CN109672639B - Signal demodulation method based on machine learning - Google Patents
Signal demodulation method based on machine learning Download PDFInfo
- Publication number
- CN109672639B CN109672639B CN201910058921.6A CN201910058921A CN109672639B CN 109672639 B CN109672639 B CN 109672639B CN 201910058921 A CN201910058921 A CN 201910058921A CN 109672639 B CN109672639 B CN 109672639B
- Authority
- CN
- China
- Prior art keywords
- signal
- layer
- encoder
- demodulation
- signal demodulation
- 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.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/0008—Modulated-carrier systems arrangements for allowing a transmitter or receiver to use more than one type of modulation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L1/00—Arrangements for detecting or preventing errors in the information received
- H04L1/0001—Systems modifying transmission characteristics according to link quality, e.g. power backoff
- H04L1/0036—Systems modifying transmission characteristics according to link quality, e.g. power backoff arrangements specific to the receiver
Landscapes
- Engineering & Computer Science (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Quality & Reliability (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention relates to a signal demodulation method based on machine learning, which relates to the demodulation of a single modulation mode signal and a plurality of mixed modulation mode signals. The invention aims to solve the problems that different hardware circuits are needed to demodulate and signals with unknown modulation modes are difficult to demodulate when signals with different modulation modes are demodulated. The process is as follows: acquiring an original baseband signal, processing the baseband signal to obtain a noise signal, constructing a denoising autoencoder, and denoising the noise signal to obtain a denoised signal; secondly, obtaining a trained denoising self-encoder network and parameters; thirdly, constructing a signal demodulation model based on a convolutional neural network; fourthly, obtaining a trained signal demodulation model; and fifthly, cascading the trained denoising self-encoder network and the signal demodulation model together to demodulate a single modulation signal and a signal mixed by multiple modulation modes. The invention is used for the field of signal demodulation under a Gaussian white noise channel.
Description
Technical Field
The invention relates to the field of signal demodulation under a Gaussian white noise channel, and relates to the demodulation of a single modulation mode signal and a plurality of mixed modulation mode signals.
Background
With the continuous development of communication technology, people have stronger and stronger requirements on the interconnection of everything, and communication spectrum resources and bandwidths are more and more strained. Generally, it is thought that the utilization rate of the spectrum and the bandwidth is improved by increasing the modulation order to increase the amount of information carried by a transmission symbol, but as the modulation order increases, the demodulation difficulty of a receiver becomes greater and the error rate becomes higher, and then a coherent demodulation method is generally adopted to reduce the error rate. However, in an actual communication system, conditions of a wireless channel are very complex, so that it is difficult for a receiving end to accurately extract information such as carrier phase, time delay and the like, which greatly increases complexity of related demodulation.
The conventional signal demodulation process is divided into two steps of synchronization and demodulation, which first needs to perform synchronization parameter estimation through a series of algorithms, and then demodulates the signal under the condition that the synchronization parameters are known. In order to provide a receiver with real-time carrier extraction and tracking, a communication system needs to adopt a plurality of feedback loops, and generally requires that received data be continuous, but under some conditions, the existing communication system cannot meet the above requirements, even if the data are continuous, and because the locking of the feedback loops needs a certain time, when data is transmitted at a high rate, the receiver has difficulty in achieving carrier synchronization and timing synchronization even if the feedback loops exist. In conventional signal demodulation, the residual synchronization error is usually ignored, which will significantly degrade the demodulation performance.
In order to adapt to different hardware devices and communication environments, different signal modulation modes, such as PSK, QAM, FSK, etc., are generally selected. Modulation and demodulation schemes for radio signals play an important role in wireless communications for the efficient transmission of low bit error rate data. To achieve better performance, many modulation and demodulation techniques have been fully demonstrated by the predecessors. However, if different modulation signals are faced according to the conventional method, different receiver circuits are required for demodulation, resulting in a great increase in the complexity of the receiver circuits.
Disclosure of Invention
The invention aims to solve the problems that different hardware circuits are needed to demodulate signals of different modulation modes and signals with unknown modulation modes are difficult to demodulate in the prior art, and provides a signal demodulation method based on machine learning.
