CN113111720A - Electromagnetic modulation signal denoising method and system based on deep learning - Google Patents
Electromagnetic modulation signal denoising method and system based on deep learning Download PDFInfo
- Publication number
- CN113111720A CN113111720A CN202110284247.0A CN202110284247A CN113111720A CN 113111720 A CN113111720 A CN 113111720A CN 202110284247 A CN202110284247 A CN 202110284247A CN 113111720 A CN113111720 A CN 113111720A
- Authority
- CN
- China
- Prior art keywords
- signal
- noise
- data
- generator
- signals
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 230000006870 function Effects 0.000 claims abstract description 52
- 238000012549 training Methods 0.000 claims abstract description 36
- 238000012545 processing Methods 0.000 claims abstract description 18
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 230000003042 antagnostic effect Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000013527 convolutional neural network Methods 0.000 claims description 4
- 238000009826 distribution Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 238000001914 filtration Methods 0.000 abstract description 9
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Signal Processing (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Filters That Use Time-Delay Elements (AREA)
Abstract
A method and a system for denoising electromagnetic modulation signals based on deep learning comprise the following steps: s1, making a target signal and noise signal data set; s2, a filtering denoising method and a signal enhancement method are used, and a data set is expanded and optimized; s3, respectively defining model structures and loss functions of the generator and the discriminator, and training the model until the model is stable; and S4, outputting a denoising result. The generator loss function mentioned in the method combines and uses the output loss of the discriminator, the minimum absolute value deviation and the continuity difference value, and particularly considers the continuity characteristic of the generated denoising signal. The invention also comprises an electromagnetic modulation signal denoising system based on deep learning, which consists of a data processing module, a training module and an output module which are connected in sequence. The invention combines a filtering denoising method not based on learning and a denoising method based on learning, can adaptively learn the characteristics of signals, realizes signal denoising, and has better universality on signal denoising.
Description
Technical Field
The invention relates to the field of unsupervised learning of electromagnetic signals, in particular to an electromagnetic modulation signal denoising method and system based on deep learning.
Background
Electromagnetic signals refer to electromagnetic waves, such as radio signals, that propagate in free space. In signal processing, the presence of noise is a common problem for any signal type. The neural network is a learning-based signal denoising method, and does not need precise modeling of signals and noises and optimal parameter adjustment. This approach is particularly popular in the field of learning-based images. Likewise, deep neural networks have achieved good results in terms of audio and speech processing. In contrast, learning-based methods have a rather limited impact on low dimensional signals (e.g. modulation signals). For low-dimensional signals, non-learning methods such as filtering, wavelet transformation and empirical mode decomposition are mainly used in the previous denoising work. Linear denoising methods rely on filtering, but they show limitations when the signal and noise share the same spectrum. In wavelet transforms, their performance depends on the selection of predefined basis functions, which may not reflect the properties of the signal. Empirical mode decomposition is a data-driven method, but it is difficult to decompose the signal into distinct frequency components.
There are papers: the paper "adaptive Signal Denoising with Encoder-Decoder Networks", accepted in EUSIPCO in 2020 (28 th european Signal processing conference in 2020), proposes an Encoder-Decoder architecture to denoise a Signal, considering the Denoising task as a distributed alignment between clean and noisy signals, the Signal being represented by a series of measurements. The goal is to align the clean and noisy signal potential representations for a given two signals by the encoder. However, the method does not study the modulation signal, and does not consider the combination of a filtering and denoising method which is not based on learning and a method based on learning, so that the signal denoising effect is improved.
There are patents: the technical scheme disclosed in the patent with the application number of CN201710614277.7 is based on an object detection method and device for generating a countermeasure network, in which the method performs size transformation on an input original image to obtain a first image, and performs filtering and denoising processing on the first image to obtain a second image; inputting the second image into a production countermeasure network. The method aims at object recognition in the image field, improves the object recognition rate by combining filtering denoising and generation countermeasure network on the image, does not relate to the signal field, and uses a specific network structure and a loss function which are different from those of the method.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, and provides an electromagnetic modulation signal denoising method and system based on deep learning, which can denoise electromagnetic modulation signals and improve the signal-to-noise ratio of the signals.
