An underwater acoustic physical layer network coding intelligent coding mapping method

Abstract

A method for intelligent coding and mapping of underwater acoustic physical layer network coding is proposed, relating to the field of underwater acoustic communications. Combining the BP algorithm with deep learning, a superposition signal intelligent coding and mapping method (FBENet) with low coupling to channel information is proposed. The feature extraction module effectively extracts abstract features of the input information through a convolutional neural network (CNN). The bidirectional equalization module utilizes Bi-LSTM network neurons to equalize inter-carrier interference (ICI) in a time-frequency dual-extended channel, correcting the posterior probability distribution of subcarriers by simulating the sum-product message passing process of the BP algorithm. The probability output module outputs the probability distribution of the transmitting codeword combination for each subcarrier through a fully connected network for subsequent network coding. The network model's loss function uses a multi-classification cross-entropy loss function. During the training phase, as the model calculates parameter updates, the difference between the sample and the output decreases until it reaches a threshold, indicating successful training.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic communication, and in particular to an underwater acoustic physical layer network coding intelligent coding mapping method for receiving decoding mapping conditions of a physical layer network coding relay node in an underwater acoustic channel. Background Art

[0002] In 2006, researchers combined channel coding and network coding in the paper "Hot topic: Physical-layer network coding," proposing physical-layer network coding (PNC). This research encodes the information from two sources into an XOR sequence and details the coding mapping scheme for BPSK and QPSK modulation. It superimposes the information from different sources in space, then uses coding mapping to calculate it into an XOR sequence and broadcast it. Each node can then recover the data sent by other nodes using the network coding signal and local source information. In recent years, PNC technology has been widely studied for its potential to improve network throughput. It breaks the shackles of multi-hop communication that relies on point-to-point store-and-forward. In the typical application scenario of two-way relay channel (TWRC), its end-to-end communication rate can reach the theoretical upper limit, approximately the point-to-point communication rate, perfectly matching the demand for underwater acoustic network applications that urgently need to increase throughput. The implementation of PNC technology relies heavily on the coding and mapping of superimposed signals by relay nodes, a process that is extremely complex due to the dual-spread time-frequency channel in underwater environments. Chinese patent CN108880629 A discloses a PNC coding and mapping scheme designed for flat fading channels based on a bidirectional relay model. Chinese patent CN114257343 A discloses a linear coding method for constellation points. Chinese patent CN102983947 B proposes a collaborative communication method that combines space-time coding and physical layer network coding.

[0003] The methods used in the aforementioned public literature are based on PNC coding in terrestrial wireless network scenarios. However, in underwater networks, acoustic wave propagation speeds are only 1500 m / s, far slower than the propagation speed of electromagnetic waves. Furthermore, the influence of the time-frequency dual-spread channel causes information to exhibit severe multipath and Doppler effects during transmission. Signals inevitably experience propagation delays, resulting in multipath effects. Furthermore, underwater nodes are subject to dynamic movement due to factors such as ocean currents and surface fluctuations, causing Doppler effects on signal reception. Under the influence of the time-frequency dual-spread channel, signals are subject to severe ICI interference, making traditional PNC coding schemes unsuitable. Therefore, it is imperative to design a coding mapping method that can adapt to complex channels and exhibit strong generalization capabilities. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent coding mapping method for underwater acoustic physical layer network coding, that is, an intelligent coding mapping method designed through popular deep learning, in response to the limitations of the coding mapping scheme of the traditional factor graph of the physical layer network coding relay node in the time-frequency dual extension channel. The present invention improves the traditional algorithm into an intelligent coding mapping scheme, and through the powerful nonlinear fitting ability of deep learning and combined with the traditional BP algorithm, designs a feature extraction and bidirectional equalization network (Feature Extraction and Bidirectional Equalization Network, FBENet) to achieve low coupling with channel information, and has good effect. It can better replace the traditional BP algorithm and can better meet the needs of underwater networking technology communication.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An intelligent coding mapping method for underwater acoustic physical layer network coding comprises the following steps:

[0007] 1) The PNC system relay receives the superimposed signal and performs data preprocessing, splitting the superimposed signal vector and the estimated channel vector into real and imaginary parts. The data vector is then sent to the normalization layer and the concatenate layer for standard splicing into a two-dimensional matrix. The processed data is then transmitted to the feature extraction module.

[0008] 2) The feature extraction module uses a convolution kernel to perform a convolution operation on the input two-dimensional matrix data to extract abstract features of the input information, and splices the obtained specific features into a feature map. The feature information is input into the bidirectional equalization module through the residual block structure. The main path of the residual block structure contains three convolutional layer networks and an activation function ReLU. The shortcut path of the residual block structure has only one convolution layer. The results of the two paths are added and output through the activation function ReLU.