A signal demodulation method based on machine learning comprises the following specific processes:
acquiring an original baseband signal, processing the baseband signal to obtain a noise signal, constructing a denoising autoencoder, and denoising the noise signal to obtain a denoised signal;
the denoising self-encoder comprises an encoder and a decoder;
the specific process is as follows:
constructing an encoder of a denoising autoencoder, and extracting the characteristics of the acquired modulation signals;
step two, constructing a decoder of the denoising autoencoder, and performing denoising restoration on the signal with the extracted characteristics to obtain a signal subjected to denoising processing;
step two, training the neural networks in the encoder and the decoder in the step one, and optimizing parameters to obtain a trained denoising self-encoder network and parameters;
step three, constructing a signal demodulation model based on a convolutional neural network;
step four, performing supervised training on the signal demodulation model based on the convolutional neural network constructed in the step three, initializing parameters of the signal demodulation model by adopting Gaussian distribution, and continuously adjusting and optimizing the parameters to obtain a trained signal demodulation model;
and step five, cascading the denoised self-encoder network trained in the step two and the signal demodulation model trained in the step four together (the networks are cascaded together in sequence), and demodulating the signal mixed by a single modulation signal and multiple modulation modes.
The invention has the beneficial effects that:
1. the invention designs a signal demodulation model based on the convolutional neural network, the model can realize the demodulation function on various single signals, and meanwhile, the model has good performance on resisting the influence of a Gaussian white noise channel. For BPSK modulated signals, the bit error rate is close to 10 at a signal-to-noise ratio of 0dB-1The error code rate has been reduced to 10 when the SNR is higher than 9dB-4The following. For 4-PAM modulation signals, the error rate of model demodulation reaches 10 under the signal-to-noise ratio of 15dB-3Left and right. For FSK modulated signals, the convolutional neural network based demodulation model is comparable toThe error code performance of coherent demodulation theory has some differences, but the precondition of the two is different, the error code curve of coherent demodulation theory is obtained on the premise that the frequency interval is equal to the code element transmission rate, and the convolutional neural network model is obtained on the premise that the frequency interval is equal to half of the code element transmission rate, so the frequency spectrum utilization rate of the design of the invention is higher, by carrying out matlab simulation on the incoherent demodulation of FSK signals under the same condition, experiments show that the frequency interval is reduced for the incoherent demodulation, and the error code performance is greatly reduced, the invention takes the optimal incoherent demodulation as the contrast, under the same condition, the demodulation performance of the convolutional neural network model is improved by 3.5 dB compared with the incoherent demodulation performance, compared with the optimal incoherent reception, the performance of the model of the invention is superior to the optimal incoherent detection when the signal-to-noise ratio is lower than 11dB, at signal to noise ratios above 11dB, the performance is slightly lower than optimal incoherent detection. Meanwhile, the demodulation model can demodulate various signals after being trained, compared with other demodulation modes based on learning, the model of the invention does not need to manually design features, compared with the demodulation mode based on non-learning, the model of the invention can demodulate various signals without replacing hardware and algorithms, only the model is trained in advance, and the signals with unknown modulation modes can be demodulated, only the model is trained in advance.
2. The signal demodulation model designed by the invention can uniformly demodulate BPSK, FSK and 4-PAM modulation signals, which cannot be realized by the traditional method and is also a unique characteristic of a deep learning-based method. Under the condition of 10dB signal-to-noise ratio, the error rate of hybrid demodulation can still reach 10-2Left and right, comparing the single demodulation results of different signals, it can be found that the signal demodulation model has independence when demodulating various signals, that is, the demodulation of various signals does not affect each other, and the main factor affecting the performance is the signal-to-noise ratio of the signal.
3. The invention adopts a denoising autoencoder as an improved method of pre-training. The scheme of using the denoising self-encoder as the pre-training can greatly enhance the anti-noise interference performance of the network, thereby improving the performance of the algorithm. When various modulation signals are uniformly demodulated, the performance of the improved method of the denoising self-encoder is improved by 1-2dB compared with the method of a random initialization model.