In order to achieve the purpose, the invention provides the following scheme:
the invention provides an electromagnetic modulation signal denoising method and system based on deep learning, wherein the method comprises the following steps:
s1, taking the 18dB 9 modulation type data in the public data set as target signals, and adding quantitative Gaussian noise into the target signals to obtain corresponding 12dB noise signals;
s2, inputting the noise signal in S1 into a low-pass filter, outputting the noise signal as a filtered signal, correspondingly forming a data pair by the target signal in S1 and the newly obtained filtered signal, and then expanding the data volume by ten times by using a signal enhancement method;
s3, constructing and generating a confrontation network model, respectively defining loss functions of a generator and a discriminator, taking the data pairs in S2 as the input of the discriminator, taking the filtered signals as the input of the generator, and minimizing the difference between the generated signals and the target signals through confrontation training so as to obtain the generator for realizing signal denoising;
and S4, extracting the trained generator, inputting the filtered signal, and outputting a de-noising signal.
Further, the S1 specifically includes:
s1.1, adopting a public data set RML2016.10a, extracting 8910 data with the signal-to-noise ratio of 18dB and the modulation types of 8PSK, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM as target signals;
and S1.2, adding quantitative Gaussian noise in the extracted target signal data to obtain corresponding 12dB noise signal data.
Further, the S2 specifically includes:
s2.1, setting a low-pass filter to be 8-order, setting the parameter to be 0.2, inputting the noise signal obtained in the S1.2 into the low-pass filter, and outputting the noise signal which is the filtered signal;
s2.2, loading the filtered signals one by one, and then respectively rotating by 90 degrees, rotating by 180 degrees, rotating by 270 degrees, turning over up and down, turning over left and right, adding weak random noise parameters of (0,0.0005), adding weak random noise parameters of (0,0.0010), adding weak random noise parameters of (0,0.0015) and adding weak random noise parameters of (0,0.0020) to obtain new signals, so that the data volume is enlarged by ten times;
and S2.3, expanding the data volume of the target signal by ten times by using a copy function, splicing all signals mentioned in the S2.2, and forming a data pair by the target signal and the filtered signals one by using an enumeration function, wherein the data structure is (None, 4, 128).
Further, the S3 specifically includes:
s3.1, defining a generator model;
s3.2, defining a loss function of the generator model, wherein the loss function of the generator comprises the output loss of the discriminator, the minimum absolute value deviation and a continuity difference value, the continuity difference value (Gtv _ loss) is the average value of the p-th power of the absolute value of the difference value of the first 127 data minus the second 127 data of the I-path signal (Q-path signal) generated by the generator, the continuity degree of the points generating the signals is reflected, and the lower the numerical value is, the better the continuity is;
the loss function (G _ loss) of the generator combines three loss functions:
s3.3, defining a discriminator model;
s3.4, defining a loss function of the discriminator model;
s3.5, defining a model optimizer, and performing model optimization by adopting an Adam optimizer and an attenuation learning rate;
s3.6, performing countermeasure training, inputting corresponding data pairs into a discriminator when the discriminator discriminates for the first time, expecting the discriminator to discriminate the data pairs of the target signal and the filtered signal as true, and discriminating the data pairs of the de-noised signal and the filtered signal generated by the generator as false; inputting the filtered signal data into a generator to generate a new de-noising signal, and carrying out second-time judger judgment, wherein the new de-noising signal and the filtered signal data are input into a judger to be expected to be judged to be true; repeating the steps to enable the data distribution of the de-noising signal and the target signal to be more and more like; and finally, finishing training when the true and false discrimination probabilities of the discriminator on the data pairs are the same.
Further, the generator loss specifically includes: the discriminator outputs loss, minimum absolute value deviation and continuity difference.
Further, the S4 specifically includes:
s4.1, taking out the generator network model defined in S3.1, and loading generator model parameters stored by the training module into the generator network model to obtain a complete generator model for denoising;
and S4.2, inputting the noise signal or the filtered signal into a generator model, and outputting a denoising signal.
A system for electromagnetic modulation signal denoising based on deep learning, comprising: the device comprises a data processing module, a training module and an output module;
the data processing module takes the 9 modulation type data of 18dB in the public data set as target signals, then adds quantitative Gaussian noise into the target signals to obtain corresponding 12dB noise signals, inputs the noise signals into a low-pass filter, outputs the noise signals into filtered signals, correspondingly forms data pairs by the target signals and the newly obtained filtered signals, and then enlarges the data volume by ten times by using a signal enhancement method;
the training module inputs data in the data processing module into a training module, trains the antagonistic network model, iterates for multiple times until the generated antagonistic network model is stably trained, and stores the model;
the output module loads a generator model in the generation confrontation network model stored by the training module, inputs a noise signal to be tested or a filtered signal into the generator model obtained by the training module and outputs a denoising signal;
the data processing module, the training module and the output module are sequentially linked.
