OFDM underwater acoustic communication receiving method based on deep learning

Abstract

The invention relates to the technical field of underwater acoustic communication, in particular to an OFDM underwater acoustic communication receiving method based on deep learning, and the method comprises the steps: 1, building an OFDM system, and generating a data set; 2, constructing a channel estimation model based on a CNN (convolutional neural network); step 3, constructing a signal demodulation model based on the long and short term memory network BiLSTM; step 4, setting hyper-parameters of the CNN network and the BiLSTM network; 5, training the channel estimation model and the signal demodulation model by using the data set; step 6, counting an accuracy rate curve, a network loss curve and training time, if the accuracy rate curve is not fitted or the network loss curve is not converged, resetting hyper-parameters, and carrying out step 4, otherwise, carrying out step 7; and 7, inputting the test set into a deep learning assisted OFDM underwater acoustic communication receiving model, and carrying out online deployment. And the precision and the accuracy of channel estimation can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic communication, and in particular to an OFDM underwater acoustic communication receiving method based on deep learning. Background Art

[0002] Underwater acoustic channel (UWA) has serious transmission loss, frequency-related environmental noise, significant multipath effect and Doppler effect, which are mainly due to the special properties of underwater acoustic propagation media and the complexity of underwater environment. Underwater acoustic channel modeling and prediction are very difficult, and the bandwidth of underwater acoustic channel is limited, with millisecond delay spread, which seriously reduces the signal reception performance and limits the rate of underwater acoustic communication, resulting in a high bit error rate at the receiving end and interference with the received signal. OFDM (Orthogonal Frequency Division Multiplexing) technology has become a key technology in the field of modern communications due to its excellent performance in high-speed data transmission, robustness in multipath propagation environment and efficient use of spectrum.

[0003] OFDM is widely used in UWA communications due to its good performance against inter-symbol interference (ISI) and multipath fading. In order to demodulate the correct original symbols at the receiving end, the use of a suitable channel estimation algorithm is crucial to the performance of the UWA OFDM system. Currently, the least squares (LS) algorithm is a popular algorithm used today. Its low complexity is a significant feature of LS, but it also ignores the interference between subcarriers, which makes LS show poor bit error rate performance when the number of pilots and cyclic prefixes is small.

[0004] In fast-changing and complex UWA channels, especially in actual UWA environments, receivers of traditional algorithms have problems such as large pilot overhead, high complexity, and poor robustness. At the same time, they cannot match the channel well, thus reducing the accuracy of data recovery. Summary of the invention

[0005] In view of the problems existing in the prior art, the purpose of the present invention is to provide an OFDM underwater acoustic communication receiving method based on deep learning to improve the accuracy of data recovery.

[0006] To achieve the above object, the technical solution adopted by the present invention is: an OFDM underwater acoustic communication receiving method based on deep learning, comprising the following steps: Step 1: Build an OFDM system and generate a data set; the initial OFDM signal transmitted by the transmitter undergoes channel preprocessing to generate the data set, and the receiver receives the OFDM signal after passing through the channel; Step 2: Construct a channel estimation model based on convolutional neural network (CNN); the expression is obtained based on the input data: The channel frequency response CFR of the leading subcarrier is obtained, and the channel frequency response of the leading subcarrier is split as the input of the convolutional neural network CNN, where: is a known pilot symbol, is the received pilot symbol, is the channel frequency response of the pilot subcarrier; Step 3: Construct a signal demodulation model based on the long short-term memory network BiLSTM; use the traditional channel equalization algorithm to obtain the expression: The initial data of the sequence Input to the signal demodulation part to accurately estimate the transmitted data symbols, where To obtain the initial data, represents the received data subcarrier, represents the CFR of the data subcarrier obtained by the proposed channel estimation network; Step 4: Set the hyperparameters of the CNN network and BiLSTM network; Step 5: Use the data set to train the channel estimation model and signal demodulation model; Step 6: Count the accuracy curve, network loss curve and training time. If the accuracy curve does not fit or the network loss curve does not converge, reset the hyperparameters and proceed to step 4. Otherwise, proceed to step 7. Step 7: Input the test set into the deep learning-assisted OFDM underwater acoustic communication receiving model for online deployment.

