An Underwater Acoustic Channel Prediction Method Based on Cross-Frequency Domain Grouping and Deep Learning

Abstract

An underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning, which belongs to the field of wireless communication. The present invention solves the problems that the existing single-tap and single-frequency point prediction methods cannot utilize the complex time-frequency relationship of the channel, and the existing deep learning methods have high computational complexity. The frequency point grouping method based on the cross-frequency correlation matrix proposed by the present invention can avoid the problems of overly complex model and excessive parameters to be trained caused by too many input features of the deep learning predictor model, resulting in high computational complexity. Moreover, the CFDG-DL channel predictor combining the FC layer and the LSTM layer proposed by the present invention can utilize the complex time-frequency relationship of the frequency domain channel and improve the performance of channel prediction. The method of the present invention can be applied to underwater acoustic channel prediction.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication, and particularly relates to an underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning. Background Technique

[0002] Underwater acoustic communication, as the main means of underwater information transmission, has been widely applied in the fields of marine scientific research, commercial development, environmental protection, military operations, etc. For a long time, the underwater acoustic channel is complex and changeable, with characteristics such as limited bandwidth, multipath transmission, and high propagation delay, which all limit the performance of underwater acoustic communication. In order to overcome the rapid change of the channel and make better use of the channel state information, channel prediction technology has become an important topic in underwater acoustic communication. Underwater acoustic channel prediction is crucial for resource allocation, multi-user communication, and network structure optimization.

[0003] In previous studies, underwater acoustic channel prediction was usually carried out in the time domain. Researchers would assume that it has a sparse structure and only predict the main paths of the channel. However, there are two problems with such an approach: First, the assumption of channel sparsity does not always hold. When the assumption fails, the performance of time-domain channel prediction will decline, and the computational complexity will also increase. Second, the underwater acoustic channel has complex time-frequency correlation. Only performing single-tap prediction on the time-domain channel will ignore this part of information and cannot further reduce the channel prediction error. Frequency-domain channel prediction can overcome the influence of whether the channel is sparse and better adapt to different channel structures. At the same time, it is found that there is cross-correlation between frequency points, and using this can further improve the performance of the channel prediction algorithm. However, the single-frequency point prediction method also has the problem of being unable to utilize the complex time-frequency relationship of the channel.

[0004] Commonly used channel prediction methods can be divided into two categories: linear and nonlinear. Linear algorithms include recursive least squares, least mean square error, exponential smoothing, and Kalman filtering algorithms. Nonlinear algorithms are mainly based on kernel-based algorithms, including kernel adaptive and support vector regression. However, these algorithms are applicable to time series prediction composed of single variables and cannot adapt to the complex data relationships in multi-variable joint prediction. In recent years, the development of deep learning algorithms has provided new ideas for channel prediction. Deep learning models represented by long short-term memory networks (LSTM) have superior performance in processing time series with long-term dependencies, and their powerful data fitting ability can capture the complex time-frequency relationship of the underwater acoustic channel.

[0005] There have been studies attempting to use deep learning networks to predict the frequency-domain channel. Researchers divided all subcarriers into several clusters, reducing the features input to the model. However, this method reduces the frequency resolution of the predicted channel to some extent. If a more detailed frequency-domain channel is needed for adaptive modulation or precoding, then it is necessary to consider predicting all frequency points. However, directly using all frequency points as the input to the predictor model will result in an overly complex predictor model with a high computational complexity.

[0006] In summary, the existing single-tap and single-frequency-point prediction methods still have the problem of being unable to utilize the complex time-frequency relationship of the channel, and the existing deep learning methods still have the problem of high computational complexity. Therefore, it is very necessary to propose a new underwater acoustic channel prediction method to solve the above problems. Summary of the Invention

[0007] The object of the present invention is to solve the problems that the existing single-tap and single-frequency-point prediction methods cannot utilize the complex time-frequency relationship of the channel and the existing deep learning methods have high computational complexity, and a method for predicting underwater acoustic channels based on cross-frequency-domain grouping and deep learning is proposed.