The technical idea of the invention is as follows: the invention discloses a deep learning signal denoising method integrating filtering denoising and generation of a countermeasure network. Secondly, the low-pass filter is used for filtering the noise signal, and a foundation is laid for subsequent deep learning denoising. Moreover, the additionally added continuity loss function is utilized to improve the connectivity of the signal generated by the generator, so that the original connection characteristics of the signal can be better reserved when the generation countermeasure network learns the characteristics of the signal. In addition, different training parameters such as different learning rates and different weight attenuations can be set, and the feasibility and the effectiveness of the method are proved by experimental results.
The invention has the beneficial effects that:
1. the method does not need to use various priori knowledge of signal data, and has better universality in signal denoising.
2. The method comprises the steps of firstly, using a filter to carry out preliminary denoising, laying a foundation for subsequent deep denoising, and simultaneously expanding a data set by using a signal enhancement method so that a generated confrontation network model can be fully trained; the generation countermeasure network uses a self-encoder structure, increasing the model learning ability.
3. The generator loss function consists of three parts: the absolute difference value of the filtered signal and the target signal, the continuity loss value of the generated signal and the discrimination similarity value of the discriminator respectively correspond to the numerical optimization in the denoising process, the data continuity optimization of the generated signal and whether the generated signal integrally accords with the characteristics of the target signal.
Drawings
FIG. 1 is a flow chart of a method and system for implementing the present invention;
FIG. 2 is a data processing flow diagram;
fig. 3 is a diagram of a generation countermeasure network model architecture.
Detailed Description
The following detailed description of embodiments of the invention is provided in connection with the accompanying drawings.
Referring to fig. 1 to 3, a method and a system for denoising an electromagnetic modulation signal based on deep learning include the following steps:
s1, using the 18dB 9 modulation type data in the public data set as target signals, adding quantitative Gaussian noise into the target signals to obtain corresponding 12dB noise signals, the concrete steps are as follows:
s1.1, adopting a public data set RML2016.10a, extracting 8910 data with the signal-to-noise ratio of 18dB and the modulation types of 8PSK, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM as target signals.
And S1.2, adding quantitative Gaussian noise in the extracted target signal data to obtain corresponding 12dB noise signal data. The noise signal data with the low signal-to-noise ratio of 12dB can be obtained by inputting the target signal data and the low signal-to-noise ratio (SNR) obtained after expected Gaussian noise is added and calculating through the following formula.
Nsignal=Anoise+Asignal (5)
AnoiseTo normalize noise, RnoiseAs random noise, AsignalIs 18db signal amplitude, PsignalIs the signal frequency, n is the signal length, VnoiseAs variance of noise, σnoiseIs the standard deviation of noise, NsignalIs a noise signal.
S2, inputting the noise signal in S1 into a low-pass filter, outputting the noise signal as a filtered signal, correspondingly forming a data pair by the target signal in S1 and the newly obtained filtered signal, and then expanding the data volume by ten times by using a signal enhancement method, wherein the specific steps are as follows:
s2.1, setting the low-pass filter to be 8-order, setting the parameter to be 0.2, and inputting the low-pass filter into the noise signal obtained in the S1.2, wherein the output is the filtered signal.
S2.2, loading the filtered signals one by one, and then respectively obtaining new signals by rotating 90 degrees, rotating 180 degrees, rotating 270 degrees, turning up and down, turning left and right, adding weak random noise parameters of (0,0.0005), adding weak random noise parameters of (0,0.0010), adding weak random noise parameters of (0,0.0015) and adding weak random noise parameters of (0,0.0020), so that the data volume is enlarged by ten times.
And S2.3, expanding the data volume of the target signal by ten times by using a copy function, splicing all signals mentioned in the S2.2, and forming a data pair by the target signal and the filtered signals one by using an enumeration function, wherein the data structure is (None, 4, 128).