[0009] 3) The bidirectional equalization module uses a Bi-LSTM network to obtain the forward and backward correlation of the input sequence, equalize the inter-carrier interference (ICI) in the time-frequency dual-extended channel, and correct the posterior probability distribution of the subcarriers by simulating the sum-product message passing process of the BP algorithm. The resulting comprehensive output state matrix is ​​used as the output of the bidirectional equalization module and transferred to the probability output module for fully connected network mapping.

[0010] 4) The probability output module uses a fully connected network as the probability output layer and uses the Softmax activation function to convert the state matrix obtained after processing by the bidirectional balancing module into the joint probability distribution of the code word combination on each subcarrier, which is used for subsequent network coding. Specifically, each row of the state matrix, i.e., the state sequence of each subcarrier, is connected to the fully connected network, the TimeDistributed layer is used to encapsulate the fully connected network to ensure that each subcarrier sequence in the matrix passes through the fully connected layer with the same structure, and the independent probability distribution of each subcarrier is obtained; the Softmax function ensures that the output probability distribution is valid, and according to different modulation modes, the output layer estimates the probability distribution of the corresponding number of code word combinations;

[0011] 5) The loss function is designed, and the cross-entropy loss function of multi-classification is used; in the training stage, the network model calculates the derivative with respect to the variable parameter, updates through the back propagation algorithm, and the difference between the sample and the output gradually decreases until the output of the loss function converges to a certain threshold, and the deep learning training is completed.

[0012] In step 1), the splitting into real and imaginary parts is performed because the data entering the PNC system is a complex sequence, and both the superimposed signal vector and the available channel estimation vector are split into real and imaginary parts, the real and imaginary parts of the signal carrying independent information respectively; the data vector is sent to the normalization layer for standardization processing to eliminate the influence of signal amplitude difference on subsequent processing; the normalized real and imaginary parts are spliced into a two-dimensional matrix using the Concatenate layer, the features are effectively flattened and input into the feature extraction module; the row dimension of the transformed two-dimensional matrix is the number of subcarriers of each OFDM symbol, and the column dimension includes the real and imaginary parts of the superimposed signal and two channels.

[0013] In step 2), the feature extraction module is essentially a CNN network, which extracts abstract information contained in the input data, and the convolution operation of the convolution kernel and the input matrix can be represented by the following formula:

[0014]

[0015] where I j [i] is the i-th output of the j-th convolution kernel, j is the total number of convolution kernels, the maximum value of i is the width of the input matrix, X is the input data, W is the weight of the convolution kernel, b is the bias term of the convolution kernel, m and n represent the position information of the corresponding input data and convolution kernel matrix, respectively, and kw and kh represent the convolution operator height.

[0016] In order to extract richer feature information, the convolutional network uses a convolutional residual block architecture. The core design of the convolutional residual block structure is dual-path transmission: the main path contains three convolutional layer networks and the activation function ReLU, while the shortcut path has only one convolution layer. The results of the two paths are added and output through the activation function ReLU. The specific performance can be seen in the following formula:

[0017]

[0018] Among them, I represents the output feature map information, r represents the superposition signal, and The estimation of channel A, Represents the estimation of channel B, Residual CNN It is represented as a convolutional residual architecture, CNN is a convolutional neural network, and ReLU is the activation function in the entire feature extraction module;

[0019] After passing through the residual network, the system inputs the feature information into the bidirectional equalization module.

[0020] In step 3), the comprehensive output state matrix can be briefly described as follows: setting a bidirectional propagation Bi-LSTM network to capture the contextual correlation of the input sequence and output the state matrix; assuming that the current scene input is the feature map I and the output is the state matrix U, the Bi-LSTM layer can be expressed as follows:

[0021] [U[0],U[1],...,U[N-1]]=BiLSTM(I[0],I[1],...,I[N-1])

[0022] Where N represents the input time, which means the total number of subcarriers in this LSTM scenario; I[N] represents the characteristic sequence data of the Nth subcarrier input at the Nth time; U[N] is the output sequence data of the Nth subcarrier at the Nth time, and its length is related to the number of neurons in the LSTM. The Bi-LSTM network consists of LSTM networks in two directions and can be expressed as follows:

[0023]

[0024] in, represents forward feedback, represents backward feedback, U′[N] and U″[N] represent the outputs of forward and backward feedback respectively, and U[N]=[U′[N],U″[N]] represents the comprehensive output state matrix.

[0025] In step 4), assume that the Bi-LSTM network in the bidirectional balancing module has N unit There are neurons, the number of subcarriers is N, and its output size is N×2N unitThe state matrix U has row dimensions corresponding to each subcarrier and column dimensions corresponding to the state sequence on the subcarrier. A fully connected network is used as the probabilistic output layer. In the fully connected network, each neuron is connected to all neurons in the previous layer to form a fully connected network structure, which can be expressed as follows:

[0026] y=f(wU[n]+b)

[0027] Among them, U[n] is the data information of the nth row of the state matrix, w is the weight vector, b is the bias vector, y is the final output vector, and f(×) is the activation function; Softmax is selected as the activation function to convert the real vector into a probability distribution, where each conversion result is between 0 and 1, and the sum of all results is 1, which is equivalent to a normalization; given a K-dimensional real vector y=(y1,y2,...,y K ), the activation function Softmax can be defined as follows:

[0028]

[0029] Among them, Softmax(y) i Represents the probability of the i-th category in the output probability distribution, y i represents the i-th data in the real number vector y, e represents the base of the natural logarithm, and K represents the dimension of the vector.