[0007] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 1, the channel preprocessing includes modulation, pilot insertion, serial-to-parallel conversion, inverse fast Fourier transform, adding cyclic prefix, parallel-to-serial conversion and fast Fourier transform.

[0008] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 2, the channel estimation model is composed of two one-dimensional convolutional layers, a self-attention module and two one-dimensional deconvolutional layers cascaded in sequence, and the ReLU function is selected as the activation function. A jump connection is set between the first one-dimensional convolutional layer and the first one-dimensional deconvolutional layer.

[0009] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 2, the input data includes a known pilot symbol and a received pilot symbol, and the step 2 further includes: Step 2-1: Split the real part and the imaginary part of the pilot subcarrier and concatenate them together as the input of the channel estimation model; Step 2-2: After extracting the underwater acoustic channel features using two one-dimensional convolutional layers, the features are transferred to the self-attention module; Step 2-3: The self-attention module captures the global features, and then reconstructs the channel frequency response of the data subcarrier through two one-dimensional deconvolution layers, which is expressed as: ,in, represents the input of the neural network, is the final output CFR, represents the function of the nth layer of CNN, , , , Represents the weight matrix of each layer of CNN, and , , , Represents the bias matrix of each layer, and A represents the attention weight matrix.

[0010] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 3, the signal demodulation model is composed of a bidirectional long short-term memory network BiLSTM and a fully connected layer, containing three layers of BiLSTM networks, each layer has 64 time steps and 64 hidden units respectively, and finally a FC layer containing 64 neurons, using a Sigmoid activation function to perform binary classification output.

[0011] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 3, the output of the binary classification using the Sigmoid activation function is expressed as: ,in, Represents the output after passing through the BiLSTM network.

[0012] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 3, the step of using the Sigmoid activation function to perform binary classification output includes: The threshold of the Sigmoid function is set to 0.5. If the output is greater than 0.5, the received symbol is considered to be 1, otherwise it is 0. The calculation formula of BiLSTM is: , , , , , ,in, , , ,and They are the input gate, forget gate, output gate and candidate state representing the current information. , , , , , , , Represents input The weight matrix of , , , Represent the bias terms of each, represents the hidden state at each time step, Indicates the unit status, and They are Sigmoid and hyperbolic tangent functions, Represents element-wise product.

[0013] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, step 5 comprises: Step 5-1: Divide the data set into a training set and a test set in a ratio of 8:2; Step 5-2: Use the Adam optimizer with a learning rate of 0.001 and cross entropy as the loss function to train the channel estimation model. The loss function is: , where y represents the true label and p represents the predicted value of the model; Step 5-3: Use the trained channel estimation model to train the signal demodulation model.

[0014] In the above-mentioned OFDM underwater acoustic communication receiving method based on deep learning, in step 7, the modulation method adopted by the OFDM underwater acoustic communication model is QPSK, the OFDM underwater acoustic communication transmission model sends a test signal, and generates a test sample after passing through the measured underwater acoustic channel, and the test sample is input into the model for prediction.

[0015] The beneficial effects of the OFDM underwater acoustic communication receiving method based on deep learning of the present invention are: the channel estimation method based on CNN can greatly improve the precision and accuracy of channel estimation; the signal detection model based on BiLSTM proposed by the present invention has higher detection accuracy and robustness, and has better performance and signal detection ability. By using jump connection, the output after the first convolution layer is directly used as the input of the first deconvolution layer, reducing the problem of feature loss that may occur during information propagation and gradient disappearance during training. An attention structure is added after the one-dimensional convolution layer to help model global features and dynamically estimate channels.

[0016] The present invention generates a data set through an OFDM system, then constructs a channel estimation model using a convolutional neural network, and uses a BiLSTM network to construct a signal demodulation model; secondly, the data set is input into the neural network, the model is trained, and the network is optimized using the Adam optimizer and the cross entropy loss function, which effectively solves the problem that the traditional underwater acoustic OFDM system cannot adapt to the rapid changes of the underwater acoustic channel and has a high demodulation bit error rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a specific embodiment of the present invention; Figure 2 This is the improved CNN+BiLSTM network structure diagram of the present invention; Figure 3 is a structural diagram of the channel estimation subnet in the present invention; Figure 4 It is a structural diagram of the signal detection subnet in the present invention; Figure 5 It is a BER comparison diagram of the present invention, traditional methods, FC-DNN, CE, CNN and LSTM under NOF channel conditions; Figure 6 It is a BER comparison diagram of the present invention, traditional methods, FC-DNN, CE, CNN and LSTM under NCS channel conditions; Figure 7 It is a BER comparison diagram of the present invention, traditional methods, FC-DNN, CE, CNN and LSTM under SIM channel conditions; Figure 8 This is a BER comparison chart of the present invention with the traditional method, F-DNN, CE, CNN and LSTM in the absence of cyclic prefix. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is described below in conjunction with specific implementation methods and drawings.

