen that the existing windowing operation cannot effectively eliminate the spectrum leakage phenomenon, and the frequency resolution is low. Summary of the invention

[0005] The present invention provides a signal spectrum enhancement method and system based on deep learning, which are used to solve the defects in the prior art that the windowing operation cannot effectively eliminate the spectrum leakage phenomenon and the frequency resolution is low.

[0006] In a first aspect, the present invention provides a signal spectrum enhancement method based on deep learning, the method comprising:

[0007] Obtaining the spectrum of the signal to be processed;

[0008] Inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model;

[0009] The spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated.

[0010] According to a signal spectrum enhancement method based on deep learning provided by the present invention, the training process of the spectrum enhancement model includes:

[0011] Building deep learning neural networks for spectrum enhancement;

[0012] Obtaining spectrum sample data with leakage eliminated and spectrum sample data with leakage, and constructing a training sample data set;

[0013] The deep learning neural network is trained using the training sample data set to obtain a spectrum enhancement model.

[0014] According to a signal spectrum enhancement method based on deep learning provided by the present invention, the deep learning neural network includes a plurality of cascaded complex domain fully connected layers, each of the complex domain fully connected layers includes a plurality of parallel complex domain neural operators, and the output of each of the complex domain fully connected layers is the output combination result of each complex domain neural operator therein.

[0015] According to a signal spectrum enhancement method based on deep learning provided by the present invention, the input of the complex domain neural operator is:

[0016] r in =x+j*y

[0017] The output of the complex domain neural operator is:

[0018] r out =x′+j*y′

[0019] Among them, x, x' are the real parts of input and output respectively, y, y' are the imaginary parts of input and output respectively, r in is a complex vector, r out is a plural number;

[0020] The relationship between the output and input of the complex domain neural operator is:

[0021]

[0022] Among them, w1, w2 are the parameters of the complex domain neural operator, b1, b2 are the biases of the complex domain neural operator, Σ means summing all elements in the vector, and σ(·) represents the activation function.

[0023] According to a signal spectrum enhancement method based on deep learning provided by the present invention, spectrum sample data with leakage eliminated and spectrum sample data with leakage are obtained to construct a training sample data set, including:

[0024] Generate an ideal spectrum through a random number generator to obtain spectrum sample data with leakage eliminated;

[0025] performing a leakage operation on the spectrum sample data from which leakage is eliminated to obtain spectrum sample data with leakage;

[0026] A training sample data set is constructed based on the leakage-eliminating spectrum sample data and the corresponding leakage spectrum sample data.

[0027] According to a signal spectrum enhancement method based on deep learning provided by the present invention, after performing a leakage operation on the spectrum sample data with leakage eliminated to obtain the spectrum sample data with leakage, the method further includes:

[0028] Gaussian white noise is added to the spectrum sample data with leakage to obtain an observed spectrum.

[0029] In a second aspect, the present invention further provides a signal spectrum enhancement system based on deep learning, the system comprising:

[0030] An acquisition module, used for acquiring the spectrum of the signal to be processed;

[0031] A processing module, used for inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model;

[0032] The spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated.

[0033] In a third aspect, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned signal spectrum enhancement methods based on deep learning are implemented.

[0034] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described deep learning-based signal spectrum enhancement methods.

[0035] In a fifth aspect, the present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned deep learning-based signal spectrum enhancement methods.

[0036] The signal spectrum enhancement method and system based on deep learning provided by the present invention realize spectrum enhancement through a pre-built spectrum enhancement model. Since the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated, it can greatly reduce or even eliminate the influence of spectrum leakage, thereby making the frequency resolution of the signal spectrum higher. Compared with the traditional windowing operation, the spectrum leakage elimination effect is more ideal. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 It is the time domain waveform of the truncated single-frequency signal;

[0039] Figure 2 It is the frequency domain waveform of the truncated single frequency signal;

[0040] Figure 3 This is the time domain waveform of the truncated signal after using the Gaussian window;

[0041] Figure 4 This is the frequency domain waveform of the truncated signal after using the Gaussian window;

[0042] Figure 5 is the ideal spectrum of a signal with two frequency components;

[0043] Figure 6 is the true spectrum of a signal with two frequency components;

[0044] Figure 7 It is a schematic diagram of the implementation flow of the signal spectrum enhancement method based on deep learning provided by the present invention;

[0045] Figure 8 It is a schematic diagram of the structure of the spectrum enhancement network;

[0046] Fig. 9 It is a schematic diagram of the data calculation principle of the complex neural operator;

[0047] Fig.10 It is a schematic diagram of the training principle of the spectrum enhancement network;

[0048] Fig.11 It is a schematic diagram of the spectrum waveform after being enhanced using the spectrum enhancement model;

[0049] Fig.12 It is a structural schematic diagram of a signal spectrum enhancement system based on deep learning provided by the present invention;

[0050] Fig.13 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] First, taking the commonly used Fourier transform as an example, the process of converting time domain signals into frequency domain is explained in detail.

[0053] If f(t) represents the time domain signal to be processed, then the Fourier transform of the time domain signal can be expressed as:

[0054]

[0055] Among them, t represents time, f represents frequency, i represents imaginary unit, π represents pi, and F(f) is the frequency spectrum of the time domain signal f(t).

[0056] However, the Fourier transform algorithm needs to be operated within the time range of -∞ to +∞, which is difficult to implement. Therefore, when the time domain signal is transformed using the algorithm, only time domain data of limited length can be transformed, so the original signal needs to be truncated.

[0057] For example, if a single-frequency signal of 20 Hz is selected, the single-frequency signal can be expressed as:

[0058] f(t)=e -i*2*π*20*t (2)

[0059] The single-frequency signal changes periodically from -∞ to +∞. Substituting f(t) into the above Fourier transform formula for Fourier transform, the spectrum of the obtained single-frequency signal is as follows:

[0060] F(f)=δ(f-20) (3)

[0061] Among them, δ(x) represents the impulse function, that is:

[0062]

[0063] That is, F(f) only has an impulse at 20Hz, and the value at other positions is 0. However, when a segment of a single-frequency signal is selected for Fourier analysis, the resulting spectrum is no longer an impulse function. For example, if a 250-ms signal is selected, Right now:

[0064]

[0065] Among them, T0=250ms. Figure 1 Shown The signal Substituting the expression into the above Fourier transform formula, we can get:

[0066]

[0067] in, for spectrum. Figure 2 Shown From the waveform, it can be seen that in addition to the component at 20Hz, the spectrum also has leakage components at other positions, generating more side lobes.

[0068] The ideal state should be The spectrum of f(t) is the same as the spectrum of f(t), so the spectrum of the latter can be obtained by Fourier transforming the former, but this is obviously impossible. This spectrum leakage phenomenon will cause the spectrum of the signal to be distorted, thus affecting the subsequent signal processing process.

[0069] In order to reduce spectrum leakage, the truncated signal is often windowed to further process the spectrum. The signal after windowing can be expressed as:

[0070]

[0071] Where wind(t) represents the window function used. Commonly used window functions include Gaussian, Hamming, Hanning, etc. Figure 3 and Figure 4 Shown The time domain and frequency domain waveforms of the signal after the Gaussian window function is added show that the side lobes are greatly reduced and the leakage phenomenon is alleviated, but at the cost of widening the main lobe. When there are multiple frequency components in the signal, this widening of the main lobe will cause adjacent frequency components to interfere with each other, and even cause weaker frequency components to be submerged in stronger frequency components. Figure 5 It shows the spectrum calculated by Fourier transform theory when there are two frequency components of 15Hz and 20Hz in the signal. Figure 6 The actual spectrum obtained by Fourier transform after the windowing operation is shown. It can be seen that the leakage of the spectrum on the main lobe causes serious interference with adjacent components.

[0072] In order to obtain a better spectrum leakage elimination effect, the present invention uses deep learning technology as a carrier, and trains a deep learning neural network to explore the influence of the time-frequency domain analysis algorithm on the signal spectrum leakage, thereby enhancing the signal spectrum obtained by the frequency domain analysis algorithm, reducing or even eliminating the spectrum leakage phenomenon.

[0073] Figure 7 The present invention shows a signal spectrum enhancement method based on deep learning, which includes:

[0074] S710: Obtain a spectrum of a signal to be processed;

[0075] S720: Inputting the signal spectrum to be processed into the spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model;

[0076] Among them, the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated.

[0077] Specifically, in this embodiment, the process of obtaining spectrum sample data with leakage eliminated and spectrum sample data with leakage and constructing a training sample data set includes:

[0078] First, a deep learning neural network for spectrum enhancement is constructed; the neural network is also called a Spectrogram Enhancement Network (SEN). The input of the network is a spectrum with leakage obtained by traditional Fourier transform, and the output is a spectrum with leakage eliminated.

[0079] The structure of the deep learning neural network (i.e., spectrum enhancement network) used for spectrum enhancement in this embodiment is as follows: Figure 8 As shown, the deep learning neural network includes four cascaded units, each unit is a complex domain fully connected layer (i.e., Complex FC) activated by an activation function (tanh function is used in this embodiment), each complex domain fully connected layer includes several parallel complex domain neural operators, and the data calculation principle of each complex domain neural operator is as follows Fig. 9 shown.

[0080] If the input of the i-th complex domain neural operator of Complex FC is recorded as:

[0081] r in =x+j*y (8)

[0082] The output is recorded as:

[0083] r out =x′+j*y′ (9)

[0084] Among them, x, x' are the real parts of the two, y, y' are the imaginary parts of the two, and r in is a complex vector, r out Is a plural number.

[0085] Then the relationship between the output and input of the complex domain neural operator is:

[0086]

[0087] Among them, w1 and w2 are parameters of the complex domain neural operator, b1 and b2 are biases of the complex domain neural operator, Σ represents the sum of all elements in the vector, σ(·) represents the activation function, and the activation function in this embodiment is the tanh function.

[0088] The expression of the tanh function is:

[0089]

[0090] The complex arrays output by all complex-domain neural operators of each complex-domain fully-connected layer are synthesized into a complex vector, which serves as the overall output of the complex-domain fully-connected layer.

[0091] In this embodiment, the four Complex FCs in the spectrum enhancement network each contain 128, 256, 256, and N complex domain neural operators, respectively, where N is the number of sampling points of the spectrum input to the network, so as to ensure that the number of sampling points of the spectrum output by the network is the same as that of the input spectrum.

[0092] It should be noted that, since the spectrum of the signal to be processed can be represented by an expression containing complex terms, such as the spectrum expression obtained by Fourier transform, i.e., formula (1), the spectrum enhancement network provided in this embodiment mainly obtains new complex terms by performing a series of operations on the complex terms in the spectrum expression, i.e., the operation process of the above-mentioned complex domain neural operator, thereby obtaining a new spectrum expression, i.e., the final enhanced spectrum.

[0093] From the perspective of a neural network, this embodiment uses a spectrum enhancement network to extract hidden layer features from the original spectrum. The hidden layer features are used to characterize the structural characteristics of the spectrum elements (i.e., the spectrum intensity distribution at each sampling time in the spectrum), and then an enhanced spectrum is constructed based on the hidden layer features.

[0094] Then, spectrum sample data with leakage eliminated and spectrum sample data with leakage are obtained to construct a training sample data set.

[0095] In order to train the above spectrum enhancement network, a large amount of sample data is required. In this embodiment, the network training is performed by artificially generating sample data. The process of data generation and model training is as follows: Fig.10 shown.

[0096] The following is a detailed description of the construction process of the training sample data set:

[0097] Step 1: Generate an ideal spectrum through a random number generator, that is, eliminate the leaked spectrum sample data; the expression of the ideal spectrum in this embodiment is:

[0098]

[0099] Among them, n is a random integer representing the number of frequency components. is a random real number, f i represents the frequency of the i-th frequency component, represents the initial phase of the ith frequency component. In this embodiment, the ideal spectrum ψ(f) has impulse responses only at a finite number of frequency components, and the values ​​at other positions are all 0.

[0100] Step 2: Perform a leakage operation on the spectrum sample data with leakage eliminated to obtain spectrum sample data with leakage.

[0101] Considering the properties of Fourier transform, time domain windowing is equivalent to convolution in the frequency domain. Therefore, the spectrum after windowing is equal to the circular convolution of the spectrum before windowing and the spectrum of the window function, that is:

[0102]

[0103] in, WIND (f) is the Fourier transform of the window function wind(t). In this embodiment, the window function can be selected from commonly used window functions, such as Gaussian, Hamming, Hanning, etc. is the circular convolution of two functions, namely:

[0104]

[0105] At this point, a training sample data set can be constructed based on the spectrum sample data without leakage and the corresponding spectrum sample data with leakage. However, considering the authenticity of the sample data, in order to simulate the noise of the signal collected in the real environment, this embodiment adds a third step.

[0106] Step 3: Add Gaussian white noise to the spectrum sample data with leakage to obtain the observed spectrum. The expression of the observed spectrum in this embodiment is as follows:

[0107] ψ measured (f) = ψ leaked (f)+N(0,1) (15)

[0108] Where N(0,1) is Gaussian white noise with a mean of 0 and a standard deviation of 1.

[0109] In this embodiment, each observed spectrum is combined with its corresponding ideal spectrum into a data pair, and the data pairs then constitute a training sample data set.