Drawings
FIG. 1 is a basic flow diagram of the signal demodulation model based on convolutional neural network of the present invention;
FIG. 2 is a diagram of a signal demodulation model based on a convolutional neural network according to the present invention;
FIG. 3 is a graph of the performance of the signal demodulation model of the present invention for demodulating FSK modulated signals over an AWGN channel;
FIG. 4 is a graph of the performance of the signal demodulation model of the present invention for the unified demodulation of three modulated signals, BPSK, FSK, and 4-PAM, in an AWGN channel;
FIG. 5 is a basic flow diagram of the signal demodulation model with a denoised autoencoder of the present invention;
FIG. 6a is a diagram of a neural network structure of the denoising autoencoder of the present invention;
FIG. 6b is a block diagram of the neural network of the decoder of the denoising autoencoder of the present invention;
FIG. 7 is a comparison graph of the unified demodulation performance curves of the signal demodulation model of the present invention after being improved by the de-noising self-encoder model for three carrier modulation signals BPSK, FSK and 4-PAM;
fig. 8 is a graph of the demodulation performance of the signal demodulation model of the present invention under AWGN channel for the 4-PAM single carrier modulation signal after being improved by the de-noising and de-noising encoder model.
Detailed Description
The first embodiment is as follows: the signal demodulation method based on machine learning of the embodiment comprises the following specific processes:
acquiring an original baseband signal, processing the baseband signal to obtain a noise signal, constructing a denoising autoencoder, and denoising the noise signal to obtain a denoised signal; as shown in fig. 5, 6a, 6 b;
the denoising self-encoder comprises an encoder and a decoder;
the specific process is as follows:
constructing an encoder of a denoising autoencoder, and extracting the characteristics of the acquired modulation signals;
step two, constructing a decoder of the denoising autoencoder, and performing denoising restoration on the signal with the extracted characteristics to obtain a signal subjected to denoising processing;
step two, training the neural networks in the encoder and the decoder in the step one, and optimizing parameters to obtain a trained denoising self-encoder network and parameters;
step three, constructing a signal demodulation model based on a convolutional neural network; as shown in fig. 1 and 2;
step four, performing supervised training on the signal demodulation model based on the convolutional neural network constructed in the step three, initializing parameters of the signal demodulation model by adopting Gaussian distribution, and continuously adjusting and optimizing the parameters to obtain a trained signal demodulation model;
and step five, cascading the denoised self-encoder network trained in the step two and the signal demodulation model trained in the step four together (the networks are cascaded together in sequence), and demodulating the signal mixed by a single modulation signal and multiple modulation modes.
The second embodiment is as follows: the difference between the first embodiment and the first embodiment is that the first step collects an original baseband signal, processes the baseband signal, and obtains a noise signal; the specific process is as follows:
different modulation modes refer to that an original signal is a baseband signal and then modulated onto a carrier, and the modulation mode of the signal is uncertain and can be PAM modulation or BPSK modulation and other different modulation modes;
step 2, sampling the communication signals added with the white gaussian noise in the step 1, wherein each symbol is 128 points, 4 communication symbols are one sample data, that is, 512 sampling points are sampled to form a group, the signals added with the white gaussian noise are randomly disordered by taking the 512 signal sampling points as a unit, the random disordering is in one-to-one correspondence with the original signals not added with the white gaussian noise, the sampling points are divided into training samples and test samples according to a ratio of 4:1, each training sample and each test sample are composed of data and a label, the data is the signals added with the white gaussian noise, and the label is the signals not added with the white gaussian noise;
the unit of the communication signal comprises symbols and bits and is related to different modulation modes;
sending the training sample into a denoising autoencoder for training to obtain stable denoising parameters;
the training samples and the test samples contain a signal after adding white gaussian noise and a signal without adding white gaussian noise.
Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: the first or second embodiment is different from the first or second embodiment in that the encoder of the denoising autoencoder is constructed in the first step, and the specific process is as follows:
step one, adopting a convolutional neural network to construct a network structure of an encoder of a denoising self-encoder, wherein the network structure comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and an output layer;
the first convolution layer and the second convolution layer use a RELU activation function, and the form is g (z) ═ max {0, z }, namely when the input is less than 0, the output is 0, and when the input is more than zero, the output is equal to the input;
the length of the first convolution kernel is set to 5 x 5, and the number of convolution kernels is set to 16;
the length of the second convolution kernel is set to 5 x 5, and the number of convolution kernels is set to 32;
the step length of the first pooling layer is 2, and the width of the pooling is 3;
the step length of the second pooling layer is 2, and the width of the pooling is 3;
and step one and two, inputting the training samples obtained in the step 2 into an encoder of the constructed denoising autoencoder, and obtaining the output h of the encoder, wherein the output h is f (x).