S3, constructing and generating a confrontation network model, respectively defining loss functions of a generator and a discriminator, taking the data pair in S2 as the input of the discriminator, taking the filtered signal as the input of the generator, and minimizing the difference between the generated signal and the target signal through confrontation training so as to obtain the generator for realizing signal denoising, wherein the specific steps are as follows:
and S3.1, defining a generator model. The encoder uses a 6-layer one-dimensional convolutional neural network and uses an activation function with parameters after each layer. And finally changing the input data structure from (None, 2, 128) to (None, 128, 2). The decoder uses 6 layers of one-dimensional inverse convolution neural network, and uses activation function with parameters after each layer; each layer of input is a combination of random noise (or the last layer of deconvolution output data) and the output data of the corresponding layer of the first half network. The final producer output data structure is still (None, 2, 128).
S3.2, defining a loss function of the generator model. The generator loss function includes the discriminator output loss, the minimum absolute value deviation, and the continuity difference. Discriminator output loss, i.e., discriminator output value (f (x)i) The sum of the squares of the differences from the target value (1) is minimized:
the minimum absolute value deviation is the de-noised signal data (G (x)) generated by the generatori') and target signal data (S (x)i) ) is minimized.
The continuity difference (Gtv _ loss) is the average of the p-th power of the absolute value of the difference between the first 127 data minus the second 127 data of the I-path signal (Q-path signal) generated by the generator, and represents the degree of continuity of the points at which the signal is generated, with lower values being better in continuity.
The loss function (G _ loss) of the generator combines the above three loss functions as:
and S3.3, defining a discriminator model, adopting a 6-layer one-dimensional convolutional neural network, using a normalization and activation function after each layer, and adding a Dropout layer after the 3 rd convolutional layer. And then connected with 1 layer of one-dimensional convolution neural network, 1 linear layer and a sigmoid layer. Finally, the data structure is changed from input (None, 4, 128) to output scalar.
And S3.4, defining a loss function of the discriminator model. The discriminator mainly uses the least square error, which is the output value (f (x) of the discriminatori) Is minimized from the squared sum of the difference from the target value (1/0).
When data pairs of a target signal and a noise signal are input, the target value 1 is true, and the loss function is:
when the data pair of the denoised signal and the noise signal is input, the target value 0 is false, and the loss function is:
and S3.5, defining a model optimizer, and performing model optimization by adopting an Adam optimizer and an attenuation learning rate.
And S3.6, performing countermeasure training, inputting corresponding data pairs into the discriminator when the discriminator performs discrimination for the first time, expecting the discriminator to discriminate the data pairs of the target signal and the filtered signal as true, and discriminating the data pairs of the de-noised signal and the filtered signal generated by the generator as false. And inputting the filtered signal data into a generator to generate a new de-noising signal, and carrying out second-time judger judgment, wherein the new de-noising signal and the filtered signal data are expected to be judged to be true when being input into the judger. The above steps are repeated, so that the data distribution of the de-noising signal and the target signal is more and more like. And finally, finishing training when the true and false discrimination probabilities of the discriminator on the data pairs are the same.
S4, extracting the trained generator, inputting the filtered signal or noise signal, and outputting a denoising signal, wherein the specific steps are as follows:
s4.1, taking out the generator network model defined in S3.1, and loading generator model parameters stored by the training module into the generator network model to obtain a complete generator model for denoising;
and S4.2, inputting the noise signal or the filtered signal into a generator model, and outputting a denoising signal.
