Underwater sound adaptive transmission method based on channel prediction

Abstract

The invention discloses an underwater acoustic adaptive transmission method based on channel prediction, and belongs to the technical field of communication. The method comprises the following steps of: 1, designing a transmitting link of an underwater acoustic self-adaptive transmission system based on channel prediction; and 2, designing a receiving link of the underwater acoustic adaptive transmission system based on channel prediction. According to the method, chirp parameters are dynamically selected on a sending link based on a predicted channel state, then adaptive transmission is carried out on an imitated radio frequency division multiplexing signal, channel state information is estimated through embedded pilot frequency at a receiving end, channel characteristics are fully learned through a channel prediction model of Transform, and the channel state information is subjected to channel prediction. And the obtained accurate future channel state information is used for balancing and self-adaptive parameter adjustment, and the flexibility and the transmission reliability of underwater acoustic communication are improved by the method provided by the invention. The method is very suitable for a half-duplex-based underwater acoustic adaptive transmission system, and the excellent performance of the method is fully embodied.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and relates to adaptation, channel estimation, channel prediction, half-duplex communication, and AFDM (Affine Frequency Division Multiplexing) technology. Background Art

[0002] With the rapid development of fields such as marine resource exploration and development, environmental monitoring, and national defense security, underwater acoustic communication, as a core technology for achieving reliable underwater information transmission, has become increasingly important. However, underwater acoustic channels have inherent characteristics such as low propagation speed, extremely limited available bandwidth, complex background noise, severe multipath effects, and significant Doppler frequency shifts caused by the relative motion of the transmitter and receiver and the fluctuations of the seawater medium. These characteristics together constitute a typical dual-selective (time-frequency dual-selective) fading channel, which seriously restricts the transmission rate, spectrum efficiency, link reliability, and adaptability of underwater acoustic communication systems to dynamic environments. In particular, in modern underwater communication applications that pursue high speed, low latency, and high reliability, traditional feedback-based adaptive schemes that rely solely on instantaneous channel state information (CSI) are unable to effectively cope with the impact of channel information obsolescence caused by the rapid time-varying underwater acoustic channels. Channel prediction technology is of great significance in combating outdated channel state information. This technology uses a neural network model to predict the channel state at future moments and uses the predicted future CSI to guide the adaptive adjustment of the transmitter's parameters. It can effectively compensate for the feedback loop delay and make the system parameter configuration closer to the actual channel conditions during data transmission, thereby significantly improving link reliability and spectrum efficiency. On the other hand, AFDM, as an emerging multi-carrier modulation technology, uses linear frequency modulation (Chirp) signals generated by the Discrete Affine Fourier Transform (DAFT) as subcarriers, demonstrating excellent characteristics for combating double-selective fading channels. Its core advantage is that by flexibly adjusting the two chirp parameters of the DAFT, it can match the channel's delay and Doppler spread characteristics, thereby achieving performance close to full diversity in time-varying channels. Summary of the Invention

[0003] The purpose of the present invention is to address the technical problems existing in the background technology and design an underwater acoustic adaptive transmission method based on channel prediction to accurately obtain the future channel state to improve the flexibility and reliability of transmission. The present invention is the first to propose an underwater acoustic adaptive transmission method based on channel prediction, which incorporates channel prediction into the half-duplex AFDM system framework. First, accurate channel states are obtained through channel estimation and stored in the CSI database. Then, the time correlation between channel states is learned by using a Transformer network prediction model. Finally, the transmission parameters are dynamically adjusted according to the future underwater acoustic channel state, thereby improving the flexibility and transmission reliability of the underwater acoustic communication system and fully demonstrating its superior performance.

[0004] In order to achieve the above object, the technical solution of the present invention is:

[0005] An underwater acoustic adaptive transmission method based on channel prediction comprises the following steps:

[0006] Step 1: Design of the transmission link of the underwater acoustic adaptive transmission system based on channel prediction;

[0007] Step 2: Design scheme of receiving link of underwater acoustic adaptive transmission system based on channel prediction;

[0008] The specific steps in step 1 are as follows:

[0009] Step 1.1: Switch the transceiver switch to the transmit / receive mode of node A.

[0010] Step 1.2: Constellation mapping is performed on the bits to be transmitted, and the transmission parameters are dynamically adjusted according to the predicted future channel state.

[0011] Step 1.3: Map the modulation symbols onto subcarriers to form an affine Fourier domain transmit signal. Then, based on the predicted channel state, dynamically select parameters and construct the DAFT matrix A.

[0012] Step 1.4: Use IDAFT matrix A -1 The affine Fourier domain transmit signal is transformed into the time domain, and a Chirp Cyclic Prefix (CPP) is added. The AFDM time domain signal is then transmitted on the transducer according to the traditional scheme.

[0013] The specific steps in step 2 are summarized as follows:

[0014] Step 2.1: Switch the transceiver switch to the receiving mode of Node B, and collect the noisy received signal y(t) at the receiving end of Node B;

[0015] Step 2.2: Remove CPP and transform the signal into the affine domain using the DAFT matrix A.

[0016] Step 2.3: Use the pilot signal to estimate the current effective channel matrix and store the CSI state in the channel state database;

[0017] Step 2.4: Extract the historical CSI sequence from the database and input it into the neural network prediction model to output the channel state at the future moment;

[0018] Step 2.5: Based on the channel prediction results, the received signal is equalized and the bit stream is demodulated. At the same time, the predicted channel state is fed back to the transmitter for parameter adjustment in the next transmission cycle.

[0019] Furthermore, in step 1.2, based on the future predicted channel state, information such as the delay-Doppler complex gain can be obtained, and the DAFT chirp parameters c1 and c2 can be adjusted. This dynamically adjusts the DAFT chirp parameters c1 and c2 to match the delay-Doppler characteristics of the channel and adaptively selects the appropriate modulation order based on different signal-to-noise ratio conditions.

[0020] Furthermore, in step 1.3, the modulation symbols are mapped to the subcarriers to form the affine Fourier domain data x, and then the DAFT matrix A is constructed according to the adjusted chirp parameters c1 and c2.

[0021]

[0022] Where F is the discrete Fourier transform matrix, Λ C Defined as

[0023]

[0024] Furthermore, in step 1.4, the IDAFT matrix A is used -1 Transform data in the affine Fourier domain

[0025]

[0026] Add the Chirp period prefix CPP to the transformed data, where CPP is:

[0027]

[0028] The above-organized AFDM data is sent as a signal on the corresponding transducer according to the traditional solution.

[0029] Furthermore, in step 2.1, the transceiver switch is switched to the node B receiving mode, and the received signal of the data transmission link is:

[0030]

[0031] Where P is the number of paths, d pis the normalized delay, ν p is the Doppler shift, ω r [n] is complex white Gaussian noise.

[0032] Furthermore, in step 2.2, the Chirp cyclic prefix of the received data is removed, and the received signal can be written as a matrix expression:

[0033] r=Hs+w (7)

[0034] in P is the number of paths, h i is the complex gain, It is the impact of CPP. Doppler shift diagonal array, is a time-delay diagonal matrix, W is Gaussian white noise;

[0035] Use DAFT matrix A to transform it, and express the input-output relationship as:

[0036]

[0037] Furthermore, in step 2.3, the corresponding pilot prior information is used for embedded channel estimation. The full diversity characteristics of the AFDM system are used to sequentially estimate the delay-Doppler and complex gain, and the effective channel matrix is ​​restored based on these parameters.

[0038] Furthermore, in step 2.4, the neural network architecture adopts a Transformer network prediction model, using an encoder-decoder approach to transform channel prediction from a sequential prediction problem into a parallel mapping problem. This step divides the estimated effective channel matrix into real and imaginary parts. The historical channel state of the P frame is input into the encoder of the Transformer network prediction model. The encoder's attention mechanism generates a key matrix K, a query matrix Q, and a value matrix V through three linear transformations, where W is a trainable linear transformation matrix:

[0039]

[0040] Calculate K T The product of Q is the attention matrix E (e) :

[0041]

[0042] The attention matrix is ​​then multiplied by the value matrix to obtain the output of the attention mechanism. At the same time, in order to avoid the gradient disappearance problem, the encoder uses residual connections and layer normalization to obtain the output:

[0043]

[0044] Where LN stands for layer normalization;

[0045] In order to further extract the channel features of each historical input, a fully connected neural network is used to obtain it, while considering the residual connection and layer normalization again. The final encoder output is:

[0046] Y(e)=LN(Z (e) +FCN(Z (e) )) (14)