Other steps and parameters are the same as those in the first or second embodiment.
The fourth concrete implementation mode: the difference between this embodiment and one of the first to third embodiments is that, in the second step, a decoder of the denoising autoencoder is constructed, and the specific process is as follows:
step two, constructing a network structure of a decoder of the denoising self-encoder by adopting a convolutional neural network, wherein the network structure comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and an output layer;
the first convolution layer and the second convolution layer use a RELU activation function, and have the form g (z) ═ max {0, z }, namely when the input is less than 0, the output is 0, and when the input is greater than zero, the output is equal to the input;
the length of the first convolution kernel is set to 5 x 5, and the number of convolution kernels is set to 16;
the length of the second convolution kernel is set to 5 x 5, and the number of convolution kernels is set to 32;
the step length of the first pooling layer is 2, and the width of the pooling is 3;
the step length of the second pooling layer is 2, and the width of the pooling is 3;
and step two, inputting the encoder output h ═ f (x) obtained in step one and two into the decoder of the denoising self-encoder constructed in step one and two to obtain the decoder output r ═ g (h (x)).
Other steps and parameters are the same as those in one of the first to third embodiments.
The fifth concrete implementation mode: the difference between this embodiment and one of the first to fourth embodiments is that, in the third step, a signal demodulation model based on a convolutional neural network is constructed, and the specific process is as follows:
the signal demodulation model network structure based on the convolutional neural network comprises an input layer, a first convolutional layer, a first pooling layer, a first activation unit, a second convolutional layer, a second pooling layer, a second activation unit, a random deactivation layer, two full-connection layers and an output layer;
the output r ═ g (h (x)) of the decoder is used as the input of the input layer of the signal demodulation model, the number of input neurons of the input layer is set to 512, which corresponds to 8-bit data of each sample and 512 sampling points; 4 communication symbols are sample data, 4 communication symbols are 8 bit symbols, and the sampled data form a sample;
the length of the convolution kernel of the first volume base layer is set to be 5 x 5, and the number of the convolution kernels is set to be 16;
the length of the convolution kernel of the second volume base layer is set to be 5 x 5, and the number of the convolution kernels is set to be 32;
the step length of the first pooling layer is 2, and the width of the pooling is 3;
the step length of the second pooling layer is 2, and the width of the pooling is 3;
the first and second activation units use the RELU activation function, which is in the form of g (z) max {0, z }, i.e. when the input is less than 0, the output is 0, and when the input is greater than zero, the output is equal to the input;
the parameter of the random inactivation layer is set to be 20%, namely only 80% of data output by the network of the previous layer is reserved, and the rest 20% is set to be 0;
the output layer output structure is a Sigmoid unit with 8 neurons and is used for outputting 8-bit demodulation bit data.
Because the number of the network layers is small, and the random inactivation layer is added in the middle, the overfitting phenomenon of the network can be effectively prevented.
Other steps and parameters are the same as those in one of the first to fifth embodiments.
The sixth specific implementation mode: the difference between the first embodiment and the fifth embodiment is that in the fourth step, supervised training is performed on the signal demodulation model based on the convolutional neural network, which is constructed in the third step, parameters of the signal demodulation model are initialized by adopting gaussian distribution, and the parameters are continuously adjusted and optimized to obtain a trained signal demodulation model; the specific process is as follows:
performing supervised training on the signal demodulation model based on the convolutional neural network constructed in the step three by using a random gradient descent algorithm, and setting the learning rate to be 0.0001;
when the signal-to-noise ratio is greater than or equal to 10dB, initializing a signal demodulation model parameter by Gaussian distribution, enabling the parameter of the demodulation model to tend to be stable through iteration and optimization, and enabling the error rate after demodulation to be reduced to the lowest;
and taking the obtained signal demodulation model parameters as the initialized signal demodulation model parameters when the signal-to-noise ratio is lower than 10dB, and enabling the demodulation performance of the signal demodulation model to be optimal, namely the error function is not reduced any more and tends to be stable through iteration and optimization to obtain the trained signal demodulation model.