A system for electromagnetic modulation signal denoising based on deep learning, comprising: the device comprises a data processing module, a training module and an output module;
the data processing module takes the 9 modulation type data of 18dB in the public data set as target signals, then adds quantitative Gaussian noise into the target signals to obtain corresponding 12dB noise signals, inputs the noise signals into a low-pass filter, outputs the noise signals into filtered signals, correspondingly forms data pairs by the target signals and the newly obtained filtered signals, and then enlarges the data volume by ten times by using a signal enhancement method; the method specifically comprises the following steps:
s1.1, adopting a public data set RML2016.10a, extracting 8910 data with the signal-to-noise ratio of 18dB and the modulation types of 8PSK, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM as target signals;
s1.2, adding quantitative Gaussian noise in the extracted target signal data to obtain corresponding 12dB noise signal data; inputting target signal data and low signal-to-noise ratio (SNR) obtained after expected Gaussian noise is added, and calculating by the following formula to obtain noise signal data with low SNR of 12 dB;
Nsignal=Anoise+Asignal (5)
Anoiseto normalize noise, RnoiseAs random noise, AsignalIs 18db signal amplitude, PsignalIs the signal frequency, n is the signal length, VnoiseAs variance of noise, σnoiseIs the standard deviation of noise, NsignalAs noise signals
S2.1, setting a low-pass filter to be 8-order, setting the parameter to be 0.2, inputting the noise signal obtained in the S1.2 into the low-pass filter, and outputting the noise signal which is the filtered signal;
s2.2, loading the filtered signals one by one, and then respectively rotating by 90 degrees, rotating by 180 degrees, rotating by 270 degrees, turning over up and down, turning over left and right, adding weak random noise parameters of (0,0.0005), adding weak random noise parameters of (0,0.0010), adding weak random noise parameters of (0,0.0015) and adding weak random noise parameters of (0,0.0020) to obtain new signals, so that the data volume is enlarged by ten times;
and S2.3, expanding the data volume of the target signal by ten times by using a copy function, splicing all signals mentioned in the S2.2, and forming a data pair by the target signal and the filtered signals one by using an enumeration function, wherein the data structure is (None, 4, 128).
The training module inputs data in the data processing module into a training module, trains the antagonistic network model, iterates for multiple times until the generated antagonistic network model is stably trained, and stores the model; the method specifically comprises the following steps:
s3.1, defining a generator model; the encoder uses a 6-layer one-dimensional convolutional neural network and uses an activation function with parameters after each layer. And finally changing the input data structure from (None, 2, 128) to (None, 128, 2). The decoder uses 6 layers of one-dimensional inverse convolution neural network, and uses activation function with parameters after each layer; each layer of input is a combination of random noise (or the last layer of deconvolution output data) and the output data of the corresponding layer of the first half network. The data structure output by the final generator is still (None, 2, 128);
s3.2, defining a loss function of the generator model, wherein the loss function of the generator comprises the output loss of the discriminator, the minimum absolute value deviation and a continuity difference value, the continuity difference value (Gtv _ loss) is the average value of the p-th power of the absolute value of the difference value of the first 127 data minus the second 127 data of the I-path signal (Q-path signal) generated by the generator, the continuity degree of the points generating the signals is reflected, and the lower the numerical value is, the better the continuity is;
the loss function (G _ loss) of the generator combines three loss functions:
s3.3, defining a discriminator model; a 6-layer one-dimensional convolution neural network is adopted, a normalization and activation function is used after each layer, and a Dropout layer is added after the 3 rd convolution layer; and then connected with 1 layer of one-dimensional convolution neural network, 1 linear layer and a sigmoid layer. Finally, the data structure is changed from input (None, 4, 128) to output scalar;
s3.4, defining a loss function of the discriminator model; the discriminator mainly uses the least square error to output the value (f (x) of the discriminatori) Minimization of the sum of squares of the differences from the target value (1/0);
when data pairs of a target signal and a noise signal are input, the target value 1 is true, and the loss function is:
when the data pair of the denoised signal and the noise signal is input, the target value 0 is false, and the loss function is:
s3.5, defining a model optimizer, and performing model optimization by adopting an Adam optimizer and an attenuation learning rate;
s3.6, performing countermeasure training, inputting corresponding data pairs into a discriminator when the discriminator discriminates for the first time, expecting the discriminator to discriminate the data pairs of the target signal and the filtered signal as true, and discriminating the data pairs of the de-noised signal and the filtered signal generated by the generator as false; inputting the filtered signal data into a generator to generate a new de-noising signal, and carrying out second-time judger judgment, wherein the new de-noising signal and the filtered signal data are input into a judger to be expected to be judged to be true; repeating the steps to enable the data distribution of the de-noising signal and the target signal to be more and more like; and finally, finishing training when the true and false discrimination probabilities of the discriminator on the data pairs are the same.
The output module loads generator model parameters stored by the training module and used for generating a confrontation network model, inputs noise signals or filtered signals to be tested into the generator model obtained by the training module and outputs denoising signals; the method specifically comprises the following steps:
s4.1, taking out the generator network model defined in S3.1, and loading generator model parameters stored by the training module into the generator network model to obtain a complete generator model for denoising;
and S4.2, inputting the noise signal or the filtered signal into a generator model, and outputting a denoising signal.