[0019] Example 1 A deep learning-based OFDM underwater acoustic communication receiving method comprises the following steps: Step 1: Build an OFDM system and generate a data set; the initial OFDM signal transmitted by the transmitter undergoes channel preprocessing to generate the data set, and the receiver receives the OFDM signal after passing through the channel.

[0020] Channel preprocessing includes modulation, pilot insertion, serial-to-parallel conversion, inverse fast Fourier transform, adding cyclic prefix, parallel-to-serial conversion and fast Fourier transform.

[0021] Step 2: Construct a channel estimation model based on convolutional neural network (CNN); the expression is obtained based on the input data: The channel frequency response CFR of the leading subcarrier is obtained, and the channel frequency response of the leading subcarrier is split as the input of the convolutional neural network CNN, where: is a known pilot symbol, is the received pilot symbol, is the channel frequency response of the pilot subcarrier.

[0022] The channel estimation model consists of two one-dimensional convolutional layers, a self-attention module and two one-dimensional deconvolutional layers cascaded in sequence. The ReLU function is selected as the activation function, and a jump connection is set between the first one-dimensional convolutional layer and the first one-dimensional deconvolutional layer.

[0023] The input data includes known pilot symbols and received pilot symbols, including: Step 2-1: Split the real and imaginary parts of the pilot subcarrier and concatenate them together as input to the channel estimation model.

[0024] Step 2-2: After extracting the underwater acoustic channel features using two one-dimensional convolutional layers, they are transferred to the self-attention module.

[0025] Step 2-3: The self-attention module captures the global features, and then reconstructs the channel frequency response of the data subcarrier through two one-dimensional deconvolution layers, which is expressed as: ,in, represents the input of the neural network, is the final output CFR, represents the function of the nth layer of CNN, , , , Represents the weight matrix of each layer of CNN, and , , , Represents the bias matrix of each layer, and A represents the attention weight matrix.

[0026] Step 3: Construct a signal demodulation model based on the long short-term memory network BiLSTM; use the traditional channel equalization algorithm to obtain the expression: The initial data of the sequence Input to the signal demodulation part to accurately estimate the transmitted data symbols, where To obtain the initial data, represents the received data subcarrier, represents the CFR of the data subcarrier obtained by the proposed channel estimation network.

[0027] The signal demodulation model consists of a bidirectional long short-term memory network BiLSTM and a fully connected layer. It contains three layers of BiLSTM networks, each with 64 time steps and 64 hidden units. The last layer is a FC layer with 64 neurons, using the Sigmoid activation function for binary classification output. It can be expressed as: ,in, Represents the output after passing through the BiLSTM network.

[0028] Including: setting the threshold of the Sigmoid function to 0.5, if the output is greater than 0.5, the received symbol is considered to be 1, otherwise it is 0; the calculation formula of BiLSTM is: , , , , , ,in, , , ,and They are the input gate, forget gate, output gate and candidate state representing the current information. and denote the weight matrix and bias respectively, , , , , , , , Represents input The weight matrix is ​​used to calculate the impact on the input gate, forget gate, output gate, and current information candidate state respectively, and , , , Represent the bias terms of each, represents the hidden state at each time step, Indicates the unit status, and They are Sigmoid and hyperbolic tangent functions, Represents element-wise product.

[0029] Step 4: Set the hyperparameters of the CNN network and BiLSTM network.

[0030] Step 5: Use the data set to train the channel estimation model and signal demodulation model.

[0031] Step 5-1: Divide the data set into a training set and a test set in a ratio of 8:2.

[0032] Step 5-2: Use the Adam optimizer with a learning rate of 0.001 and cross entropy as the loss function to train the channel estimation model. The loss function is

[0033] , where y represents the true label and p represents the predicted value of the model.

[0034] Step 5-3: Use the trained channel estimation model to train the signal demodulation model.

[0035] Step 6: Calculate the accuracy curve, network loss curve, and training time. If the accuracy curve does not fit or the network loss curve does not converge, reset the hyperparameters and proceed to step 4. Otherwise, proceed to step 7.

[0036] Step 7: Input the test set into the deep learning-assisted OFDM underwater acoustic communication receiving model for online deployment. The modulation mode adopted by the OFDM underwater acoustic communication receiving model is QPSK. The OFDM underwater acoustic communication receiving model sends test signals under different signal-to-noise ratios, generates training samples in the measured underwater acoustic channel environment, inputs them into the trained underwater acoustic communication receiving model, and outputs the predicted signal.

[0037] Example 2 An OFDM underwater acoustic communication receiver based on deep learning includes the following steps: Step 1: The transmitter randomly generates a binary bit stream and maps the binary information into OFDM baseband symbols, then generates a baseband OFDM modulated signal through fast inverse Fourier transform, converts the modulated signal and inserts a cyclic prefix before transmitting it by the transmitter; the receiver receives the signal and generates a data set after preprocessing.

[0038] Step 2: Construct an underwater acoustic OFDM channel estimation model based on CNN.

[0039] The channel estimation network consists of two one-dimensional convolutional layers, a self-attention module and two one-dimensional deconvolutional layers cascaded in sequence; the filters of the first two one-dimensional convolutional layers are 32 and 64, the convolution kernels are 3, and the stride is 1; the filters of the last two one-dimensional deconvolutional layers are 32 and 64, the convolution kernels are 3, and the stride is 1; and a jump connection is added between the first one-dimensional convolutional layer and the first one-dimensional deconvolutional layer to reduce the possible feature loss during information propagation and the gradient disappearance during training. An attention structure is added after the one-dimensional convolutional layer to help model global features and dynamically estimate the channel.

[0040] The channel frequency response (CFR) of the pilot subcarrier is calculated using the known pilot symbols and the received pilot symbols. The real and imaginary parts of the pilot subcarrier are split and concatenated together as the input of the channel estimation network. After the underwater acoustic channel features are extracted using a one-dimensional convolutional layer, they are transmitted to the attention module, which captures the global features and then reconstructs the channel frequency response of the data subcarrier through a one-dimensional deconvolutional layer.

[0041] Step 3: Build a signal demodulation model based on BiLSTM.

[0042] First, the traditional demodulation method ZF algorithm is used for initial equalization, and then the received data subcarrier, the data after initial equalization, and the CFR of the data subcarrier obtained after step 4 are used as the input of BiLSTM for accurate demodulation. A three-layer BiLSTM network is used, and the number of hidden units of the three-layer BiLSTM is 64; then the fully connected layer is connected, and Sigmoid is used as the activation function.

[0043] Step 4: Set the hyperparameters required for the entire deep learning-based OFDM receiver.

[0044] Step 5: Use the data set obtained in step 1 to train the channel estimation model and signal demodulation model.

[0045] Both networks are trained using the Adam optimizer, with a learning rate of 0.001 and cross entropy (CE) as the loss function. The loss function can be expressed as:

[0046] Among them, y represents the true label (the value is 0 or 1), and p represents the predicted value of the model.

[0047] Step 6: Calculate the accuracy curve, network loss curve, and training time. If the accuracy curve does not fit or the loss curve does not converge, reset the hyperparameters and proceed to step 4. Otherwise, proceed to step 7.

[0048] Step 7: Use the acquired data to test the deep learning-assisted OFDM underwater acoustic communication receiver.

[0049] Based on the OFDM system, test signals are sent under different signal-to-noise ratios, and training samples are generated in the measured underwater acoustic channel environment. The samples are input into the trained signal detection model, and the predicted signals are output. The predicted signals are further compared with the sent test signals to test the performance of the trained model.

[0050] The modulation method used by the OFDM underwater acoustic communication model is QPSK. The OFDM underwater acoustic communication transmitter sends a test signal and generates a test sample after passing through the measured underwater acoustic channel. The test sample is input into the model for prediction.

[0051] Example 3 A deep learning-based OFDM underwater acoustic communication receiving method, an underwater acoustic orthogonal frequency division multiplexing (OFDM) receiver based on deep learning, thereby replacing the channel estimation, channel equalization and signal detection parts of the underwater acoustic OFDM receiving end, and recovering the original bit data sent in an end-to-end manner.

[0052] like Figure 1 As shown, the receiver comprises the following steps: Step 1: Build an OFDM system. The transmitter transmits the initial signal through modulation, pilot insertion, serial-to-parallel conversion, inverse fast Fourier transform, cyclic prefix addition, and parallel-to-serial conversion to generate the required data set. The receiver receives the OFDM signal after passing through the channel.

[0053] The received data is first preprocessed, which includes operations such as removing cyclic prefix, fast Fourier transform, and parallel-to-serial conversion.

[0054] Step 2: Construct an OFDM channel estimation network based on a convolutional neural network (CNN). Using the idea of ​​image super-resolution, the channel estimation network is constructed using 2 convolutional layers, 2 deconvolutional layers, a skip connection (SC) structure, and a self-attention structure.

[0055] The channel frequency response (CFR) characteristics of the pilot subcarrier can be calculated by the known pilot symbols sent and the received pilot symbols. The CFR calculation expression of the pilot subcarrier can be expressed as: ,in, is the known transmitted pilot symbol, is the received pilot symbol. Since the CFR of the pilot subcarrier is in complex form, the matrix obtained by separating the real part and the imaginary part and then concatenating them is selected as the input of the channel estimation network.

[0056] Two convolutional layers are used to extract channel features, the attention module models global features, and then the deconvolution layer is used to restore the CFR of the data subcarrier. ReLU is selected as the activation function for both the convolutional layer and the deconvolution layer, and its expression is: , using the output after the first convolutional layer directly as the input of the first deconvolutional layer can reduce the problem of feature loss that may occur during message propagation and gradient disappearance during network training.

[0057] The output after the convolution layer, attention layer, and deconvolution layer can be expressed as: ,in, is the final output CFR, represents the function of the nth layer of CNN, , , , Represents the weight matrix of each layer of CNN, and , , , Represents the bias matrix of each layer, and A represents the attention weight matrix, which is used to weight the features.

[0058] Step 3: Construct a signal demodulation network based on a bidirectional long short-term memory network (BiLSTM). In the signal demodulation network, the traditional channel equalization algorithm ZF algorithm is first performed on the input data to obtain the initial data.

[0059] The initial data obtained after the ZF algorithm can be expressed as: ,in, represents the received data subcarrier, represents the CFR of the data subcarrier obtained by the proposed channel estimation network. Then the sequence The signal is input to the signal demodulation part to further accurately estimate the transmitted data symbols. The signal demodulation part consists of a bidirectional long short-term memory (BiLSTM) network and a fully connected layer, which contains three layers of BiLSTM networks, each with 64 time steps and 64 hidden units, and finally a FC layer containing 64 neurons, using the Sigmoid activation function to output binary classification.

[0060] The output after Sigmoid processing can be expressed as: ,in, Represents the output after passing through the BiLSTM network. Set the threshold of the Sigmoid function to 0.5. If the output is greater than 0.5, the received symbol is considered to be "1", otherwise it is "0".

[0061] The calculation formula of BiLSTM is: , , , , , ,in, , , ,and They are the input gate, forget gate, output gate and candidate state representing the current information. , , , , , , , Represents input The weight matrix is ​​used to calculate the impact on the input gate, forget gate, output gate, and current information candidate state respectively, and , , , Represent the bias terms of each, represents the hidden state at each time step, Indicates the unit status, and denote the weight matrix and bias respectively, and They are Sigmoid and hyperbolic tangent functions, Represents element-wise product.

[0062] Step 5: Use the data set to train and test the channel estimation network and signal demodulation network. First, train the channel estimation subnet for 2000 iterations, and then use the trained channel estimation network to train the demodulation network for 5000 iterations. Both use the Adam optimizer, with a learning rate of 0.001 and a cross entropy loss function.

[0063] Step 6: Calculate the accuracy curve, network loss curve, and training time. If the accuracy curve does not fit or the loss function does not converge, reset the hyperparameters and proceed to step 5. If the accuracy curve fits and the loss function converges, proceed to step 7.

[0064] Step 7: Input the test set into the OFDM receiver model based on deep learning for online deployment. This model is compared with traditional technologies LS, MMSE, existing technologies FC-DNN, CNN, CE and LSTM.

[0065] Example 4 A deep learning-based OFDM underwater acoustic communication reception method is compared with traditional technologies LS, MMSE, existing technologies FC-DNN, CNN, CE and LSTM. The OFDM system parameters are shown in Table 1, and the experimental condition parameters are shown in Table 2.

[0066] Table 1: OFDM system parameters .

[0067] Table 2: Experimental conditions and parameters .

[0068] Based on the OFDM system parameters and channel NOF experimental conditions, the system bit error rate comparison is compared with LS, MMSE, FC-DNN, CNN, CE and LSTM, as shown in Figure 5. When the signal-to-noise ratio is <20dB, all the above deep learning-based methods have higher bit error rates than the traditional LS-ZF and MMSE-ZF methods. When the bit error rate is >20dB, the FC-DNN method composed of a fully connected layer has a higher bit error rate than the MMSE-ZF method. The method proposed in the present invention is superior to other methods when the signal-to-noise ratio is between 0-30dB.

[0069] Based on the OFDM system parameters and channel NCS experimental conditions, the system bit error rate comparison is compared with LS, MMSE, FC-DNN, CNN, CE and LSTM, as shown in Figure 6. The method proposed in the present invention is superior to other methods in the entire signal-to-noise ratio range. Similar to the NOF channel, when the signal-to-noise ratio is >20dB, the bit error rate of the traditional method is lower than that of other deep learning-based methods.

[0070] Based on the OFDM system parameters and channel SIM experimental conditions, the system bit error rate comparison is compared with LS, MMSE, FC-DNN, CNN, CE and LSTM, as shown in Figure 7. Compared with the first two channels, the traditional channel estimation and demodulation method has a higher bit error rate in the SIM channel. It is speculated that the channel created by SIM is more complex than the NCS and NOF channels. The method based on deep learning is not greatly affected by the bit error rate in more complex channels, and the difference is more obvious when the signal-to-noise ratio is large. From the above results, it can be seen that the OFDM receiver proposed in the present invention is superior to other deep learning methods, especially under long-distance communication conditions, and maintains robustness.

[0071] Based on the above experimental conditions, the system bit error rate comparison is shown in Figure 8 compared with LS, MMSE, FC-DNN, CNN, CE and LSTM. In the absence of CP, the OFDM receiver based on deep learning is better than the traditional OFDM receiver, especially when the signal-to-noise ratio is high, the advantage of the receiver based on deep learning is more obvious. The receiver proposed in the present invention has a lower bit error rate than other receivers based on deep learning. Even in the absence of CP, the receiver in this paper can effectively perform channel estimation and achieve low bit error rate demodulation.

[0072] In summary, the OFDM receiver based on improved CNN combined with BiLSTM proposed in the present invention can be applied in the field of underwater acoustic wireless communication. The present invention adopts an innovative logical design. First, a data set is generated through the OFDM system, and then a convolutional neural network is used to build a channel estimation model, and a signal demodulation model is built using a BiLSTM network; secondly, the data set is input into the neural network, the model is trained, and the network is optimized using the Adam optimizer and the cross entropy loss function, and finally compared with the LS, MMSE, FC-DNN, CNN, CE and LSTM models. The present invention effectively solves the problem that the traditional underwater acoustic OFDM system cannot adapt to the rapid changes of the underwater acoustic channel and has a high demodulation bit error rate. The simulation results show that under different channel conditions and with fewer cyclic prefixes, the performance of this method is better than that of LS, MMSE, FC-DNN, CNN, CE and LSTM models.