[0110] Finally, the deep learning neural network is trained using the training sample data set to obtain the spectrum enhancement model.

[0111] The observed spectrum ψ measured (f) is used as the input of the spectrum enhancement network, and the ideal spectrum ψ(f) is used as the expected output of the spectrum enhancement network. The Back Propagation algorithm is used for training to obtain the spectrum enhancement model.

[0112] In the model training process, in order to determine whether the trained model has reached the optimal state, this embodiment uses the spectrum enhancement network to measured (f) Enhanced spectrum of actual output The spectrum enhancement network is optimized by comparing the enhanced spectrum with the ideal spectrum ψ(f) and analyzing the amplitude error between the enhanced spectrum and the ideal spectrum through the MSE loss function.

[0113] This embodiment uses 5000 epochs to train the spectrum enhancement network. Each epoch generates 100 batches of training data. Each batch contains 128 samples. Each sample is a data pair (ψ(f), ψ measured (f)).

[0114] After obtaining the spectrum enhancement model, the spectrum obtained by Fourier transform can be directly used as input data and input into the spectrum enhancement model. The output of the spectrum enhancement model is the spectrum with reduced leakage effect, that is, the enhanced signal spectrum. Fig.11 Shows the Figure 6 The spectrum waveform shown is the spectrum waveform after being enhanced using the spectrum enhancement model. It can be seen that the leakage effect is suppressed to a great extent.

[0115] The signal spectrum enhancement system based on deep learning provided by the present invention is described below. The signal spectrum enhancement system based on deep learning described below and the signal spectrum enhancement method based on deep learning described above can refer to each other.

[0116] Fig.12 A signal spectrum enhancement system based on deep learning provided by an embodiment of the present invention is shown, and the system includes:

[0117] An acquisition module 121 is used to acquire a spectrum of a signal to be processed;

[0118] The processing module 122 is used to input the signal spectrum to be processed into the spectrum enhancement model to obtain the enhanced signal spectrum output by the spectrum enhancement model;

[0119] Among them, the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated.

[0120] In this embodiment, the processing module 122 processes the signal spectrum with leakage based on the spectrum enhancement model to obtain an enhanced signal spectrum, that is, a spectrum with reduced leakage. Specifically, the training process of the spectrum enhancement model includes:

[0121] Firstly, a deep learning neural network for spectrum enhancement, namely spectrum enhancement network, is constructed.

[0122] Then, spectrum sample data with leakage eliminated and spectrum sample data with leakage are obtained to construct a training sample data set;

[0123] Finally, the deep learning neural network is trained using the training sample data set to obtain the spectrum enhancement model.

[0124] In this embodiment, the deep learning neural network includes multiple cascaded complex domain fully connected layers, each complex domain fully connected layer includes several parallel complex domain neural operators, and the output of each complex domain fully connected layer is the output combination result of each complex domain neural operator therein.

[0125] It should be noted that the sample data used to train the network in this embodiment is obtained in an artificially generated manner. First, the ideal spectrum is obtained through a random number generator, and then a leakage operation is performed on the ideal spectrum to obtain a leakage spectrum. Next, Gaussian white noise is added to the leakage spectrum to obtain an observed spectrum. The observed spectrum is used as the input of the network, and the ideal spectrum is used as the output of the network. The deep learning neural network is trained to obtain the final spectrum enhancement model.

[0126] After the spectrum enhancement model is constructed, the spectrum obtained by Fourier transform can be directly enhanced using the spectrum enhancement model to obtain a corresponding spectrum with reduced leakage.

[0127] Of course, the deep learning-based signal spectrum enhancement system provided in this embodiment is not only suitable for signal spectra with leakage obtained by Fourier transform, but also suitable for signal spectra with leakage obtained by other time-frequency domain analysis algorithms, such as spectra obtained by wavelet transform. The process and principle of spectrum enhancement processing are the same.

[0128] It can be seen that the signal spectrum enhancement system based on deep learning provided by the embodiment of the present invention realizes spectrum enhancement by using a pre-built spectrum enhancement model through a processing module. Since the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated, it can greatly reduce or even eliminate the influence of spectrum leakage, thereby making the frequency resolution of the signal spectrum higher. Compared with traditional windowing operations, the system has a better effect in eliminating spectrum leakage.

[0129] Fig.13 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.13 As shown, the electronic device may include: a processor 131, a communication interface 132, a memory 133 and a communication bus 134, wherein the processor 131, the communication interface 132 and the memory 133 communicate with each other through the communication bus 134. The processor 131 may call the logic instructions in the memory 133 to execute a signal spectrum enhancement method based on deep learning, the method comprising: obtaining a signal spectrum to be processed; inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model; wherein the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data without leakage.

[0130] In addition, the logic instructions in the above-mentioned memory 133 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0131] On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the signal spectrum enhancement method based on deep learning provided by the above methods, and the method includes: obtaining a signal spectrum to be processed; inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model; wherein the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated.

[0132] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the signal spectrum enhancement method based on deep learning provided by the above-mentioned methods, the method comprising: obtaining a signal spectrum to be processed; inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model; wherein the spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data with leakage eliminated.

[0133] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0134] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A signal spectrum enhancement method based on deep learning, characterized in that: include: Obtaining the spectrum of the signal to be processed; Inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model; The spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data without leakage; The training process of the spectrum enhancement model includes: Building deep learning neural networks for spectrum enhancement; Obtaining spectrum sample data with leakage eliminated and spectrum sample data with leakage, and constructing a training sample data set; The deep learning neural network is trained by using the training sample data set to obtain a spectrum enhancement model; The deep learning neural network includes a plurality of cascaded complex domain fully connected layers, each of the complex domain fully connected layers includes a plurality of parallel complex domain neural operators, and the output of each of the complex domain fully connected layers is a combination result of the outputs of the complex domain neural operators therein; The input of the complex domain neural operator is: r in =x+j*y The output of the complex domain neural operator is: r out =x ′ +j*y ′ Among them, x,x ′ are the real parts of the input and output, y,y ′ are the imaginary parts of input and output respectively, r in is a complex vector, r out is a plural number; The relationship between the output and input of the complex domain neural operator is: Among them, w1, w2 are the parameters of the complex domain neural operator, b1, b2 are the biases of the complex domain neural operator, Σ means summing all elements in the vector, and σ(·) represents the activation function.

2. The signal spectrum enhancement method based on deep learning according to claim 1, characterized in that: Obtain the spectrum sample data with leakage eliminated and the spectrum sample data with leakage, and construct a training sample data set, including: Generate an ideal spectrum through a random number generator to obtain spectrum sample data with leakage eliminated; performing a leakage operation on the spectrum sample data from which leakage is eliminated to obtain spectrum sample data with leakage; A training sample data set is constructed based on the leakage-eliminating spectrum sample data and the corresponding leakage spectrum sample data.

3. The signal spectrum enhancement method based on deep learning according to claim 2, characterized in that: After performing a leakage operation on the spectrum sample data with leakage eliminated to obtain the spectrum sample data with leakage, the method further includes: Gaussian white noise is added to the spectrum sample data with leakage to obtain an observed spectrum.

4. A signal spectrum enhancement system based on deep learning, characterized in that: include: An acquisition module, used for acquiring the spectrum of the signal to be processed; A processing module, used for inputting the signal spectrum to be processed into a spectrum enhancement model to obtain an enhanced signal spectrum output by the spectrum enhancement model; The spectrum enhancement model is obtained by training a deep learning neural network based on spectrum sample data with leakage and spectrum sample data without leakage; The training process of the spectrum enhancement model includes: Building deep learning neural networks for spectrum enhancement; Obtaining spectrum sample data with leakage eliminated and spectrum sample data with leakage, and constructing a training sample data set; The deep learning neural network is trained by using the training sample data set to obtain a spectrum enhancement model; The deep learning neural network includes a plurality of cascaded complex domain fully connected layers, each of the complex domain fully connected layers includes a plurality of parallel complex domain neural operators, and the output of each of the complex domain fully connected layers is a combination result of the outputs of the complex domain neural operators therein; The input of the complex domain neural operator is: r in =x+j*y The output of the complex domain neural operator is: r out =x ′ +j*y ′ Among them, x,x ′ are the real parts of the input and output, y,y ′ are the imaginary parts of input and output respectively, r in is a complex vector, r out is a plural number; The relationship between the output and input of the complex domain neural operator is: Among them, w1, w2 are the parameters of the complex domain neural operator, b1, b2 are the biases of the complex domain neural operator, Σ means summing all elements in the vector, and σ(·) represents the activation function.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the signal spectrum enhancement method based on deep learning as described in any one of claims 1 to 3 are implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the signal spectrum enhancement method based on deep learning as described in any one of claims 1 to 3 are implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the signal spectrum enhancement method based on deep learning as described in any one of claims 1 to 3 are implemented.

Images