The data processing module, the training module and the output module are sequentially linked.
The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the equivalents thereof as may occur to those skilled in the art upon consideration of the present inventive concept.
Claims (7)
1. A deep learning-based electromagnetic modulation signal denoising method is characterized by comprising the following steps: the method comprises the following steps:
s1, taking the 18dB 9 modulation type data in the public data set as target signals, and adding quantitative Gaussian noise into the target signals to obtain corresponding 12dB noise signals;
s2, inputting the noise signal in S1 into a low-pass filter, outputting the noise signal as a filtered signal, correspondingly forming a data pair by the target signal in S1 and the newly obtained filtered signal, and then expanding the data volume by ten times by using a signal enhancement method;
s3, constructing and generating a confrontation network model, respectively defining loss functions of a generator and a discriminator, taking the data pairs in S2 as the input of the discriminator, taking the filtered signals as the input of the generator, and minimizing the difference between the generated signals and the target signals through confrontation training so as to obtain the generator for realizing signal denoising;
and S4, extracting the trained generator, inputting the filtered signal, and outputting a de-noising signal.
2. The electromagnetic modulation signal denoising method based on deep learning of claim 1, wherein: the S1 specifically includes:
s1.1, adopting a public data set RML2016.10a, extracting 8910 data with the signal-to-noise ratio of 18dB and the modulation types of 8PSK, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM as target signals;
s1.2, adding quantitative Gaussian noise in the extracted target signal data to obtain corresponding 12dB noise signal data; the noise signal data with the low signal-to-noise ratio of 12dB can be obtained by inputting the target signal data and the low signal-to-noise ratio (SNR) obtained after expected Gaussian noise is added and calculating through the following formula.
Nsignal=Anoise+Asignal (5)
AnoiseTo normalize noise, RnoiseAs random noise, AsignalIs 18db signal amplitude, PsignalIs the signal frequency, n is the signal length, VnoiseAs variance of noise, σnoiseIs the standard deviation of noise, NsignalIs a noise signal.
3. The electromagnetic modulation signal denoising method based on deep learning of claim 1, wherein: the S2 specifically includes:
s2.1, setting a low-pass filter to be 8-order, setting the parameter to be 0.2, inputting the noise signal obtained in the S1.2 into the low-pass filter, and outputting the noise signal which is the filtered signal;
s2.2, loading the filtered signals one by one, and then respectively rotating by 90 degrees, rotating by 180 degrees, rotating by 270 degrees, turning over up and down, turning over left and right, adding weak random noise parameters of (0,0.0005), adding weak random noise parameters of (0,0.0010), adding weak random noise parameters of (0,0.0015) and adding weak random noise parameters of (0,0.0020) to obtain new signals, so that the data volume is enlarged by ten times;
and S2.3, expanding the data volume of the target signal by ten times by using a copy function, splicing all signals mentioned in the S2.2, and forming a data pair by the target signal and the filtered signals one by using an enumeration function, wherein the data structure is (None, 4, 128).