[0008] The technical solution adopted by the present invention to solve the above technical problems is as follows:

[0009] A method for predicting underwater acoustic channels based on cross-frequency-domain grouping and deep learning, the method specifically includes the following steps:

[0010] Step 1. At the nth sampling time point t n , use an underwater acoustic communication machine to receive the time-domain passband signal, and then obtain the true value of the frequency-domain channel at the nth sampling time point t n . After discretizing the obtained frequency-domain channel, record the frequency-domain response value of the kth frequency point at the nth sampling time point t n as H(kΔf,t n ), where Δf is the frequency-domain sampling period, k = 1, 2,..., K, and K is the total number of frequency points;

[0011] Construct a cross-frequency-domain coherence matrix of the frequency-domain channel according to the frequency-domain response values of each frequency point at the nth sampling time point and the frequency-domain response values of each frequency point at the historical n - P sampling time points;

[0012] Step 2. After grouping the K frequency points according to the constructed cross-frequency-domain coherence matrix, divide the K frequency points into L groups;

[0013] Step 3. Build a CFDG-DL channel predictor for each group of frequency points;

[0014] Step 4: Input the frequency-domain response values corresponding to each group of frequency points into the corresponding CFDG-DL channel predictor to obtain the channel prediction result at time (n + 1)T. g The channel prediction result at time (n + 1)T.

[0015] Further, the t n = nT g , and T g represents the time interval between adjacent sampling time points.

[0016] Further, construct a cross-frequency-domain coherence matrix of the frequency-domain channel according to the frequency-domain response values of each frequency point at the nth sampling time point and the frequency-domain response values of each frequency point at the previous n - P sampling time points; the element ρ(g, k) in the gth row and kth column of the cross-frequency-domain coherence matrix is:

[0017]

[0018] where ρ(g, k) represents the coherence between the gth frequency point and the kth frequency point, H(kΔf, t i ) is the frequency-domain response value of the kth frequency point at the ith sampling time point t i , H(gΔf, t i ) is the frequency-domain response value of the gth frequency point at the ith sampling time point t i , g = 1, 2,..., K, H * (gΔf, t i ) represents the conjugate transpose of H(gΔf, t i ), P represents the total number of sampling time points used when constructing the cross-frequency-domain coherence matrix of the frequency-domain channel, and |·| represents taking the absolute value.

[0019] Further, in Step 2, group the K frequency points according to the constructed cross-frequency-domain coherence matrix, and the grouping criterion is:

[0020] Criterion 1): The frequency points within the same group are adjacent in order.

[0021] Criterion 2): The number of frequency points in each group is less than the frequency point threshold.

[0022] Criterion 3): The coherence between each pair of frequency points within each group is greater than the set coherence threshold.

[0023] Group the K frequency points according to Criterion 1), Criterion 2), and Criterion 3) to obtain the frequency point grouping result.

[0024] Further, the CFDG-DL channel predictor includes an input layer, a fully connected layer, an LSTM layer, and an output layer.

[0025] Further, the specific process of the fourth step is as follows:

[0026] For the l-th group of frequency points, the frequency-domain response value at the n-th sampling time point of the l-th group of frequency points and the frequency-domain response values at the previous n - P' sampling time points are used as the input to the CFDG-DL channel predictor corresponding to the l-th group of frequency points, and the input to the CFDG-DL channel predictor corresponding to the l-th group of frequency points is denoted as c l , c l After being passed from the input layer to the FC layer, the output of the FC layer is:

[0027]

[0028] Among them, represents the output of the FC layer, b represents all the parameters of the FC layer, and f FC represents the operation process of the FC layer;

[0029] After extracting the correlation between frequency points in the FC layer, the extracted frequency-point correlation is input into the LSTM layer, and the operation process of the LSTM layer is defined as follows:

[0030]

[0031] Among them, f LSTM is the operation process of the LSTM layer, b' represents all the parameters of the LSTM layer, represents the output of the LSTM layer;

[0032] The output layer predicts the channel based on the output of the LSTM layer:

[0033]

[0034] Among them, softmax(·) is the activation function, U represents the weight, v represents the bias, represents the frequency-domain channel prediction result according to the l-th group of frequency points;

[0035] Similarly, after obtaining the channel prediction results for each group of frequency points respectively, the frequency-domain channel prediction result c at the (n + 1)-th sampling time point is obtained OUT :

[0036]

[0037] The beneficial effects of the present invention are:

[0038] The frequency point grouping method based on the cross-frequency correlation matrix proposed by the present invention can avoid the problems of overly complex models and excessive parameters to be trained caused by too many input features of the deep learning predictor model, resulting in high computational complexity. Moreover, the CFDG-DL channel predictor combining the FC layer and the LSTM layer proposed by the present invention can utilize the complex time-frequency relationship of the frequency-domain channel and improve the performance of channel prediction. Description of the Drawings

[0039] Figure 1 It is a flowchart of an underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning according to the present invention;

[0040] Figure 2 It is a structural diagram of the CFDG-DL channel predictor;

[0041] Figure 3 It is a comparison chart of channel prediction errors between the method (CFDG-DL) of the present invention and the traditional time-domain single-tap (TD-ST) prediction method and frequency-domain single-frequency point (FD-SFP) prediction method on a public dataset. Detailed Embodiments

[0042] Detailed Embodiment 1: Combine Figure 1 This embodiment is described. An underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning described in this embodiment specifically includes the following steps:

[0043] Step 1: At the nth sampling time point t n Use an underwater acoustic communication machine to receive the time-domain passband signal, and then obtain the true value of the frequency-domain channel at the nth sampling time point t n After discretizing the obtained frequency-domain channel, at the nth sampling time point t n The frequency-domain response value of the kth frequency point is denoted as H(kΔf, t n ), where Δf is the frequency-domain sampling period, k = 1, 2,..., K, and K is the total number of frequency points;

[0044] According to the frequency-domain response values of each frequency point at the nth sampling time point and the frequency-domain response values of each frequency point at the previous n - P sampling time points (the previous n - P sampling time points refer to the n - P sampling time points immediately preceding the nth sampling time point, and the frequency-domain response values of each frequency point at the previous sampling time points are obtained in the same way as the nth sampling time point), construct the cross-frequency domain coherence matrix of the frequency-domain channel;

[0045] Step 2: After grouping the K frequency points according to the constructed cross-frequency domain coherence matrix, divide the K frequency points into L groups;

[0046] Step 3: Build a CFDG-DL channel predictor for each group of frequency points;

[0047] Step 4: Input the frequency-domain response values corresponding to each group of frequency points into the corresponding CFDG-DL channel predictor to obtain the channel prediction result at time (n + 1)T g (i.e., obtain the channel prediction result at the (n + 1)-th sampling time point).

[0048] The present invention prevents the CFDG-DL channel predictor from being too complex by grouping the frequency points. It also solves the problems that the traditional single-tap or single-frequency-point prediction method cannot utilize the complex time-frequency relationship of the channel and has a too high computational complexity, reduces the error of underwater acoustic channel prediction, and thus improves the performance of underwater acoustic communication. At the same time, the CFDG-DL channel predictor proposed by the present invention solves the problems of model mismatch and increased prediction error caused by the sparsity failure of the time-domain sparse channel. For the current sampling time point, the data of the current sampling time point and the data of the previous P - 1 sampling time points before the current sampling time point can be used to construct the cross-frequency-domain coherence matrix, and then the channel prediction result for the next sampling time point can be obtained.

[0049] Specific Embodiment 2: The difference between this embodiment and Specific Embodiment 1 is that the t n = nT g , where T g represents the time interval between adjacent sampling time points.

[0050] Other steps and parameters are the same as those in Specific Embodiment 1.

[0051] Specific Embodiment 3: The difference between this embodiment and Specific Embodiment 1 or 2 is that, according to the frequency-domain response values of each frequency point at the n-th sampling time point and the frequency-domain response values of each frequency point at the previous n - P sampling time points, a cross-frequency-domain coherence matrix of the frequency-domain channel is constructed; the element ρ(g,k) in the g-th row and k-th column of the cross-frequency-domain coherence matrix is:

[0052]

[0053] where ρ(g,k) represents the coherence between the g-th frequency point and the k-th frequency point, H(kΔf,t i ) is the frequency-domain response value of the k-th frequency point at the i-th sampling time point t i , H(gΔf,t i ) is the frequency-domain response value of the g-th frequency point at the i-th sampling time point t i , g = 1, 2,..., K, and H * (gΔf,t i ) represents H(gΔf,t i)'s conjugate transpose, P represents the total number of sampling time points used when constructing the cross-frequency domain coherence matrix of the frequency domain channel, |·| represents taking the absolute value; the angular brackets <·> denote ensemble averaging, f g = gΔf, f k = kΔf, where t is time.

[0054] Other steps and parameters are the same as those in the first or second specific implementation manner.

[0055] The cross-frequency domain coherence matrix is a diagonal matrix. In the frequency domain cross-coherence matrix, the correlation between different frequency points presents a block structure. According to this property, all frequency points can be divided into several groups.

[0056] Specific implementation manner four: The difference between this implementation manner and one of the first to third specific implementation manners is that in step 2, according to the constructed cross-frequency domain coherence matrix, K frequency points are grouped, and the grouping criterion used is:

[0057] Criterion 1): The frequency points within the same group are adjacent in sequence (that is, there cannot be discontinuous points within the same group);

[0058] Criterion 2): The number of frequency points within each group is less than the frequency point threshold. By setting the threshold, it can be avoided that the predictor model becomes too complex;

[0059] Criterion 3): The coherence between each frequency point within each group is greater than the set coherence threshold. In order to meet the requirement of the frequency point threshold, the coherence between the frequency points within the group needs to be greater than the coherence threshold. For example, if the first group includes frequency points 1, 2, and 3, then it should be satisfied that: the coherence between frequency point 1 and frequency point 2 is greater than the coherence threshold, the coherence between frequency point 1 and frequency point 3 is greater than the coherence threshold, and the coherence between frequency point 2 and frequency point 3 is greater than the coherence threshold, that is, the coherence between any two frequency points within the group is greater than the coherence threshold;

[0060] Group the K frequency points according to Criterion 1), Criterion 2), and Criterion 3) to obtain the frequency point grouping result.

[0061] Other steps and parameters are the same as those in one of the first to third specific implementation manners.

[0062] In this implementation manner, by analyzing the coherence block structure on the diagonal of the matrix, all frequency points are grouped according to the block result. The K frequency points are divided into L groups. The frequency points in the first group are the frequency points from 1 to k1, the frequency points in the second group are the frequency points from k1 + 1 to k2,..., and the frequency points in the Lth group are the frequency points from k L-1 + 1 to K.

[0063] Specific implementation manner five: Combine Figure 2Describe this embodiment. The difference between this embodiment and any one of the first to fourth specific embodiments is that the CFDG-DL channel predictor includes an input layer, a fully connected layer (FC), an LSTM layer, and an output layer.

[0064] Other steps and parameters are the same as any one of the first to fourth specific embodiments.

[0065] The input layer sizes of the CFDG-DL channel predictors corresponding to each group of frequency points are different, and the input layer size of each CFDG-DL channel predictor is the same as the number of frequency points within each group. For different groups, the structures of the FC layer and the LSTM layer in the CFDG-DL channel predictor are not fixed, and the number of layers and the number of units will be adaptively adjusted according to the hyperparameter optimization algorithm. The range of the number of FC layers is n f1 ~n f2 , the number of LSTM layers is n l1 ~n l2 . The number of units in each layer of FC and LSTM ranges from n u1 ~n u2 . After setting the learning rate, training batch size, and number of training epochs of the CFDG-DL channel predictor, use the divided frequency-domain channel sample data training set and validation set to train and validate the model, and then use the trained CFDG-DL channel predictor for channel prediction.

[0066] Specific embodiment six: The difference between this embodiment and any one of the first to fifth specific embodiments is that the specific process of step four is as follows:

[0067] For the l-th group of frequency points (the corresponding frequency range is k l-1 +1~k l ), use the frequency-domain response value of the l-th group of frequency points at the n-th sampling time point and the frequency-domain response values at the previous n - P' sampling time points as the input of the CFDG-DL channel predictor corresponding to the l-th group of frequency points, and denote the input of the CFDG-DL channel predictor corresponding to the l-th group of frequency points as c l (the value of n - P' can be set according to the actual situation. In the present invention, the value is set to 3, and the previous n - P' sampling time points refer to the n - P' sampling time points immediately before the n-th sampling time point). After c l is passed from the input layer to the FC layer, the output of the FC layer is:

[0068]

[0069] Among them, represents the output of the FC layer, b represents all the parameters of the FC layer, and f FC represents the operation process of the FC layer;

[0070] After extracting the correlation between frequency points in the FC layer, the extracted frequency point correlation is input into the LSTM layer. The operation process of the LSTM layer is defined as follows:

[0071]

[0072] Among them, f LSTM is the operation process of the LSTM layer, and b′ represents all the parameters of the LSTM layer, represents the output of the LSTM layer;

[0073] The output layer predicts the channel according to the output of the LSTM layer:

[0074]

[0075] Among them, softmax(·) is the activation function, U represents the weight, and v represents the bias, represents the prediction result of the frequency-domain channel according to the l-th group of frequency points;

[0076] Similarly, after obtaining the channel prediction results of each group of frequency points respectively, the frequency-domain channel prediction result c at the (n + 1)-th sampling time point is obtained OUT :

[0077]

[0078] Other steps and parameters are the same as those in any one of the specific embodiments one to five.