[0047] For the decoder, since the channel state of the next L frames needs to be predicted, the decoder input needs to be zero-padded. In order to avoid unnecessary feature learning of the zero-padded part, the mask operation is introduced and defined as:

[0048] Mask(X)=X+tril(-inf) (15)

[0049] Where tril(-inf) represents the lower triangular matrix, and each lower triangular element is negative infinity;

[0050] Then the corresponding K, Q, and V matrices are calculated through the attention mechanism, and finally the output of the mask-attention mechanism is obtained:

[0051] Z (d) =LN(H(d)+V (d) E (d) ) (16)

[0052] Among them E (d) is the mask attention matrix:

[0053]

[0054] In order to make full use of the features provided by the encoder to extract historical information, the feature Y (e) The key matrix K and the value matrix V are calculated, and the query matrix Q is obtained by the decoder Z (d) Calculated, such a full attention mechanism can be established by multiplying the value matrix with the attention matrix to establish a feature Y (e) , taking into account the residual connection and layer normalization, we get the full attention mechanism output:

[0055]

[0056] Finally, through the fully connected neural network, the original decoder zero-filled input is overwritten with the predicted value, and the predicted output is finally obtained:

[0057]

[0058] Furthermore, in step 2.5, the channel equalization is performed based on the predicted channel state using a zero-forcing equalization method. The zero-forcing equalization method is:

[0059]

[0060] The predicted channel state is fed back to the transmitter for parameter adjustment in the next transmission cycle.

[0061] Advantages and beneficial effects of the present invention:

[0062] ① The present invention is based on an underwater acoustic adaptive transmission method for channel prediction, which uses a Transformer network prediction model to parallelly estimate future channel states, and has the characteristics of low complexity and high accuracy. ② The present invention dynamically adjusts transmission parameters based on accurate predicted channel states, and has the characteristics of flexibility and high transmission reliability. ③ The present invention adopts AFDM, which has excellent characteristics for combating double-selective fading channels, thereby achieving full diversity under time-varying channels, effectively improving link transmission stability and spectrum efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of the design scheme for the transmitting link and receiving link of the underwater acoustic half-duplex adaptive transmission system based on channel prediction;

[0064] Figure 2 Schematic diagram of the channel prediction design scheme based on the Transformer network prediction model. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is further illustrated by the following embodiments.

[0066] The present invention provides an underwater acoustic adaptive transmission method based on channel prediction, the scheme of which is as follows Figure 1 The specific steps are as follows:

[0067] Step 1: Design of transmitting and receiving links of underwater acoustic half-duplex adaptive transmission system based on channel prediction;

[0068] Step 2: Design of a channel prediction model based on Transformer.

[0069] In the first step, the designed underwater acoustic transmission method adopts AFDM, which has excellent characteristics against double-selective fading channels, thereby achieving full diversity under time-varying channels, effectively improving link transmission stability and spectrum efficiency. The schematic diagram is shown in FIG. Figure 1shown. At the transmitter end of an underwater acoustic adaptive transmission system based on channel prediction, the bits to be transmitted are first constellated and the transmission parameters are dynamically adjusted based on the predicted future channel state. The modulation symbols are then mapped onto subcarriers to form an affine Fourier domain transmit signal, and a discrete affine Fourier transform (DAFT) matrix is ​​constructed based on the channel parameters. The affine Fourier domain transmit signal is then transformed to the time domain using the IDAFT matrix, and a chirp cyclic prefix (CPP) is added to the transformed data. Finally, the organized AFDM time domain signal is transmitted over the transducer. At the data receiver end of the underwater acoustic adaptive transmission system based on AFDM, the CPP is first removed from the noisy received signal and the signal is transformed to the affine domain using the DAFT matrix. The current effective channel matrix is ​​then estimated using pilot signals, and the CSI state is stored in a channel state database. The historical CSI sequence is then extracted from the database and input into a Transformer network prediction model to obtain the channel state at the future time. Finally, based on the channel prediction results, the received signal is zero-forcing equalized and the bit stream is demodulated. The predicted channel state is fed back to the transmitter for parameter adjustment in the next transmission cycle.