4. The electromagnetic modulation signal denoising method based on deep learning of claim 1, wherein: the S3 specifically includes:
s3.1, defining a generator model; the encoder uses a 6-layer one-dimensional convolutional neural network and uses an activation function with parameters after each layer. And finally changing the input data structure from (None, 2, 128) to (None, 128, 2). The decoder uses 6 layers of one-dimensional inverse convolution neural network, and uses activation function with parameters after each layer; each layer of input is a combination of random noise (or the last layer of deconvolution output data) and the output data of the corresponding layer of the first half network. The data structure output by the final generator is still (None, 2, 128);
s3.2, defining a loss function of the generator model, wherein the loss function of the generator comprises the output loss of the discriminator, the minimum absolute value deviation and a continuity difference value, the continuity difference value (Gtv _ loss) is the average value of the p-th power of the absolute value of the difference value of the first 127 data minus the second 127 data of the I-path signal (Q-path signal) generated by the generator, the continuity degree of the points generating the signals is reflected, and the lower the numerical value is, the better the continuity is;
the loss function (G _ loss) of the generator combines three loss functions:
s3.3, defining a discriminator model; a 6-layer one-dimensional convolution neural network is adopted, a normalization and activation function is used after each layer, and a Dropout layer is added after the 3 rd convolution layer; and then connected with 1 layer of one-dimensional convolution neural network, 1 linear layer and a sigmoid layer. Finally, the data structure is changed from input (None, 4, 128) to output scalar;
s3.4, defining a loss function of the discriminator model; the discriminator mainly uses the least square error to output the value (f (x) of the discriminatori) Minimization of the sum of squares of the differences from the target value (1/0);
when data pairs of a target signal and a noise signal are input, the target value 1 is true, and the loss function is:
when the data pair of the denoised signal and the noise signal is input, the target value 0 is false, and the loss function is:
s3.5, defining a model optimizer, and performing model optimization by adopting an Adam optimizer and an attenuation learning rate;
s3.6, performing countermeasure training, inputting corresponding data pairs into a discriminator when the discriminator discriminates for the first time, expecting the discriminator to discriminate the data pairs of the target signal and the filtered signal as true, and discriminating the data pairs of the de-noised signal and the filtered signal generated by the generator as false; inputting the filtered signal data into a generator to generate a new de-noising signal, and carrying out second-time judger judgment, wherein the new de-noising signal and the filtered signal data are input into a judger to be expected to be judged to be true; repeating the steps to enable the data distribution of the de-noising signal and the target signal to be more and more like; and finally, finishing training when the true and false discrimination probabilities of the discriminator on the data pairs are the same.
5. The electromagnetic modulation signal denoising method based on deep learning of claim 4, wherein: the generator loss specifically includes: the discriminator outputs loss, minimum absolute value deviation and continuity difference.
6. The electromagnetic modulation signal denoising method based on deep learning of claim 1, wherein: the S4 specifically includes:
s4.1, taking out the generator network model defined in S3.1, and loading generator model parameters stored by the training module into the generator network model to obtain a complete generator model for denoising;
and S4.2, inputting the noise signal or the filtered signal into a generator model, and outputting a denoising signal.
7. The electromagnetic modulation signal denoising system based on deep learning is characterized by comprising the following modules: the device comprises a data processing module, a training module and an output module;
the data processing module takes the 9 modulation type data of 18dB in the public data set as target signals, then adds quantitative Gaussian noise into the target signals to obtain corresponding 12dB noise signals, inputs the noise signals into a low-pass filter, outputs the noise signals into filtered signals, correspondingly forms data pairs by the target signals and the newly obtained filtered signals, and then enlarges the data volume by ten times by using a signal enhancement method;
the training module inputs data in the data processing module into a training module, trains the antagonistic network model, iterates for multiple times until the generated antagonistic network model is stably trained, and stores the model;
the output module loads a generator model in the generation confrontation network model stored by the training module, inputs a noise signal to be tested or a filtered signal into the generator model obtained by the training module and outputs a denoising signal;
the data processing module, the training module and the output module are sequentially linked.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110284247.0A CN113111720B (en) | 2021-03-17 | 2021-03-17 | Electromagnetic modulation signal denoising method and system based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110284247.0A CN113111720B (en) | 2021-03-17 | 2021-03-17 | Electromagnetic modulation signal denoising method and system based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113111720A true CN113111720A (en) | 2021-07-13 |
CN113111720B CN113111720B (en) | 2024-06-11 |
Family
ID=76711640