[0030] In step 5), assume that there are M classes of targets to be predicted and there are only P samples in total. The label value of the i-th sample is set to The predicted value is Then the loss function is expressed as follows:

[0031]

[0032] Where, and Represent the jth data value of the predicted value sequence and label value sequence respectively.

[0033] Compared with the prior art, the present invention has the following outstanding technical effects and advantages:

[0034] For a three-node PNC relay system, this paper integrates traditional algorithms with deep learning to innovatively propose the FBENet model. This model effectively extracts input data features and converges quickly, achieving improved end-to-end transmission rates. Specifically, under BPSK modulation, its performance not only matches but even surpasses baseline methods under ideal channel response conditions. This paper also investigates intelligent coding and mapping methods for superimposed signals at the PNC end, which has constructive significance for improving underwater PNC systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a block diagram of the relay system of the intelligent coding mapping method described in the present invention.

[0036] Figure 2 This is the structural diagram of the FBENet model described in the present invention.

[0037] Figure 3 This is a flow chart of the superimposed signal and channel real and imaginary part division according to the present invention.

[0038] Figure 4 Schematic diagram of the convolution process of the convolution operator of the present invention.

[0039] Figure 5 This is a typical residual network structure diagram described in the present invention.

[0040] Figure 6 This is a diagram of the internal structure of the LSTM network neurons described in the present invention.

[0041] Figure 7 This is the Bi-LSTM network structure diagram described in the present invention.

[0042] Figure 8 Schematic diagram of the fully connected network in the probability output module of the present invention.

[0043] Figure 9 This is a fully connected network using Timedistributed encapsulation with BPSK as an example described in the present invention.

[0044] Figure 10 This is the loss curve of the FBENet model described in the present invention, taking BPSK as an example. In the figure, the horizontal axis is the number of iterations (epoch); the vertical axis is the value of the loss function (Loss).

[0045] Figure 11 This is the loss curve of the FBENet model described in the present invention, taking QPSK as an example. In the figure, the horizontal axis is the number of iterations (epoch); the vertical axis is the value of the loss function (Loss).

[0046] Figure 12 This is a performance comparison chart of the FBENet-based coding and mapping scheme and the BP algorithm-based coding and mapping scheme, taking BPSK as an example. In the chart, the horizontal axis is the signal-to-noise ratio (SNR) and the vertical axis is the bit error rate (BER).

[0047] Figure 13 This is a performance comparison chart of the FBENet-based code mapping scheme and the BP algorithm-based code mapping scheme, taking QPSK as an example. In the chart, the horizontal axis is the signal-to-noise ratio (SNR) and the vertical axis is the bit error rate (BER). DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0049] This paper improves the traditional algorithm into an intelligent coding mapping scheme: through the powerful nonlinear fitting ability of deep learning and combined with the traditional BP algorithm, a feature extraction and bidirectional equalization network FBENet is designed.

[0050] Overall part of the smart mapping encoding method:

[0051] like Figure 1 and Figure 2 As shown, the present invention is an improvement and adjustment based on the coding mapping scheme of the traditional PNC system, that is, an intelligent coding mapping scheme (FBENet) is introduced to solve the problem of strong coupling between the traditional factor graph scheme and channel information. Figure 1 This paper intuitively describes the location of problems encountered by the PNC system's relay end and the specific deployment of the intelligent mapping and coding scheme. After the superimposed signal is synchronized at the PNC relay receiver, OFDM demodulation is performed. The resulting superimposed signal and channel estimation matrix are fed into the FBENet model to output the joint posterior probability of codeword combinations. This probability information is then transferred to the CNC module for decoding and encoding, which outputs an XOR sequence. Finally, at the transmitter, channel coding, constellation mapping, OFDM modulation, and synchronization signal insertion are performed before the XOR information is broadcast to the terminal receiver. Figure 2 The design process framework of the entire FBENet network is briefly explained. Estimation of channel B The channel estimation matrix (i.e., obtained through the least squares (LS) algorithm) is then concatenated into a two-dimensional matrix using the Batch Normalization and Concatenate layers. The processed data is then transmitted to the feature extraction module. The feature extraction module uses a residual network to extract feature map information, employing a CNN (Conv1D) network and the ReLU function. The feature map is then fed into the bidirectional equalization module. The bidirectional equalization module, based on a Bi-LSTM network, simulates the sum-product message passing process of the BP algorithm to correct the posterior probability distribution of the subcarriers, obtaining a comprehensive output state matrix as the output for transfer to the probability output module. The probability output module encapsulates the fully connected network (Dense) using the TimeDistributed layer and uses the Softmax function to output the joint posterior probability of the codeword combination.