[0073] The above embodiments are only for illustrating the inventive concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A deep learning-based OFDM underwater acoustic communication receiving method, characterized in that: The following steps are involved: Step 1: Build an OFDM system and generate a data set; The initial OFDM signal transmitted by the transmitting end is subjected to channel preprocessing to generate the data set, and the receiving end receives the OFDM signal after passing through the channel; Step 2: Construct a channel estimation model based on convolutional neural network (CNN); the expression is obtained based on the input data: The channel frequency response CFR of the leading subcarrier is obtained, and the channel frequency response of the leading subcarrier is split as the input of the convolutional neural network CNN, where: is a known pilot symbol, is the received pilot symbol, is the channel frequency response of the pilot subcarrier; Step 3: Construct a signal demodulation model based on the long short-term memory network BiLSTM; use the traditional channel equalization algorithm to obtain the expression: The initial data of the sequence Input to the signal demodulation part to accurately estimate the transmitted data symbols, where To obtain the initial data, represents the received data subcarrier, represents the CFR of the data subcarrier obtained by the proposed channel estimation network; Step 4: Set the hyperparameters of the CNN network and BiLSTM network; Step 5: Use the data set to train the channel estimation model and signal demodulation model; Step 6: Count the accuracy curve, network loss curve and training time. If the accuracy curve does not fit or the network loss curve does not converge, reset the hyperparameters and proceed to step 4. Otherwise, proceed to step 7. Step 7: Input the test set into the deep learning-assisted OFDM underwater acoustic communication receiving model for online deployment.

2. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 1 is characterized in that: In step 1, the channel preprocessing includes modulation, pilot insertion, serial-to-parallel conversion, inverse fast Fourier transform, adding a cyclic prefix, parallel-to-serial conversion and fast Fourier transform.

3. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 1 is characterized in that: In step 2, the channel estimation model is composed of two one-dimensional convolutional layers, a self-attention module and two one-dimensional deconvolutional layers cascaded in sequence, the ReLU function is selected as the activation function, and a jump connection is set between the first one-dimensional convolutional layer and the first one-dimensional deconvolutional layer.

4. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 3 is characterized in that: In step 2, the input data includes known pilot symbols and received pilot symbols, and step 2 further includes: Step 2-1: Split the real part and the imaginary part of the pilot subcarrier and concatenate them together as the input of the channel estimation model; Step 2-2: After extracting the underwater acoustic channel features using two one-dimensional convolutional layers, the features are transferred to the self-attention module; Step 2-3: The self-attention module captures the global features, and then reconstructs the channel frequency response of the data subcarrier through two one-dimensional deconvolution layers, which is expressed as: ,in, represents the input of the neural network, is the final output CFR, represents the function of the nth layer of CNN, , , , Represents the weight matrix of each layer of CNN, and , , , Represents the bias matrix of each layer, and A represents the attention weight matrix.

5. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 1, characterized in that: In step 3, the signal demodulation model is composed of a bidirectional long short-term memory network BiLSTM and a fully connected layer, containing three layers of BiLSTM networks, each layer has 64 time steps and 64 hidden units, and finally a FC layer containing 64 neurons, using the Sigmoid activation function to output binary classification.

6. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 5 is characterized in that: In step 3, the output of the binary classification using the Sigmoid activation function is expressed as: ,in, Represents the output after passing through the BiLSTM network.

7. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 5 is characterized in that: In step 3, the steps of using the Sigmoid activation function to perform binary classification output include: The threshold of the Sigmoid function is set to 0.

5. If the output is greater than 0.5, the received symbol is considered to be 1, otherwise it is 0. The calculation formula of the forward BiLSTM is: , ， ， ， ， ,in, , , ,and They are the input gate, forget gate, output gate and candidate state representing the current information. , , , , , , , Represents input The weight matrix of , , , Represent the bias terms of each, represents the hidden state at each time step, Indicates the unit status, and They are Sigmoid and hyperbolic tangent functions, Represents element-wise product.

8. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 1, characterized in that: Step 5 includes: Step 5-1: Divide the data set into a training set and a test set in a ratio of 8:2; Step 5-2: Use the Adam optimizer with a learning rate of 0.001 and cross entropy as the loss function to train the channel estimation model. The loss function is: , where y represents the true label and p represents the predicted value of the model; Step 5-3: Use the trained channel estimation model to train the signal demodulation model.

9. The OFDM underwater acoustic communication receiving method based on deep learning according to claim 1, characterized in that: In step 7, the modulation mode adopted by the OFDM underwater acoustic communication model is QPSK, the OFDM underwater acoustic communication transmission model sends a test signal, and generates a test sample after passing through the measured underwater acoustic channel, and the test sample is input into the model for prediction.

Images