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110284247.0A Active CN113111720B (en) | 2021-03-17 | 2021-03-17 | Electromagnetic modulation signal denoising method and system based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113111720B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114091538A (en) * | 2021-11-23 | 2022-02-25 | 吉林大学 | Signal feature-based loss-discriminating convolutional neural network intelligent noise reduction method |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109118435A (en) * | 2018-06-15 | 2019-01-01 | 广东工业大学 | A kind of depth residual error convolutional neural networks image de-noising method based on PReLU |
CN109242798A (en) * | 2018-09-14 | 2019-01-18 | 南昌大学 | A kind of Poisson denoising method based on three cross-talk network representations |
CN110619885A (en) * | 2019-08-15 | 2019-12-27 | 西北工业大学 | Method for generating confrontation network voice enhancement based on deep complete convolution neural network |
CN110879254A (en) * | 2018-09-05 | 2020-03-13 | 哈尔滨工业大学 | Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network |
CN111178260A (en) * | 2019-12-30 | 2020-05-19 | 山东大学 | Modulation signal time-frequency diagram classification system based on generation countermeasure network and operation method thereof |
CN112380939A (en) * | 2020-11-05 | 2021-02-19 | 浙江工业大学 | Deep learning signal enhancement method based on generation countermeasure network |
CN112488934A (en) * | 2020-11-26 | 2021-03-12 | 杭州电子科技大学 | Finger vein image denoising method based on CS-TCGAN |
-
2021
- 2021-03-17 CN CN202110284247.0A patent/CN113111720B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109118435A (en) * | 2018-06-15 | 2019-01-01 | 广东工业大学 | A kind of depth residual error convolutional neural networks image de-noising method based on PReLU |
CN110879254A (en) * | 2018-09-05 | 2020-03-13 | 哈尔滨工业大学 | Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network |
CN109242798A (en) * | 2018-09-14 | 2019-01-18 | 南昌大学 | A kind of Poisson denoising method based on three cross-talk network representations |
CN110619885A (en) * | 2019-08-15 | 2019-12-27 | 西北工业大学 | Method for generating confrontation network voice enhancement based on deep complete convolution neural network |
CN111178260A (en) * | 2019-12-30 | 2020-05-19 | 山东大学 | Modulation signal time-frequency diagram classification system based on generation countermeasure network and operation method thereof |
CN112380939A (en) * | 2020-11-05 | 2021-02-19 | 浙江工业大学 | Deep learning signal enhancement method based on generation countermeasure network |
CN112488934A (en) * | 2020-11-26 | 2021-03-12 | 杭州电子科技大学 | Finger vein image denoising method based on CS-TCGAN |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114091538A (en) * | 2021-11-23 | 2022-02-25 | 吉林大学 | Signal feature-based loss-discriminating convolutional neural network intelligent noise reduction method |
CN114091538B (en) * | 2021-11-23 | 2024-04-30 | 吉林大学 | Intelligent noise reduction method for discrimination loss convolutional neural network based on signal characteristics |
Also Published As
Publication number | Publication date |
---|---|
CN113111720B (en) | 2024-06-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110619885B (en) | Method for generating confrontation network voice enhancement based on deep complete convolution neural network | |
CN111242862B (en) | Multi-scale fusion parallel dense residual convolution neural network image denoising method | |
CN110706181B (en) | Image denoising method and system based on multi-scale expansion convolution residual error network | |
CN110428849B (en) | Voice enhancement method based on generation countermeasure network | |
CN109890043B (en) | Wireless signal noise reduction method based on generative countermeasure network | |
CN112418014A (en) | Modulation signal identification method based on wavelet transformation and convolution long-short term memory neural network | |
CN111653288A (en) | Target person voice enhancement method based on conditional variation self-encoder | |
CN112581973B (en) | Voice enhancement method and system | |
CN109949821B (en) | Method for removing reverberation of far-field voice by using U-NET structure of CNN | |
CN110490816B (en) | Underwater heterogeneous information data noise reduction method | |
CN112598593B (en) | Seismic noise suppression method based on non-equilibrium depth expectation block log-likelihood network | |
CN111899750B (en) | Speech enhancement algorithm combining cochlear speech features and hopping deep neural network | |
CN115442191B (en) | Communication signal noise reduction method and system based on relative average generation countermeasure network | |
CN110808057A (en) | Voice enhancement method for generating confrontation network based on constraint naive | |
CN111816200B (en) | Multi-channel speech enhancement method based on time-frequency domain binary mask | |
CN113111720A (en) | Electromagnetic modulation signal denoising method and system based on deep learning | |
CN112380939A (en) | Deep learning signal enhancement method based on generation countermeasure network | |
CN110310658B (en) | Voice separation method based on voice signal processing | |
CN116405100B (en) | Distortion signal restoration method based on priori knowledge | |
CN111681649B (en) | Speech recognition method, interaction system and achievement management system comprising system | |
CN112331232A (en) | Voice emotion recognition method combining CGAN spectrogram denoising and bilateral filtering spectrogram enhancement | |
CN111683026B (en) | CMA-SAE-based underwater acoustic channel signal processing method and system | |
TWI749547B (en) | Speech enhancement system based on deep learning | |
CN113066483B (en) | Sparse continuous constraint-based method for generating countermeasure network voice enhancement | |
Karthik et al. | Performance estimation based recurrent-convolutional encoder decoder for speech enhancement |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant |