Multi-dimensional feature aggregation network channel estimation method based on AUV wireless communication system

Abstract

The invention discloses a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system, and the method comprises the following steps: collecting channel frequency response and receiving signal data of an AUV in an electromagnetic environment when the AUV works on a water surface, and making a data set; constructing a multi-dimensional feature aggregation network model based on a self-attention mechanism; performing offline training on the multi-dimensional feature aggregation network by using the data set; performing data preprocessing on the received signal of the AUV to obtain input data; and loading the network model after offline training, and inputting the input data into the network for channel estimation. According to the multi-dimensional feature aggregation network channel estimation method based on the AUV wireless communication system, a self-attention mechanism in deep learning is used for building a multi-dimensional feature aggregation network model, the estimation performance is far higher than that of a traditional channel estimation method, the space-time complexity of the network is very low, and the method is easy to implement. The method can be applied to an AUV wireless communication system in an offline training mode.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication signal processing, and in particular to a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system. Background Art

[0002] AUV is an unmanned, automatically controlled underwater vehicle. When AUV communicates on the surface, there are many requirements for predicting the communication environment, among which the estimation of channel frequency response is particularly important. Based on the results of channel estimation, AUV can not only compensate for signal fading to ensure the reliability of communication, but also adaptively adjust the modulation and coding scheme of the transmitter to adapt to the current channel environment. Therefore, channel estimation is an indispensable link in the surface communication of AUV.

[0003] Traditional channel estimation methods mainly include pilot-based LS estimation and MMSE (Minimum Mean Square Error) estimation. LS estimation ignores the influence of noise and involves only one division operation, which is simple and easy to implement. However, with the increasing complexity of the electromagnetic environment at sea and the increasing intensity of electromagnetic games in strong confrontation environments, its estimation performance does not meet the requirements of AUV surface communications. MMSE estimation is very complex, involves two matrix inversions, and requires accurate channel statistics and noise variance. This makes MMSE estimation very difficult to implement in reality, although it has extremely high estimation performance.

[0004] In recent years, with the development of deep learning, more and more researchers have used deep learning methods to achieve channel estimation and achieved good results. Existing deep learning-based channel estimation methods, such as ChannelNet and CRCENet, have much higher estimation performance than traditional channel estimation methods. However, on the one hand, these networks are designed for channel estimation of land communications, and their communication environment is not as complex and changeable as the marine communication environment. On the other hand, the computational complexity of these networks is extremely high, making it difficult to match the hardware equipment of AUVs. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system, and the time and space complexity of the network model is low, which can be applied to the AUV through offline training, achieving an estimation performance far superior to that of traditional channel estimation methods, thereby solving the above-mentioned technical problems existing in the prior art.

[0006] To achieve the above object, the present invention provides a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system, comprising the following steps:

[0007] S1. Collect the channel frequency response and received signal data in the electromagnetic environment where the AUV is located when working on the water surface, extract the pilot signal from the received signal, and upsample the pilot signal using the interpolation method based on DPA estimation as the input data of the model to be trained. The channel frequency response in the collected data is used as the training label and corresponds to the received signal data to form a data set.

[0008] S2. Construct a multi-dimensional feature aggregation network model based on the self-attention mechanism; wherein the multi-dimensional feature aggregation network model FACENet based on the self-attention mechanism includes a multi-dimensional feature aggregation module, a feature processing module and an upsampling module;

[0009] S3, in the multi-dimensional feature aggregation network model established in S2, use the data set collected by S1 for offline training to obtain a trained network model;

[0010] S4. When the AUV is actually working on the water surface, data preprocessing is performed on the receiving end of the AUV to obtain input data of the network model trained in S3.

[0011] S5. Input the input data obtained in S4 into the network model trained in S3, estimate the real channel frequency response of the environment where the AUV is located, and complete the channel estimation.

[0012] Preferably, in step S1, up-sampling the pilot signal based on the interpolation method estimated by DPA comprises the following steps:

[0013] Perform LS estimation at the pilot position for the first OFDM symbol and the eighth OFDM symbol;

[0014] The LS estimation of the first eighth OFDM symbol is used as a virtual preamble, and the DPA estimation of the remaining OFDM symbols is performed using the virtual preamble;

[0015] Place the DPA estimation result in the corresponding position of the real channel matrix and perform linear interpolation in the OFDM symbol and subcarrier dimensions;

[0016] The interpolation result is separated into real and imaginary parts as input data for the model to be trained.

[0017] Preferably, in step S2, the multidimensional feature aggregation module aggregates the features of the input data in multiple dimensions, including a spatial feature aggregation block and a channel feature aggregation block; the spatial feature aggregation block extracts the spatial features of the input data from the time direction and the frequency direction; the channel feature aggregation block extracts the channel features of the input data from the channel direction.

[0018] Preferably, the spatial feature aggregation block includes a spatial self-attention block and a feed-forward network;

[0019] The spatial self-attention block consists of three parallel branches. The first branch divides the input data into 72 patches with a feature dimension of 16 along the frequency direction, and then implements a multi-head attention mechanism to extract the spatial features of the input data from the frequency dimension. The second branch divides the input data into 14 patches with a feature dimension of 16 along the time direction, and then implements a multi-head attention mechanism to extract the spatial features of the input data from the time dimension. The third branch extracts local features from the input data through a convolutional layer. The first two branches fuse the spatial features of the frequency dimension and the time dimension through matrix multiplication, and the results are added to the third branch to obtain the output of the spatial self-attention block.

[0020] The feed-forward network uses a LayerNorm layer, two linear layers, and an activation layer to process the features extracted by the spatial self-attention block.

[0021] Preferably, the channel feature aggregation block includes a channel self-attention block and a feedforward network; the channel self-attention block consists of two parallel branches, the first branch average-pools the input data and then implements a multi-head attention mechanism to extract the channel features of the input data from the channel dimension; the second branch extracts local features from the input data through a convolutional layer. The feedforward network uses a LayerNorm layer, two linear layers, and an activation layer to process the features extracted by the two branches.

[0022] Preferably, the feature processing module further processes the extracted features, first using a convolution layer to preprocess the features extracted by the multi-dimensional feature aggregation module, and then using four consecutive RBs for feature processing. The output of the last RB will pass through a convolution layer to obtain the result of feature processing, and the outputs of the two convolution layers are connected through a residual structure.

[0023] Preferably, the RB is formed by cascading a convolutional layer, an activation layer whose activation function is Gelu, and a convolutional layer, and the input and output of the RB are connected by a residual structure.

[0024] Preferably, the upsampling module restores the processed features to the target size using the Pixel Shuffle upsampling method.

[0025] Preferably, in step S4, the process of data preprocessing of the received signal by the AUV at the receiving end is consistent with the step of up-sampling the pilot signal using the interpolation method estimated by DPA.

[0026] Preferably, in step S5, the AUV loads the FACENet model that has been trained offline, inputs the received data that has undergone data preprocessing into FACENet, and estimates the actual channel frequency response of the current environment in which the AUV is located through model calculation.

[0027] Therefore, the present invention adopts the above-mentioned multi-dimensional feature aggregation network channel estimation method based on the AUV wireless communication system, and its technical effects are as follows:

[0028] (1) The present invention uses the self-attention mechanism in deep learning to build a multi-dimensional feature aggregation network for channel estimation. The network aggregates the spatial and channel features of the input data by alternating between spatial self-attention and channel self-attention, and processes these features through a deep residual network. Considering that the channel has a strong correlation in time and frequency, spatial self-attention obtains and fuses spatial features from the directions of time and frequency respectively. Such a network design combines the characteristics of deep learning and OFDM signals themselves, can better extract the features of the input data, and makes the network's estimation performance far superior to traditional channel estimation, and compared with other deep learning-based channel estimation schemes, it also presents better performance.

[0029] (2) The present invention uses an interpolation method based on DPA estimation to upsample the pilot signal. The OFDM symbol containing the pilot is used as a virtual preamble, and the virtual preamble is used to perform DPA estimation on the remaining OFDM symbols. Linear interpolation is performed in two dimensions: OFDM symbol and subcarrier. This allows the constructed network to support flexible pilot patterns and maximize the use of pilot signals.

[0030] (3) The multi-dimensional feature aggregation network constructed by the present invention has low spatiotemporal complexity. First, the time required to complete an OFDM subframe channel estimation is 1.6863ms. It can be seen that the delay caused by the calculation is allowed in the communication environment of the real AUV, and the operations performed in the network model are all multiplication, division, addition and subtraction, so the implementation in hardware is not particularly difficult. Secondly, the number of model parameters of the multi-dimensional feature aggregation network is 66,746, which is significantly reduced compared to other channel estimation networks based on deep learning such as ChannlNet (the number of model parameters is 678,658) and CRCENet (the number of model parameters is 961,238). Therefore, the multi-dimensional feature aggregation network constructed by the present invention is suitable for application in AUV.

[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1A flowchart of implementing deep learning channel estimation on an AUV according to the present invention;

[0033] Figure 2 A schematic diagram of a pilot scheme used in the present invention;

[0034] Figure 3 A schematic diagram of the network architecture of the multi-dimensional feature aggregation network constructed by the present invention;

[0035] Figure 4 A schematic diagram of the network architecture of a multi-dimensional feature aggregation module in a multi-dimensional feature aggregation network;

[0036] Figure 5 Schematic diagram of the network architecture of the spatial self-attention block in the multi-dimensional feature aggregation module;

[0037] Figure 6 Schematic diagram of the network architecture of the channel self-attention block in the multi-dimensional feature aggregation module;

[0038] Figure 7 is an interpolation flow chart of an interpolation scheme based on DPA estimation;

[0039] Figure 8 It is a simulation flow chart of the channel estimation method based on the multi-dimensional feature aggregation network;

[0040] Fig. 9 It is the MSE performance simulation diagram of various channel estimation methods;

[0041] Fig.10 The BER performance simulation diagram of various channel estimation methods. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0043] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0044] Embodiment 1

[0045] like Figure 1 As shown, the present invention provides a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system, comprising the following steps:

[0046] S1. Collect the channel frequency response and received signal data in the electromagnetic environment where the AUV is located when working on the water surface, use the channel frequency response in the collected data as a training label, and correspond it to the received signal data to form a data set;

[0047] The environment in which the data set is collected is the surface of the sea area where the AUV works. Each type of sea area can be defined as a type of channel model. The data set under each type of channel model covers the received signal data and channel frequency response data under different signal-to-noise ratios (5-25dB, with 5dB intervals). The received signal data is collected according to the AUV receiving end, and the channel frequency response data is collected according to a dedicated channel measurement device. In the collected data, the pilot signal is extracted from the received signal, and the pilot signal is upsampled using an interpolation method based on DPA estimation as the input data of the model to be trained, and the corresponding channel frequency response is used as the label. The pilot signal extracted from the received signal is distributed in a 32×2 manner in the first OFDM symbol and the eighth OFDM symbol of the 72×14 time-frequency grid, as shown in Figure 2. Figure 2 As shown. The pilot signal is sent by a shore-based transmitter or other AUV transmitter together with the transmitted data, and its value is 1+1j. The interpolation method based on DPA estimation is used to upsample the pilot signal, including the following steps:

[0048] S11. For the first OFDM symbol and the eighth OFDM symbol, LS estimation is performed at the pilot position using formula (1):

[0049]

[0050] In formula (1), X p is the pilot signal sent, Y p is the received signal at the pilot, is the LS estimation result at the pilot frequency;

[0051] For example, the position of the pilot in the OFDM subframe is known to both communicating parties (shore-based and AUV, AUV and AUV) before communication. In this embodiment, the pilot is located at the 1st, 3rd, 5th, ..., 71st subcarrier positions in the 1st OFDM symbol in the OFDM subframe and the 2nd, 4th, 6th, ..., 72nd subcarrier positions in the 8th OFDM symbol. The overall distribution is in a 32×2 grid form in a 72×14 time-frequency grid, as shown in FIG. Figure 2 shown.

[0052] S12, taking the LS estimation of the first eighth OFDM symbol as a virtual preamble, and using the virtual preamble to perform DPA estimation on the remaining OFDM symbols, the DPA estimation is shown in equations (2) and (3):

[0053]

[0054] In formulas (2) and (3), y i is the i-th received OFDM symbol, is the DPA estimate of the ith OFDM symbol, yeqi is the ith OFDM symbol to be equalized, d i is the equalized OFDM symbol y eqi is mapped to the nearest constellation point.

[0055] The production of a virtual leader includes the following specific steps:

[0056] The corresponding LS estimates are calculated using formula (1) for the 1st, 3rd, 5th, ..., 71st subcarrier positions in the 1st OFDM symbol, the channel estimation values ​​at the 2nd, 4th, 6th, ..., 72nd subcarrier positions are set to 0, and the processed 1st OFDM symbol is used as the 1st virtual preamble; similarly, the corresponding LS estimates are calculated using formula (1) for the 2nd, 4th, 6th, ..., 72nd subcarrier positions in the 8th OFDM symbol, the channel estimation values ​​at the 1st, 3rd, 5th, ..., 71st subcarrier positions are set to 0, and the processed 8th OFDM symbol is used as the 2nd virtual preamble.

[0057] S13, placing the estimated result in the corresponding position of the real channel matrix, and performing linear interpolation in the OFDM symbol and subcarrier dimensions

[0058] The following specific steps are involved in DPA estimation at other locations:

[0059] The DPA estimation is performed on the 1st, 3rd, 5th, ..., 71st subcarrier positions on the 2nd to 14th OFDM symbols as shown in formulas (2) and (3). is the first virtual preamble produced; similarly, the DPA estimation is performed on the 2nd, 4th, 6th, ..., 72nd subcarrier positions on the 9th to 14th OFDM symbols as shown in formulas (2) and (3). At this time, This is the second virtual preamble created; for other positions involving DPA estimation, their channel estimation values ​​are all set to 0.

[0060] In step S104, linear interpolation is as shown in formula (4):

[0061]

[0062] Among them, (x1, y1) and (x2, y2) are the two known data points closest to the interpolation point (x, y). By implementing linear interpolation in both the OFDM symbol and subcarrier directions, the completed DPA estimation result can be upsampled to the scale of the complete OFDM subframe, that is, 72×14.

[0063] S14, separate the real part and the imaginary part of the interpolation result and use it as the third dimension of the input data, so that the input of the model is As input data for the model to be trained.

[0064] S2. Construct a multi-dimensional feature aggregation network model based on self-attention mechanism suitable for AUV;

[0065] The multi-dimensional feature aggregation network model FACENet based on the self-attention mechanism includes a multi-dimensional feature aggregation module, a feature processing module and an upsampling module. Figure 3 As shown in the figure, the multi-dimensional feature aggregation module aggregates the features of the input data in multiple dimensions to extract the input features to the greatest extent; the feature processing module further processes the extracted features; and the upsampling module restores the processed features to the target size.

[0066] The multi-dimensional feature aggregation module is the core of the multi-dimensional feature aggregation network. This module aggregates the features of the input data from multiple dimensions. The module consists of a continuous spatial feature aggregation block (SFAB) and a channel feature aggregation block (CFAB), such as Figure 4 shown.

[0067] SFAB consists of a spatial self-attention block (S-SAB) and a feed-forward network (FFN), such as Figure 5 As shown in Figure 1, its core lies in S-SAB. S-SAB takes into account that the channel has a strong correlation in time and frequency, and spatial self-attention obtains and fuses spatial features from the directions of time and frequency respectively. To this end, S-SAB first uses the idea of ​​visual transformer (VIT) to divide the input data into several patches along the directions of time and frequency, and then implements self-attention on these patches. Finally, through matrix multiplication, the features extracted in these two directions can be fused to obtain the features of the input data in the spatial dimension. Similarly, CFAB consists of a channel self-attention block (C-SAB) and an FNN, and its core lies in C-SAB. C-SAB averages the spatial features extracted by S-SAB, and implements self-attention in the channel direction to obtain the features in the channel dimension. Note that C-SAB completes downsampling while performing the average pooling operation, which enables subsequent feature processing to be performed on a small scale, reducing the parameters and computational complexity of the model. Considering that the features extracted by the self-attention mechanism are global features, we implemented a parallel convolution layer in S-SAB and C-SAB to obtain local features. The local features obtained by convolution are added to the corresponding elements of the global features obtained by self-attention to obtain the corresponding outputs of S-SAB and C-SAB.

[0068] The spatial self-attention block consists of three parallel branches. The first branch divides the input data into 72 patches with a feature dimension of 16 along the frequency direction, and then implements a multi-head attention mechanism to extract the spatial features of the input data from the frequency dimension. The second branch divides the input data into 14 patches with a feature dimension of 16 along the time direction, and then implements a multi-head attention mechanism to extract the spatial features of the input data from the time dimension. The third branch extracts local features from the input data through a convolutional layer. The first two branches fuse the spatial features of the frequency dimension and the time dimension through matrix multiplication, and the results are added to the third branch to obtain the output of the spatial self-attention block. The feedforward network uses a LayerNorm layer, two linear layers, and an activation layer to process the features extracted by the spatial self-attention block.

[0069] The channel feature aggregation block consists of a channel self-attention block and a feed-forward network. Figure 6 As shown in Figure 1, the channel self-attention block consists of two parallel branches. The first branch average-pools the input data and then implements a multi-head attention mechanism to extract the channel features of the input data from the channel dimension. The second branch passes the input data through a convolutional layer to extract local features. The feedforward network uses a LayerNorm layer, two linear layers, and an activation layer to process the features extracted by the two branches.

[0070] Spatial Self-Attention Block (S-SAB): Figure 5 As shown, S-SAB has three branches. The first two branches are used to obtain the spatial features of the input data (from the time and frequency directions respectively), and the last branch is used to obtain the local features of the input data. The first branch first divides it into K features with N dimensions along the frequency direction e The set of patches is The specific operation of patch division is: input data After a convolution kernel size of (1, I), the number of filters is N e , the convolutional layer with a stride of 1, the output Refactored to X f The frequency self-attention layer (F-SA) will focus on X along the frequency direction. f Implement multi-head self-attention. First, X f It will be mapped to a size of K×N by the following formula e The query(Q), key(K) and value(V) matrices are:

[0071] Q f =X f W Q ,K f =X fW K ,V f =X f W V , (5)

[0072] in, To omit the linear mapping of the bias, the subscript f denotes the frequency of its implementation of multi-head self-attention.

[0073] Next, Q f ,K f ,V f Separation into N head Size: The dimension of each head is d = N e / N head The output of the ith head is It is calculated by the following formula:

[0074]

[0075] Finally, the output Y of F-SA is obtained by connecting the outputs of all heads and performing linear mapping f :

[0076]

[0077] in, To omit the linear mapping of the deviation, It represents the extraction of spatial features in the frequency direction of the input data.

[0078] For input data The second branch first divides it into I feature dimensions N along the time direction e The set of patches is Note that the convolution kernel size used in the patch division is (K, 1). The temporal self-attention layer (T-SA) will t Implementing multi-head self-attention, the steps are exactly the same as F-SA. It is worth noting that This is consistent with the Q f ,K f ,V f ,Y f The subscript t indicates the time of multi-head self-attention implementation.

[0079] In order to implement the spatial feature Y extracted by self-attention along the frequency f And the spatial feature Y extracted by self-attention along time t To perform fusion, a Reshape operation is required first to obtain and Then matrix multiplication can be performed to obtain spatial features As shown below:

[0080]

[0081] The third branch implements a convolution kernel with size (N kernel ,N kernel ), the number of filters is N e , a convolutional layer with a step size of 1 is used to extract the local features of the input data. The result after convolution is

[0082] Finally, through and By adding the corresponding elements, we can get the output of S-SA.

[0083] Feedforward Network (FFN): SFAB includes an LN layer, an FNN, and a residual structure in addition to S-SAB. FFN includes two linear mappings with omitted biases. and 1 activation layer GeLU. The output H of S-SAB S-SAB After FFN, the output of SFAB is

[0084] Channel Self-Attention Block (C-SAB): Figure 6 As shown, C-SAB has two branches. The first branch is used to obtain the channel features of the input data, and the second branch is used to obtain the local features of the input data. S , the first branch first performs average pooling with a pooling size of N d ×N d , the step length is N d ×N d The result after pooling will be reconstructed as As the input of channel self-attention (C-SA). C-SA uses formulas (5) to (7) to calculate X C is processed, where the resulting matrix Linear mapping with omitted bias Output of C-SA The channel features of the input data are extracted. Another branch implements a convolution kernel with a size of N d ×N d , the number of filters is N e , the step length is N d ×N d The convolution layer is used to extract the local features of the input data. The result after convolution is Finally, Y C Refactored to and Add the corresponding elements to get the output of C-SAB

[0085] Like S-SAB, C-SAB is also connected to an identical LN layer, FNN and residual structure. C-SAB , the output of CFAM is

[0086] The feature processing module further processes the extracted features. First, a convolution layer is used to preprocess the features extracted by the multi-dimensional feature aggregation module. Then, four consecutive RBs are used for feature processing. The output of the last RB will be passed through a convolution layer to obtain the result of feature processing. The outputs of these two convolution layers are connected through a residual structure. For example, the feature processing module consists of N RB RBs, each RB consists of two identical convolutional layers, an activation layer, and a residual structure. The convolution kernel size of the convolutional layer is N kernel ×N kernel , the number of filters is N e The activation function of the activation layer is ReLU. The residual structure adds the input and output of RB. There are two convolutional layers before and after the feature processing module, which are exactly the same as the convolutional layers in RB. For the output H of module 101 C After passing through the first convolution layer, it will be used as the input of the feature processing module. The output of this module passes through the second convolution layer and is added to the input through a residual structure to obtain the processed features.

[0087] The upsampling module restores the processed features to the target size using the pixel shuffle upsampling method. For example, after the pixel shuffle method is completed, H will be restored to the size of the real channel matrix through the upsampling module. Before the pixel shuffle, there is a convolution kernel size of N kernel ×N kernel , the number of filters is N e ×N d The convolution layer provides enough depth features for pixel shuffle. After pixel shuffle, there is a convolution kernel with size N. kernel ×N kernel , a convolutional layer with 2 filters is used to make the number of output channels consistent with the number of label channels. The output of the upsampling layer is

[0088] For the multi-dimensional feature aggregation network, we set the parameters in the network framework as follows: N e=16, N head =2, N kernel =3,N d =4,N RB = 4. Finally, the total number of network training parameters is 66,746.

[0089] S3, in the multi-dimensional feature aggregation network model established in S2, the data set collected in S1 is used for offline training to obtain a trained network model. For example, FACENet is trained offline under the hyperparameters of a learning rate of 0.001, an optimizer of Adam, a loss function of L1 loss, a batch size of 128, and a training cycle of 100 to obtain a trained weight.

[0090] S4. When the AUV is actually working on the water surface, the received signal is preprocessed at the receiving end of the AUV to obtain the input data of the network model trained in S3. The process of the AUV preprocessing the received signal at the receiving end is consistent with S11 to S14.

[0091] S5. Input the input data obtained in S4 into the network model trained in S3, estimate the real channel frequency response of the environment where the AUV is located, and complete the channel estimation.

[0092] Embodiment 2

[0093] like Figure 8 As shown, a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system is also disclosed, including the following steps:

[0094] Step 1: Preparation of training data set

[0095] Matlab 2023a was used to simulate the marine channel environment, and a training data set was produced by simulation. Considering the SISO-OFDM system, for each time slot in an OFDM subframe, there are 14 OFDM symbols, and each OFDM symbol has 72 subcarriers. The pilots are arranged in a 36×2 grid in the transmission symbol. The modulation mode of the signal in the system is divided into QPSK. The channel model is a marine multipath channel model, the subcarrier frequency is 2.1GHz, the carrier frequency is 15kHz, the sampling frequency is 1.08MHz, and the AUV speed is 50km / h. The training data set is generated under the above system parameters. The signal-to-noise ratio (SNR) is 5dB to 25dB at intervals of 5dB, and 10,000 samples are generated for each signal-to-noise ratio, for a total of 50,000 channels. The ratio of the training set to the validation set in the training data set is 3:1.

[0096] Step 2: Offline training of multi-dimensional feature aggregation network;

[0097] The multi-dimensional feature aggregation network FACENet is trained offline on NVIDIA GeForce RTX 3080, Tensorflow-2.5.0 with hyperparameters of learning rate 0.001, optimizer Adam, loss function L1 loss, batch size 128, and training cycle 100 to obtain the trained weights.

[0098] Step 3: Preparation of test data set

[0099] The system parameters for generating the test data set are consistent with the training data set except for the AUV speed. SNR is in the range of 0dB to 30dB with 5dB intervals, and 2500 samples are generated for each signal-to-noise ratio, with a total of 17,500 channels. For each sample under each signal-to-noise ratio, the receiving end speed is randomly selected from 0 to 50km / h, that is, the Doppler frequency range of these 2500 samples is 0 to 97Hz.

[0100] Step 4: Channel estimation through multi-dimensional feature aggregation network

[0101] Use the trained FACENet to test the test data set and get the channel estimation result. Figure 8 , 9 As shown in the figure, the minimum mean square error (MSE) and system bit error rate (BER) of channel estimation using FACENet are shown. CRCENet, Channelformer, ChannelNet and ReEsNet in the figure are other deep learning channel estimation schemes, and LMMSE, LS and DPA-interpolation are non-deep learning channel estimation schemes for comparison. Fig. 9 As can be seen from , FACENet has the smallest MSE and the best estimation performance under the full signal-to-noise ratio. Fig.10 As can be seen in Figure 2, FACENet has the smallest BER and the highest system reliability under the full signal-to-noise ratio.

[0102] Therefore, the present invention adopts the above-mentioned multi-dimensional feature aggregation network channel estimation method based on the AUV wireless communication system. The network model involved has low time and space complexity and can be applied to the AUV through offline training to achieve an estimation performance far superior to the traditional channel estimation method, thereby solving the above-mentioned technical problems existing in the prior art.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-dimensional feature aggregation network channel estimation method based on AUV wireless communication system, characterized in that: The following steps are involved: S1. Collect the channel frequency response and received signal data in the electromagnetic environment where the AUV is located when working on the water surface, extract the pilot signal from the received signal, and upsample the pilot signal using the interpolation method based on DPA estimation as the input data of the model to be trained. The channel frequency response in the collected data is used as the training label and corresponds to the received signal data to form a data set. S2. Construct a multi-dimensional feature aggregation network model based on the self-attention mechanism; wherein the multi-dimensional feature aggregation network model FACENet based on the self-attention mechanism includes a multi-dimensional feature aggregation module, a feature processing module and an upsampling module; S3, in the multi-dimensional feature aggregation network model established in S2, use the data set collected by S1 for offline training to obtain a trained network model; S4. When the AUV is actually working on the water surface, data preprocessing is performed on the receiving end of the AUV to obtain input data of the network model trained in S3. S5. Input the input data obtained in S4 into the network model trained in S3, estimate the real channel frequency response of the environment where the AUV is located, and complete the channel estimation.

2. According to claim 1, a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system is characterized in that: In step S1, upsampling the pilot signal based on the interpolation method estimated by DPA includes the following steps: Perform LS estimation at the pilot position for the first OFDM symbol and the eighth OFDM symbol; The LS estimation of the first eighth OFDM symbol is used as a virtual preamble, and the DPA estimation of the remaining OFDM symbols is performed using the virtual preamble; Place the DPA estimation result in the corresponding position of the real channel matrix and perform linear interpolation in the OFDM symbol and subcarrier dimensions; The interpolation result is separated into real and imaginary parts as input data for the model to be trained.

3. According to claim 1, a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system is characterized in that: In step S2, the multi-dimensional feature aggregation module aggregates the features of the input data in multiple dimensions, including a spatial feature aggregation block and a channel feature aggregation block; The spatial feature aggregation block extracts the spatial features of the input data from the time direction and frequency direction; The channel feature aggregation block extracts the channel features of the input data from the channel direction.

4. A multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system according to claim 3, characterized in that: The spatial feature aggregation block consists of a spatial self-attention block and a feed-forward network; The spatial self-attention block consists of three parallel branches. The first branch divides the input data into 72 patches with a feature dimension of 16 along the frequency direction, and then implements a multi-head attention mechanism to extract the spatial features of the input data from the frequency dimension; The second branch divides the input data into 14 patches with a feature dimension of 16 along the temporal direction, and then implements a multi-head attention mechanism to extract the spatial features of the input data from the temporal dimension; The third branch passes the input data through a convolutional layer to extract local features ； The first two branches fuse the spatial features of the frequency dimension and the time dimension through matrix multiplication, and the result is added to the third branch to obtain the output of the spatial self-attention block; The feed-forward network uses a LayerNorm layer, two linear layers, and an activation layer to process the features extracted by the spatial self-attention block.

5. According to claim 3, a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system is characterized in that: The channel feature aggregation block includes a channel self-attention block and a feed-forward network; the channel self-attention block consists of two parallel branches. The first branch average-pools the input data and then implements a multi-head attention mechanism to extract the channel features of the input data from the channel dimension; The second branch passes the input data through a convolutional layer to extract local features; The feed-forward network uses a LayerNorm layer, two linear layers, and an activation layer to process the features extracted by the two branches.

6. According to a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system according to claim 1, it is characterized in that: The feature processing module further processes the extracted features. First, a convolution layer is used to preprocess the features extracted by the multi-dimensional feature aggregation module. Then four consecutive RBs are used for feature processing. The output of the last RB will pass through a convolution layer to obtain the result of feature processing, and the outputs of these two convolution layers are connected through a residual structure.

7. A multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system according to claim 6, characterized in that: RB is composed of a convolutional layer, an activation layer with Gelu activation function, and a convolutional layer cascaded, and the input and output of RB are connected by a residual structure.

8. According to claim 1, a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system is characterized in that: The upsampling module restores the processed features to the target size using the Pixel Shuffle upsampling method.

9. According to claim 1, a multi-dimensional feature aggregation network channel estimation method based on an AUV wireless communication system is characterized in that: In step S4, the process of data preprocessing of the received signal by the AUV at the receiving end is consistent with the step of up-sampling the pilot signal using the interpolation method estimated by the DPA.

10. The multi-dimensional feature aggregation network channel estimation method based on the AUV wireless communication system according to claim 1 is characterized in that: In step S5, the AUV loads the FACENet model that has been trained offline, and inputs the received data that has undergone data preprocessing into FACENet. Through model calculation, the actual channel frequency response of the current AUV environment can be estimated.

Images