[0052] The combined part of the superimposed signal and the real and imaginary parts of the channel:

[0053] like Figure 3The figure shows how the PNC system relay inputs the superimposed signal and estimated channel into the network. The superimposed signal received by the PNC system relay enters the FEBNet network as a complex sequence. Since this data is input to the network as a complex sequence, the real and imaginary parts of the superimposed signal vector and the estimated channel vector must be separated before feature extraction. This data is then normalized using a batch normalization layer and then concatenated into a two-dimensional matrix using a concatenate layer before being fed into the feature extraction module. The resulting two-dimensional matrix has row dimensions corresponding to the number of subcarriers per OFDM symbol, and column dimensions containing the superimposed signal and the real and imaginary parts of the two channels.

[0054] Feature extraction module:

[0055] The feature extraction module of the present invention is designed to effectively extract abstract features from the input information. In the BP algorithm, the essence of calculating the initial confidence LLR is actually to extract the distance between each constellation point on the constellation diagram corresponding to the superimposed signal. However, the superimposed signal and the estimated channel transfer function carry abstract information. The traditional factor graph scheme can only extract the linear features therein. The present invention uses CNN to extract abstract features and has better generalization ability. The CNN network can extract the abstract information contained in the input data. The convolution operation of its convolution kernel and the input matrix can be expressed by the following formula:

[0056]

[0057] Among them, I j [i] is the i-th output of the j-th convolution kernel, j is the total number of convolution kernels, the maximum value of i is the width of the input matrix, X is the input data, W is the weight of the convolution kernel, b is the bias term of the convolution kernel, m and n represent the position information of the corresponding input data and convolution kernel matrix, respectively, and kw and kh represent the height of the convolution operator. The convolution process is similar to filtering. Each convolution kernel can extract specific features of the matrix, and finally the acquired specific features are spliced ​​into a feature map. In order to extract richer feature information, the residual network adopts the convolution residual block architecture. The convolution residual block can fuse networks of different depths, which means that when the network extracts data information features, it can not only retain shallow network information but also learn new features from deeper network structures.

[0058] The two-dimensional matrix obtained by concatenating the concatenated layers will first pass through the convolution layer. Figure 4 Briefly demonstrate the process of extracting input sequence features by the convolution operator in FBENet. Considering that the information input is a digital signal matrix, for the sake of simplicity, Figure 4In this example, a two-dimensional matrix is ​​represented by a one-dimensional vector. Each element in the vector contains the superimposed signal and channel information for each subcarrier. The convolution kernel is also represented by a vector. The convolution process is essentially similar to filtering, and each convolution kernel can extract specific features. The figure assumes a convolution operator height of kh = 3. When kh > 1, the CNN's sliding convolution mechanism is similar to the interaction between subcarriers, effectively extracting the correlation between adjacent subcarriers. The relationship between the size of kh and the ICI interference depth d is:

[0059] kh=2*d+1

[0060] In the formula, * represents convolution; each convolution operator obtains subcarrier features after convolution with the input vector sequence, and a single operator convolution results in a feature sequence. After multiple convolution operators complete convolution, the resulting feature sequences are spliced ​​into feature map I, which is used by the subsequent bidirectional equalization module to balance the information between subcarriers. To achieve better performance, the convolutional network structure uses convolutional residual blocks for feature extraction. Convolutional residual blocks fuse networks of different depths, not only preserving the original information of the shallow network but also obtaining deeper information through multiple convolutions.

[0061] The core design of a typical convolutional residual block is to transmit data through two paths: a main path consisting of multiple convolutional layers and activation functions, and a shortcut path using skip connections that directly superimposes the input with the output of the main path. These two paths are combined at the end through an addition operation and then activated using a ReLU function to form the output of the residual block. This arrangement superimposes the deep features learned in the main path with the shallow features in the shortcut, adjusting the information distribution of the feature map input. It performs convolution operations to adjust the input dimension to match the output of the main path. The skip path generally does not include an activation function because activation functions introduce nonlinearities that can distort information and cause vanishing gradients. This skip connection strategy preserves more input information or shallow features, reducing the loss of original information during the convolution process while the network learns deep, abstract features, and mitigates the impact of noise, thereby enhancing the robustness of the model. Figure 5 The residual network designed by the present invention includes a main path and a shortcut path. The main path contains three convolutional layers and an activation function, while the shortcut path has only one convolutional layer. The results of the two paths are added and output through the activation function. The specific process is as follows:

[0062]

[0063] In the formula, I represents the output feature map information, r represents the superposition signal, and The estimation of channel A, Represents the estimation of channel B, Residual CNNThe CNN is a convolutional neural network, and the ReLU is an activation function in the entire feature extraction module.

[0064] The bidirectional equalization module:

[0065] After the information in the FBENet network proposed by the application is preprocessed and extracted, it is transferred to the bidirectional equalization module for processing.

[0066] h t =f(h t-1 ,x t )

[0067] Where h t represents the output of the current time step, h t-1 represents the output of the previous time step, x t represents the input of the current time step, f(.) is a function representing the RNN model, which is a nonlinear function, and different nonlinear functions correspond to different RNN models. The above formula recursively updates the output of each time step, and this recursive structure enables the RNN to capture the dependency relationship in the sequence data. Especially under the time-frequency double expansion channel, the introduction of the RNN model helps to process the correlation and ICI interference between subcarriers, and achieves better equalization effect.

[0068] There are many variant structures of RNN network, among which the Long Short Term Memory (LSTM) has excellent performance in processing time sequence correlation scenes within a limited range, so the LSTM network is introduced to replace the traditional algorithm in the application. Under the interference of ICI, a subcarrier may be interfered by all other subcarriers, but in order to facilitate modeling, it is considered that a subcarrier is only interfered by other subcarriers within a certain range, and the rest of the interference is considered as part of the environmental noise. Different subcarriers are regarded as different time nodes input, and for this time sequence correlation scene within a limited range, the LSTM network is introduced.

[0069] The LSTM neuron introduces a memory unit and a gating mechanism to capture long-term dependencies in sequence data, Figure 6The internal structure of LSTM neurons is shown, and the calculation process is as follows:

[0070] f t =σ(W f ·[h t-1 ,x t ]+b f )(Forget Gate)

[0071] i t =σ(W i ·[h t-1 ,x t ]+b i )(Input gate)

[0072] (Candidate memory unit)

[0073] (Update memory unit)

[0074] o t =σ(W o ·[h t-1 ,x t ]+b o )(Output gate)

[0075] h t =o t Tanh(C t )(Update hidden state)

[0076] Among them, f t is the output of the forget gate at the current time step, W f is the weight of the forget gate, b f is the bias of the forget gate; i t is the output of the input gate at the current time step, W i is the weight of the input gate, b i is the bias of the input gate; is the candidate memory unit output at the current time step, W c is the weight of the memory unit, b c is the bias of the memory unit; C t is the memory unit output of the current time step, C t-1 is the output of the memory unit at the previous time step, o t is the output of the output gate, h t is the hidden state of the current time step; h t-1 Represents the hidden state of the previous time step; σ represents the activation function Sigmoid, and its mathematical expression is as follows:

[0077]

[0078] The activation function Sigmoid maps the input value between (0, 1), because it has a smooth derivative, so it is easier to optimize gradient descent in the back propagation process, and accelerate convergence; the Tanh function also acts as an activation function in the LSTM neuron structure, and its mathematical expression is as follows:

[0079]

[0080] The Tanh function maps the input value between (-1, 1), because its output is centered at zero, so it helps to alleviate the problem of gradient disappearance.

[0081] The gating mechanism of LSTM can be combined with BP algorithm, which can successfully reproduce the traditional algorithm to deep learning. The following specifically introduces the key points of the combination of LSTM network and traditional algorithm in the application.

[0082] Forget gate: decide to keep or remove which information in the memory cell at the previous moment, can eliminate irrelevant information. When the BP algorithm transmits messages, the self-information of the subcarrier and the correlation of the adjacent carriers decrease with the distance, indicating that the current subcarrier is less disturbed by other distant subcarriers. In the original traditional BP algorithm, it is determined that the subcarriers with a long distance will not interfere with the subcarriers, and should be forgotten, and the information of the subcarriers with a short distance should be retained. With the passage of time, the information of the subcarriers changes, which corresponds to the function of the LSTM forget gate.

[0083] Input gate: decide how much component of the input information at the current moment will be recorded into the memory unit. In the BP algorithm, the observation factor needs to integrate the information of adjacent nodes during message transmission, and different original information inputs have different weights on message transmission, which corresponds to the input gate mode of the LSTM neuron.

[0084] Memory cell: the core element in LSTM, whose state update depends on the joint action of the forget gate and the input gate. The forget gate is responsible for removing part of the information in the original memory cell, and the input gate determines how much new information component is added. The update of the memory cell not only integrates new and old information, but also realizes the timely forgetting of irrelevant information over a long period of time. The BP algorithm integrates its own confidence and the information transmitted by adjacent carriers, which corresponds to the deep information processing capability of the memory cell in LSTM.

[0085] Output gate: control the information transmission flow of the memory cell to the hidden state at the current moment.

[0086] Hidden state: The primary output of the LSTM at each moment captures the model's current state and can be used in subsequent tasks. The probability distribution calculated during the message passing process of the BP algorithm is actually an intermediate state quantity, corresponding to the hidden state of the LSTM neuron. This intermediate state quantity requires further mapping processing using a fully connected network to be suitable for subsequent decoding steps.

[0087] In the above description of the key LSTM architecture, information transmission within a complete LSTM network is unidirectional. In a time-frequency dual-extended channel, subcarriers are subject to interference from both adjacent subcarriers. Therefore, an LSTM network with only unidirectional feedforward memory cannot fully replace the traditional BP algorithm. Figure 7 This is a model framework diagram of a Bi-LSTM network. This invention uses a Bi-LSTM network to capture the contextual relevance of the input sequence, i.e., an LSTM network with bidirectional transmission capabilities. In the current scenario, assuming the input is a feature map I and the output is a state matrix U, the Bi-LSTM layer can be represented as follows:

[0088] [U[0],U[1],...,U[N-1]]=BiLSTM(I[0],I[1],...,I[N-1])

[0089] Where N represents the input time, which in this LSTM scenario means the total number of subcarriers; I[N] represents the characteristic sequence data of the Nth subcarrier input at the Nth time; U[N] is the output sequence data of the Nth subcarrier at the Nth time, and its length is related to the number of neurons in the LSTM. The Bi-LSTM network consists of two LSTM networks in the forward and backward directions and can be expressed as follows:

[0090]

[0091] in, represents forward feedback, represents backward feedback, U′[N] and U″[N] represent the outputs of the forward and backward feedback, respectively, and U[N] = [U′[N], U″[N]] represents the combined output state matrix. Each bidirectional propagation Bi-LSTM layer corresponds to one iteration of the traditional BP algorithm.

[0092] After obtaining the state matrix U[N]=[U′[N],U″[N]], it is transferred to the probability output module as the output of the bidirectional balancing module to map the fully connected network into a probability distribution.

[0093] Probability output module:

[0094] The purpose of the probability output module in the present invention is to transform the state matrix after equalization into the real codeword combination joint probability distribution. Assume that the Bi-LSTM network in the above bidirectional equalization module has a total of N unit There are neurons and the number of subcarriers is N, so its output size is N×2N unit The state matrix U has row dimensions corresponding to each subcarrier and column dimensions corresponding to the state sequence on the subcarrier. The purpose of designing a deep learning model is to achieve joint probability distribution, so a fully connected network is used as the probability output layer. Figure 8 Showing the fully connected network model, each neuron is associated with all neurons in the previous layer, forming a fully connected network structure. This is also the origin of the fully connected network. Its model can be expressed by the following formula:

[0095] y=f(wU[n]+b)

[0096] Among them, U[n] is the data information of the nth row of the state matrix, w is the weight vector, b is the bias vector, y is the final output vector, and f(×) is the activation function. Softmax is selected as the activation function, which can convert the real vector into a probability distribution, where each conversion result is between 0 and 1, and the sum of all results is 1, ensuring the normalization of the probability. Given a K-dimensional real vector y=(y1,y2,...,y K ), the activation function Softmax can be defined as follows:

[0097]

[0098] Among them, Softmax(y) i Represents the probability of the i-th category in the output probability distribution. i represents the i-th data in the real number vector y, e represents the base of the natural logarithm, and K represents the dimension of the vector.

[0099] Since the fully connected network only accepts vectors as input, and the output of the bidirectional feedback equalization module is a matrix, it is necessary to connect each row of the state matrix (that is, the state sequence of each subcarrier) to the fully connected network. The TimeDistributed layer is used to encapsulate the fully connected network to ensure that each subcarrier sequence in the matrix passes through the fully connected layer with the same structure. Depending on the modulation method, the data at each frequency point at the transmitter will theoretically have a specific number of combinations. For the PNC relay end, if it is BPSK, theoretically the data transmitted from the terminal node has two state quantities. When superimposed at the relay end, there will be 4 situations, and similarly, there will be 16 situations with QPSK. Figure 9 It is a fully connected network model diagram designed for BPSK modulation.

[0100] With the assistance of the probability output module, the state transition of the Bi-LSTM layer output is effectively and successfully output as the joint probability distribution of each subcarrier codeword combination, completing the core output part of the entire FBENet network and replacing the original traditional BP algorithm.

[0101] Loss function design:

[0102] In the overall network construction, a loss function needs to be designed. In this invention, a multi-classification cross entropy function is selected. Assume that there are M classes of targets to be predicted and there are only P samples. The label value of the i-th sample is set to The predicted value is Then the loss function of the FBENet network can be expressed as follows:

[0103]

[0104] Where, and Represent the jth data value of the predicted value sequence and label value sequence respectively.

[0105] During the training phase, the network model calculates derivatives with respect to variable parameters and updates them through the back-propagation algorithm. The updates gradually reduce the difference between the sample and the learned output until the output of the loss function converges to a certain threshold. Figure 10 and 11 The simulation curves for BPSK and QPSK show that the loss curves gradually converge with increasing epochs, indicating successful neural network training. The loss values ​​for the BPSK and QPSK modulation modes in the figure stabilize after 25 and 40 epochs of training, respectively, indicating that after these epochs, both the training loss and validation loss of the model have dropped to low levels, and the model has converged and stabilized.

[0106] Figure 12 and 13 The following figure shows the bit error rate comparison curve of FEBNet's intelligent coding mapping method and the baseline method. The least squares LS algorithm in the figure is a channel estimation algorithm. For OFDM systems, the LS channel estimation matrix can be expressed as follows:

[0107]

[0108] Where H LS represents the LS channel estimation matrix, Y represents the signal received by the receiver, X represents the signal sent by the transmitter, the range of k is {0, 1, 2, ..., N-1}, and N represents the number of subcarriers.

[0109] In addition, the channel transfer function with different estimation accuracy is obtained by adding noise to the channel matrix. Here, the channel error rate (CER) is used to quantify the channel estimation error:

[0110]

[0111] Among them, H μ is the actual frequency domain channel transmission matrix, Θ μ is the channel estimation error (which can be understood as H μ The average power of the additive noise on 10 is a logarithmic function with base 10, and μ∈{A,B} indicates whether the source of the information is the terminal node A or B of the PNC system. For the above formula, a larger CER value indicates a more accurate channel estimation.

[0112] Perfect H refers to a known, perfect channel situation. This means that channel estimation is no longer necessary and the BP algorithm can be used directly. As can be seen from the figure, while the LS algorithm has low computational complexity, its channel estimation is very limited, resulting in the worst performance. The CER curve improves as the CER value increases, gradually approaching the perfect channel situation.

[0113] Experiments show that the proposed relay intelligent code mapping method, FBENet, can approach or even exceed the optimal performance of the baseline method in bit error rate curves, regardless of BPSK or QPSK modulation. This demonstrates that the intelligent code mapping method can match the performance of the known perfect channel matrix, Perfect H, without requiring precise channel information.

[0114] The PNC system of the present invention uses Orthogonal Frequency Division Multiplexing (OFDM) modulation for underwater transmission. Combining the Markov model-based Belief Propagation (BP) algorithm with deep learning, the present invention proposes a superposition signal intelligent coding and mapping method with low coupling to channel information—the Feature Extraction and Bidirectional Equalization Network (FBENet), which replaces the channel estimation and signal detection modules in traditional PNC relay receivers. FBENet can be divided into three modules: the feature extraction module uses convolutional neural networks (CNNs) to effectively extract abstract features of input information; the bidirectional equalization module utilizes bidirectional long short-term memory (Bi-LSTM) neurons to equalize inter-carrier interference (ICI) in dual-frequency and time-extended channels. This module corrects the posterior probability distribution of subcarriers by simulating the sum-product message passing process of the backpropagation algorithm; the probability output module outputs the probability distribution of the transmitting codeword combination for each subcarrier through a fully connected network for subsequent network coding. The network model's loss function uses a multi-class cross-entropy loss function. During the training phase, the model calculates parameter updates, and the difference between the sample and the output decreases until it reaches a threshold, indicating successful training.

[0115] The above embodiments are only preferred embodiments of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An intelligent coding mapping method for underwater acoustic physical layer network coding, characterized in that The following steps are involved: 1) The PNC system relay receives the superimposed signal and performs data preprocessing, splitting the superimposed signal vector and the estimated channel vector into real and imaginary parts. The data vector is then sent to the normalization layer and the concatenate layer to form a two-dimensional matrix. The processed data is then transmitted to the feature extraction module. 2) The feature extraction module uses the convolution kernel to perform a convolution operation on the input two-dimensional matrix data to extract the abstract features of the input information, splices the obtained features into a feature map, and inputs the feature information into the bidirectional equalization module through the convolution residual network; 3) The bidirectional equalization module uses a Bi-LSTM network to obtain the forward and backward correlation of the input sequence, equalize the inter-carrier interference (ICI) in the time-frequency dual-extended channel, and correct the posterior probability distribution of the subcarriers by simulating the sum-product message passing process of the BP algorithm. The resulting comprehensive output state matrix is ​​used as the output of the bidirectional equalization module and transferred to the probability output module for fully connected network mapping. 4) The probability output module uses a fully connected network as the probability output layer and uses the Softmax activation function to convert the state matrix obtained after processing by the bidirectional equalization module into the joint probability distribution of codeword combinations on each subcarrier for subsequent network coding. Specifically, each row of the state matrix, that is, the state sequence of each subcarrier, is connected to the fully connected network. The fully connected network is encapsulated using the TimeDistributed layer to ensure that each subcarrier sequence in the matrix passes through the same fully connected layer structure, resulting in an independent probability distribution for each subcarrier. The Softmax function ensures that the output probability distribution is valid. Depending on the modulation method, the output layer estimates the probability distribution of the corresponding number of codeword combinations. 5) Design a loss function, using a multi-class cross-entropy loss function; during the training phase, the network model calculates the derivative with respect to the variable parameters and updates it through the back-propagation algorithm. The difference between the sample and the output gradually decreases until the output of the loss function converges to a certain threshold, completing the deep learning training.

2. The intelligent coding mapping method for underwater acoustic physical layer network coding according to claim 1, characterized in that In step 1), the data is split into real and imaginary parts. Since the data entering the PNC system is a complex sequence, the superimposed signal vector and the available channel estimation vector are split into real and imaginary parts. The real and imaginary parts of the signal carry independent information respectively. The data vector is sent to the normalization layer for normalization to eliminate the impact of signal amplitude differences on subsequent processing. The normalized real and imaginary parts are concatenated into a two-dimensional matrix using the Concatenate layer. The features are effectively flattened and input into the feature extraction module. The row dimension of the converted two-dimensional matrix is ​​the number of subcarriers per OFDM symbol, and the column dimension contains the superimposed signal and the real and imaginary parts of the two channels.

3. The intelligent coding mapping method for underwater acoustic physical layer network coding according to claim 1, characterized in that In step 2), the feature extraction module structure is essentially a CNN network. The CNN network extracts the abstract information contained in the input data. The convolution operation between the convolution kernel and the input matrix is ​​expressed as follows: Among them, I j [i] is the i-th output of the j-th convolution kernel, j is the total number of convolution kernels, the maximum value of i is the width of the input matrix, X is the input data, W is the weight of the convolution kernel, b is the bias term of the convolution kernel, m and n represent the position information of the corresponding input data and convolution kernel matrix respectively, and kw and kh represent the height of the convolution operator.

4. The intelligent coding mapping method for underwater acoustic physical layer network coding according to claim 1, characterized in that In step 2), in order to extract richer feature information, the convolutional residual network uses a convolutional residual block structure. The core design of the convolutional residual block structure is dual-path transmission: the main path contains three convolutional layer networks and the activation function ReLU, while the shortcut path has only one convolution layer. The results of the two paths are added and output through the activation function ReLU, which is specifically expressed as follows: Among them, I represents the output feature map information, r represents the superposition signal, represents the estimation of channel A, Represents the estimation of channel B, Residual CNN It is represented as a convolutional residual architecture, CNN is a convolutional neural network, and ReLU is the activation function in the entire feature extraction module.

5. The intelligent coding mapping method for underwater acoustic physical layer network coding according to claim 1, characterized in that In step 3), the comprehensive output state matrix is ​​obtained. Specifically, a bidirectional Bi-LSTM network is set to capture the contextual correlation of the input sequence and output the state matrix. Assuming that the input of the current scene is the feature map I and the output is the state matrix U, the Bi-LSTM layer is represented as follows: [U[0],U[1],...,U[N-1]]=BiLSTM(I[0],I[1],...,I[N-1]) Where N represents the input time, which means the total number of subcarriers in the LSTM scenario; I[N] represents the characteristic sequence data of the Nth subcarrier input at the Nth time; U[N] is the output sequence data of the Nth subcarrier at the Nth time, and its length is related to the number of neurons in the LSTM. The Bi-LSTM network consists of two LSTM networks in two directions, as shown below: in, represents forward feedback, represents backward feedback, U′[N] and U″[N] represent the outputs of forward and backward feedback respectively, and U[N]=[U′[N],U″[N]] represents the comprehensive output state matrix.

6. The intelligent coding mapping method for underwater acoustic physical layer network coding according to claim 1, characterized in that In step 4), assume that the Bi-LSTM network in the bidirectional balancing module has N unit There are neurons, the number of subcarriers is N, and its output size is N×2N unit The state matrix U has row dimensions corresponding to each subcarrier and column dimensions corresponding to the state sequence on the subcarrier. A fully connected network is used as the probabilistic output layer. In the fully connected network, each neuron is connected to all neurons in the previous layer, forming a fully connected network structure as follows: y=f(wU[n]+b) Among them, U[n] is the data information of the nth row of the state matrix, w is the weight vector, b is the bias vector, y is the final output vector, and f(·) is the activation function; Softmax is selected as the activation function to convert the real vector into a probability distribution, where each conversion result is between 0 and 1, and the sum of all results is 1, which is equivalent to a normalization; given a K-dimensional real vector y = (y1, y2, ..., y K ), the activation function Softmax is defined as follows: Among them, Softmax(y) i Represents the probability of the i-th category in the output probability distribution, y i Represents the i-th data in the real vector y; e represents the base of the natural logarithm, and K represents the dimension of the vector.

7. The intelligent coding mapping method for underwater acoustic physical layer network coding according to claim 1, characterized in that In step 5), the loss function is designed, assuming that there are M classes of targets to be predicted and there are only P samples in total, and the label value of the i-th sample is set to The predicted value is Then the loss function is expressed as follows: Where, and Represent the jth data value of the predicted value sequence and label value sequence respectively.

Images