[0079] As Figure 3 shown, compared with the traditional time-domain single-tap prediction and frequency-domain single-frequency point prediction, the CFDG-DL channel prediction algorithm proposed by the present invention has lower prediction error and computational complexity on the same channel public dataset.

[0080] The above examples of the present invention are only for explaining in detail the calculation model and calculation process of the present invention, rather than limiting the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation manners here. Any obvious changes or modifications derived from the technical solutions of the present invention still fall within the protection scope of the present invention.

Claims

1. An underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning, characterized in that, The method specifically includes the following steps: Step 1: At the nth sampling time point t n , use an underwater acoustic communication machine to receive the time-domain passband signal, and then obtain the true value of the frequency-domain channel at the nth sampling time point t n . After discretizing the obtained frequency-domain channel, at the nth sampling time point t n , record the frequency-domain response value of the kth frequency point as H(kΔf, t n ), where Δf is the frequency-domain sampling period, k = 1, 2,..., K, and K is the total number of frequency points; The said t n = nT g , where T g represents the time interval between adjacent sampling time points; Construct a cross-frequency-domain coherence matrix of the frequency-domain channel according to the frequency-domain response values of each frequency point at the nth sampling time point and the frequency-domain response values of each frequency point at the historical n-P sampling time points; Step 2: After grouping the K frequency points according to the constructed cross-frequency-domain coherence matrix, divide the K frequency points into L groups; Step 3: Build a CFDG-DL channel predictor for each group of frequency points; The CFDG-DL channel predictor includes an input layer, a fully connected layer, an LSTM layer, and an output layer; Step 4: Input the frequency-domain response values corresponding to each group of frequency points into the corresponding CFDG-DL channel predictor respectively to obtain the channel prediction result at time (n + 1)T g ​ 2. The underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning according to claim 1, characterized in that Construct a cross-frequency-domain coherence matrix of the frequency-domain channel according to the frequency-domain response values of each frequency point at the nth sampling time point and the frequency-domain response values of each frequency point at the historical n-P sampling time points; the element ρ(g,k) in the gth row and kth column of the cross-frequency-domain coherence matrix is: where ρ(g,k) represents the coherence between the g-th frequency point and the k-th frequency point, and H(kΔf,t i ) is the frequency-domain response value of the k-th frequency point at the i-th sampling time point t i , H(gΔf,t i ) is the frequency-domain response value of the g-th frequency point at the i-th sampling time point t i , g = 1, 2, …, K, H * (gΔf,t i ) represents the conjugate transpose of H(gΔf,t i ), P represents the total number of sampling time points used when constructing the cross-frequency-domain coherence matrix of the frequency-domain channel, and |·| represents taking the absolute value.

3. The underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning according to claim 2, wherein In the second step, when grouping the K frequency points according to the constructed cross-frequency-domain coherence matrix, the grouping criterion adopted is: Criterion 1): The frequency points within the same group are adjacent in sequence; Criterion 2): The number of frequency points in each group is less than the frequency point threshold; Criterion 3): The coherence between each frequency point within each group is greater than the set coherence threshold; Group the K frequency points according to Criterion 1), Criterion 2), and Criterion 3) to obtain the frequency point grouping result.

4. The underwater acoustic channel prediction method based on cross-frequency domain grouping and deep learning according to claim 3, wherein The specific process of the fourth step is: For the l-th group of frequency points, the frequency-domain response value of the l-th group of frequency points at the n-th sampling time point and the frequency-domain response values at the previous n - P' sampling time points are used as the inputs to the CFDG-DL channel predictor corresponding to the l-th group of frequency points, and the inputs to the CFDG-DL channel predictor corresponding to the l-th group of frequency points are denoted as c l , c l After being passed from the input layer to the FC layer, the output of the FC layer is: Among them, represents the output of the FC layer, b represents all the parameters of the FC layer, and f FC represents the operation process of the FC layer; After extracting the correlation between frequency points in the FC layer, input the extracted frequency point correlation into the LSTM layer, and define the operation process of the LSTM layer as follows: Among them, f LSTM is the operation process of the LSTM layer, and b' represents all the parameters of the LSTM layer, representing the output of the LSTM layer; The output layer predicts the channel according to the output of the LSTM layer; where softmax(·) is the activation function, U represents the weights, and v represents the biases, represents the frequency-domain channel prediction result according to the frequency points of the l-th group; Similarly, after obtaining the channel prediction results for each group of frequency points respectively, the frequency-domain channel prediction result c at the (n + 1)-th sampling time point is obtained. OUT :

Images