Because the optimization algorithm is iterative, an initial value needs to be specified, and parameters are initialized randomly by adopting Gaussian distribution.
Under the condition of high signal-to-noise ratio, firstly, initializing a signal demodulation model parameter by utilizing Gaussian distribution, enabling the parameter of the demodulation model to tend to be stable through iteration and optimization, and reducing the error rate after demodulation to the minimum;
and taking the obtained signal demodulation model parameters as initialization model parameters under other low signal-to-noise ratio conditions, and then performing iteration and optimization to enable the demodulation performance of the model to be optimal. The invention considers that in the training process of the signal with high signal-to-noise ratio, the parameters after iterative updating have certain capability of extracting features from the signal covered by the noise. When training is started with this initial value, it is equivalent to only needing to let the neural network learn the ability to extract features from signals more heavily polluted by noise, which is equivalent to reducing the difficulty of learning, thereby improving performance.
Other steps and parameters are the same as those in one of the first to fifth embodiments.
Fig. 3 is a graph showing the performance of the signal demodulation model of the present invention for demodulating FSK modulated signals in AWGN channels;
fig. 4 is a performance curve diagram of the signal demodulation model of the present invention for uniformly demodulating BPSK, FSK, and 4-PAM modulated signals in an AWGN channel;
FIG. 7 is a comparison graph of the unified demodulation performance curve for BPSK, FSK, and 4-PAM three carrier modulation signals after the signal demodulation model of the present invention is improved by the de-noising self-encoder model;
fig. 8 is a graph of the demodulation performance of the signal demodulation model of the present invention under AWGN channel for the 4-PAM single carrier modulation signal after being improved by the de-noising and de-noising encoder model.
The bit error rate of the invention is close to 10 when the signal to noise ratio is 0dB for BPSK modulation signals-1The error code rate has been reduced to 10 when the SNR is higher than 9dB-4The following. For 4-PAM modulation signals, the error rate of model demodulation reaches 10 under the signal-to-noise ratio of 15dB-3Left and right. For FSK modulation signals, although a demodulation model based on a convolutional neural network has some difference compared with the error code performance of coherent demodulation theory, the two preconditions are different, the error code curve of coherent demodulation theory is obtained on the premise that the frequency interval is equal to the code element transmission rate, and the convolutional neural network model is obtained on the premise that the frequency interval is equal to half of the code element transmission rate, so the spectrum utilization rate designed by the invention is higher, matlab simulation is carried out on the incoherent demodulation of FSK signals under the same condition, experiments show that the frequency interval is smaller for the incoherent demodulation, the error code performance is greatly reduced, the invention uses the optimal incoherent demodulation as a contrast, the demodulation performance of the convolutional neural network model is improved by 3.5 dB compared with the incoherent demodulation performance under the same condition, compared with the optimal incoherent reception, the model of the invention has better performance than the best incoherent detection when the signal-to-noise ratio is lower than 11dB, and has slightly lower performance than the best incoherent detection when the signal-to-noise ratio is higher than 11 dB. Meanwhile, the demodulation model can demodulate various signals after being trained, compared with other demodulation modes based on learning, the model of the invention does not need to manually design features, compared with the demodulation mode based on non-learning, the model of the invention can demodulate various signals without replacing hardware and algorithms, only the model is trained in advance, and the signals with unknown modulation modes can be demodulated, only the model is trained in advance.
The signal demodulation model designed by the invention can be used for carrying out BPSK, FSK and 4-PAM modulation signalsUnified demodulation, which cannot be achieved by the traditional method, is also a unique characteristic of a deep learning-based method. Under the condition of 10dB signal-to-noise ratio, the error rate of hybrid demodulation can still reach 10-2Left and right, comparing the single demodulation results of different signals, it can be found that the signal demodulation model has independence when demodulating various signals, that is, the demodulation of various signals does not affect each other, and the main factor affecting the performance is the signal-to-noise ratio of the signal.
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.
Claims (6)
1. A signal demodulation method based on machine learning, characterized by: the method comprises the following specific processes:
acquiring an original baseband signal, processing the baseband signal to obtain a noise signal, constructing a denoising autoencoder, and denoising the noise signal to obtain a denoised signal;
the denoising self-encoder comprises an encoder and a decoder;
the specific process is as follows:
constructing an encoder of a denoising autoencoder, and extracting the characteristics of the acquired modulation signals;
step two, constructing a decoder of the denoising autoencoder, and performing denoising restoration on the signal with the extracted characteristics to obtain a signal subjected to denoising processing;
step two, training the neural networks in the encoder and the decoder in the step one, and optimizing parameters to obtain a trained denoising self-encoder network and parameters;
step three, constructing a signal demodulation model based on a convolutional neural network;
step four, performing supervised training on the signal demodulation model based on the convolutional neural network constructed in the step three, initializing parameters of the signal demodulation model by adopting Gaussian distribution, and continuously adjusting and optimizing the parameters to obtain a trained signal demodulation model;
and step five, cascading the denoised self-encoder network trained in the step two and the signal demodulation model trained in the step four together to demodulate a single modulation signal and a signal mixed by multiple modulation modes.
2. A method for demodulating a signal based on machine learning according to claim 1, wherein: acquiring an original baseband signal in the first step, and processing the baseband signal to obtain a noise signal; the specific process is as follows:
step 1, collecting an original baseband signal, carrying out carrier modulation on the baseband signal to generate signals with different modulation modes, and adding Gaussian white noise to the collected carrier modulation signal according to the signal-to-noise ratio range of 0-15dB and with 1dB as a step length;
and 2, sampling the communication signals added with the white Gaussian noise in the step 1, wherein each symbol is 128 points, 4 communication symbols are one sample data, that is, 512 sampling points are sampled to form a group, the signals added with the white Gaussian noise are randomly disordered by taking the 512 signal sampling points as a unit, the random disordering is in one-to-one correspondence with the original signals not added with the white Gaussian noise, the sampling is divided into training samples and test samples according to a ratio of 4:1, each training sample and each test sample are composed of data and a label, the data is the signals added with the white Gaussian noise, and the label is the signals not added with the white Gaussian noise.
3. A method for demodulating a signal based on machine learning according to claim 1 or 2, characterized in that: the encoder of the denoising autoencoder is constructed in the steps, and the specific process is as follows:
step one, adopting a convolutional neural network to construct a network structure of an encoder of a denoising self-encoder, wherein the network structure comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and an output layer;
the first convolution layer and the second convolution layer use a RELU activation function;
the length of the first convolution layer convolution kernel is set to be N x N, and the number of the convolution kernels is set to be M;
the length of the second convolution layer convolution kernel is set to be N x N, and the number of the convolution kernels is set to be Q;
the step length of the first pooling layer is L1, and the width of pooling is L2;
the step length of the second pooling layer is L1, and the width of pooling is L2;
and step one and two, inputting the training samples obtained in the step 2 into an encoder of the constructed denoising autoencoder, and obtaining the output h of the encoder, wherein the output h is f (x).
4. A method for demodulating a signal based on machine learning according to claim 3, wherein: in the second step, a decoder of the denoising autoencoder is constructed, and the specific process is as follows:
step two, constructing a network structure of a decoder of the denoising self-encoder by adopting a convolutional neural network, wherein the network structure comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and an output layer;
the first convolution layer and the second convolution layer use a RELU activation function;
the length of the first convolution layer convolution kernel is set to be N x N, and the number of the convolution kernels is set to be M;
the length of the second convolution layer convolution kernel is set to be N x N, and the number of the convolution kernels is set to be Q;
the step length of the first pooling layer is L1, and the width of pooling is L2;
the step length of the second pooling layer is L1, and the width of pooling is L2;
and step two, inputting the encoder output h ═ f (x) obtained in step one and two into the decoder of the denoising self-encoder constructed in step one and two to obtain the decoder output r ═ g (f (x)).
5. The signal demodulation method based on machine learning according to claim 4, characterized in that: in the third step, a signal demodulation model based on a convolutional neural network is constructed, and the specific process is as follows:
the signal demodulation model network structure based on the convolutional neural network comprises an input layer, a first convolutional layer, a first pooling layer, a first activation unit, a second convolutional layer, a second pooling layer, a second activation unit, a random deactivation layer, two full-connection layers and an output layer;
the output r ═ g (f (x)) of the decoder is used as the input of the signal demodulation model input layer;
setting the length of a first volume of base layer convolution kernels to be N x N, and setting the number of convolution kernels to be M;
setting the length of the convolution kernel of the second volume of the base layer as N x N, and setting the number of the convolution kernels as Q;
the step length of the first pooling layer is L1, and the width of pooling is L2;
the step length of the second pooling layer is L1, and the width of pooling is L2;
the first activation unit and the second activation unit use the RELU activation function;
the parameter of the random inactivation layer is set to be 20%, namely only 80% of data output by the network of the previous layer is reserved, and the rest 20% is set to be 0;
the output layer output structure is a Sigmoid unit with 8 neurons and is used for outputting 8-bit demodulation bit data.
6. The signal demodulation method based on machine learning according to claim 5, characterized in that: in the fourth step, supervised training is carried out on the signal demodulation model based on the convolutional neural network constructed in the third step, Gaussian distribution is adopted to initialize parameters of the signal demodulation model, and the parameters are continuously adjusted and optimized to obtain a trained signal demodulation model; the specific process is as follows:
carrying out supervised training on the signal demodulation model based on the convolutional neural network constructed in the step three by using a random gradient descent algorithm, and setting the learning rate as eta;
when the signal-to-noise ratio is greater than or equal to 10dB, initializing a signal demodulation model parameter by Gaussian distribution, enabling the parameter of the demodulation model to tend to be stable through iteration and optimization, and enabling the error rate after demodulation to be reduced to the lowest;
and taking the obtained signal demodulation model parameters as the initialized signal demodulation model parameters when the signal-to-noise ratio is lower than 10dB, and enabling the demodulation performance of the signal demodulation model to be optimal, namely the error function is not reduced any more and tends to be stable through iteration and optimization to obtain the trained signal demodulation model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910058921.6A CN109672639B (en) | 2019-01-22 | 2019-01-22 | Signal demodulation method based on machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910058921.6A CN109672639B (en) | 2019-01-22 | 2019-01-22 | Signal demodulation method based on machine learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109672639A CN109672639A (en) | 2019-04-23 |
| CN109672639B true CN109672639B (en) | 2021-07-16 |
Family
ID=66150771
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910058921.6A Active CN109672639B (en) | 2019-01-22 | 2019-01-22 | Signal demodulation method based on machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109672639B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110233810B (en) * | 2019-06-25 | 2021-07-02 | 电子科技大学 | MSK signal demodulation method based on deep learning under mixed noise |
| CN111342896B (en) * | 2020-03-02 | 2021-04-02 | 深圳市南科信息科技有限公司 | Self-coding algorithm based on convolutional neural network |
| CN113094993B (en) * | 2021-04-12 | 2022-03-29 | 电子科技大学 | Modulation signal denoising method based on self-coding neural network |
| CN113395225B (en) * | 2021-06-10 | 2023-01-31 | 中国人民解放军战略支援部队信息工程大学 | Universal intelligent processing method and device for directly receiving communication signal waveform to bit |
| CN113630130B (en) * | 2021-07-31 | 2022-12-23 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | End-to-end digital communication demodulation method |
| CN116938658A (en) * | 2022-03-31 | 2023-10-24 | 华为技术有限公司 | Data processing method, device and system |
| CN114938248B (en) * | 2022-07-26 | 2022-11-01 | 中国海洋大学三亚海洋研究院 | Method for building and demodulating underwater wireless optical communication demodulation model |
| CN117714244B (en) * | 2024-02-05 | 2024-05-31 | 丝路梵天(甘肃)通信技术有限公司 | Wireless signal reconstruction method, device, receiver and system based on neural network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101594154A (en) * | 2009-07-14 | 2009-12-02 | 北京航空航天大学 | TCM neural network demodulation method based on code restoration |
| CN107147603A (en) * | 2017-05-05 | 2017-09-08 | 西安电子科技大学 | DBPSK demodulation methods based on multiple neural network |
| CN108282437A (en) * | 2018-01-08 | 2018-07-13 | 西安电子科技大学 | Based on simulated annealing neural network and eliminate the data detection method interfered |
| CN109067688A (en) * | 2018-07-09 | 2018-12-21 | 东南大学 | A kind of OFDM method of reseptance of data model double drive |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9717440B2 (en) * | 2013-05-03 | 2017-08-01 | The Florida International University Board Of Trustees | Systems and methods for decoding intended motor commands from recorded neural signals for the control of external devices or to interact in virtual environments |
| CN108234370B (en) * | 2017-12-22 | 2020-12-15 | 西安电子科技大学 | Communication signal modulation mode identification method based on convolutional neural network |
-
2019
- 2019-01-22 CN CN201910058921.6A patent/CN109672639B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101594154A (en) * | 2009-07-14 | 2009-12-02 | 北京航空航天大学 | TCM neural network demodulation method based on code restoration |
| CN107147603A (en) * | 2017-05-05 | 2017-09-08 | 西安电子科技大学 | DBPSK demodulation methods based on multiple neural network |
| CN108282437A (en) * | 2018-01-08 | 2018-07-13 | 西安电子科技大学 | Based on simulated annealing neural network and eliminate the data detection method interfered |
| CN109067688A (en) * | 2018-07-09 | 2018-12-21 | 东南大学 | A kind of OFDM method of reseptance of data model double drive |
Non-Patent Citations (2)
| Title |
|---|
| 《Neural Network Demodulator for Frequency Shift Keying》;Min Li,et al.;《2008 International Conference on Computer Science and Software Engineering》;20081222;全文 * |
| 《基于神经网络的FSK信号解调技术研究》;李岷;《CNKI中国知网》;20090821;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109672639A (en) | 2019-04-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109672639B (en) | Signal demodulation method based on machine learning | |
| Ye et al. | Power of deep learning for channel estimation and signal detection in OFDM systems | |
| CN109617847A (en) | A kind of non-cycle prefix OFDM method of reseptance based on model-driven deep learning | |
| CN113630130B (en) | End-to-end digital communication demodulation method | |
| Zhang et al. | Deep learning based on orthogonal approximate message passing for CP-free OFDM | |
| CN111064512B (en) | Deep learning-based optical orthogonal frequency division multiplexing modulation method and system | |
| CN113381951A (en) | MFTN joint channel estimation and equalization method under time-frequency conversion selective fading channel | |
| CN112003804B (en) | Multipath multivariate differential chaotic shift keying iterative receiving method | |
| CN112291005A (en) | Bi-LSTM neural network-based receiving end signal detection method | |
| CN112242969A (en) | Novel single-bit OFDM receiver based on model-driven deep learning | |
| CN116132239A (en) | OFDM channel estimation method adopting pre-activation residual error unit and super-resolution network | |
| CN106452652B (en) | A kind of MPI suppression method based on chaos wireless communication system | |
| Khan et al. | Symbol denoising in high order M-QAM using residual learning of deep CNN | |
| CN111277530A (en) | Modulation-demodulation method for carrier system keying | |
| CN114595729A (en) | Communication signal modulation identification method based on residual error neural network and meta-learning fusion | |
| CN114826832B (en) | Channel estimation method, neural network training method, device and equipment | |
| CN110290083A (en) | Multicarrier difference chaotic Demodulation Systems method based on low-rank matrix estimation | |
| CN113726697A (en) | OTFS symbol detection method based on confidence space dynamic decision | |
| CN112636855A (en) | OFDM signal detection method | |
| Li et al. | Deep gated recurrent unit convolution network for radio signal recognition | |
| CN109639301B (en) | Confidence estimation-based FTN (fiber to the home) equalization method | |
| Wong et al. | Deep Learning for Channel Estimation and Signal Detection in OFDM-Based Communication Systems | |
| CN112671419B (en) | Wireless signal reconstruction method, device, system, equipment and storage medium | |
| CN115913849A (en) | Electromagnetic signal identification method based on one-dimensional complex value residual error network | |
| CN115883001A (en) | Anti-interference method for wireless communication |
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 | ||
| GR01 | Patent grant |