[0070] In step 2, the present invention adopts the channel prediction technology based on Transformer, which not only improves the accuracy of channel prediction, but also effectively compensates for the feedback loop delay, making the system parameter configuration closer to the actual channel conditions during data transmission. Figure 2 As shown in Figure 2. It consists of two main parts: the encoder and the decoder. The encoder inputs past channel state information and effectively extracts channel features through the attention mechanism. The network architecture uses residual connections and layer normalization to prevent vanishing gradients. Ultimately, the features are input to the decoder's full attention mechanism. The decoder requires input of partially zero-padded channel states, which are then passed through the masked attention mechanism and then fed into the full attention mechanism. Through parallel mapping relationships, the decoder fully learns past channel features, ultimately overwriting the original zero-padded decoder input with the predicted value to obtain the predicted output.

[0071] Example 1:

[0072] A channel prediction-based underwater acoustic adaptive transmission method uses a Transformer network prediction model to concurrently estimate future channel states and dynamically adjust transmission parameters based on the channel states, improving the flexibility and robustness of underwater acoustic communication. The steps and details of each step are as follows:

[0073] Step 1: Design of the transmission link of the underwater acoustic adaptive transmission system based on channel prediction;

[0074] Step 2: Design scheme of receiving link of underwater acoustic adaptive transmission system based on channel prediction.

[0075] In the first step, since the underwater acoustic communication system adopts AFDM communication technology, the present invention uses MPSK / MQAM digital modulation technology, a bandwidth of 8 kHz, N = 256 subcarriers, and the adaptive system data transmission link takes a common single array element as an example, which can be expanded to multiple array elements. The following processing will be performed in the data transmission link:

[0076] Step 1.1. Switch the transceiver switch to Node A transmit mode.

[0077] In step 1.2, the DAFT chirp parameters c1 and c2 are dynamically adjusted based on the channel state information fed back by the receiver, such as delay, Doppler, and complex gain, to match the delay-Doppler characteristics of the channel, and the appropriate modulation order is adaptively selected based on different signal-to-noise ratio conditions.

[0078] Step 1.3: Map the modulation symbols to the subcarriers to form the affine Fourier domain data x, and then construct the DAFT matrix A based on the adjusted chirp parameters c1 and c2.

[0079]

[0080] Where F is the discrete Fourier transform matrix, Λ C Defined as

[0081]

[0082] Step 1.4: Use IDAFT matrix A -1 Transform data in the affine Fourier domain

[0083]

[0084] Add a Chirp Periodic Prefix (CPP) to the transformed data, where CPP is:

[0085]

[0086] The above-organized AFDM data is sent as a signal on the corresponding transducer according to the traditional solution.

[0087] In step 2, the Transformer network prediction model is used to predict the future channel state. The predicted channel state is used for channel equalization and fed back to the transmitter for adaptive parameter adjustment. The following processing is performed in the data receiving link:

[0088] Step 2.1: Switch the transceiver switch to the node B receiving mode. The received signal of the data transmission link is:

[0089]

[0090] Where P is the number of paths, d p is the normalized delay, ν p is the Doppler shift, ω r [n] is complex Gaussian white noise

[0091] Step 2.2: Remove the Chirp cyclic prefix of the received data. The received signal can be expressed as a matrix expression:

[0092] r=Hs+w (7)

[0093] in P is the number of paths, h i is the complex gain, It is the impact of CPP. Doppler shift diagonal array, is a diagonal matrix of time delays, and W is Gaussian white noise.

[0094] Use DAFT matrix A to transform it, and express the input-output relationship as:

[0095]

[0096] In step 2.3, the corresponding pilot prior information is used for embedded channel estimation. The full diversity characteristic of the AFDM system is utilized to sequentially estimate the delay-Doppler and complex gain, and the effective channel matrix is ​​restored based on these parameters.

[0097] In step 2.4, the estimated effective channel matrix is ​​divided into real and imaginary parts. The historical channel state of the P frame is input into the encoder of the Transformer network prediction model. The encoder's attention mechanism generates the key matrix K, the query matrix Q and the value matrix V through three linear transformations, where W is a trainable linear transformation matrix:

[0098]

[0099] Calculate K T The product of Q is the attention matrix E (e) :

[0100]

[0101] The attention matrix is ​​then multiplied by the value matrix to obtain the output of the attention mechanism. At the same time, in order to avoid the gradient disappearance problem, the encoder uses residual connections and layer normalization to obtain the output:

[0102]

[0103] Where LN stands for layer normalization.

[0104] In order to further extract the channel features of each historical input, a fully connected neural network is used to obtain it, while considering the residual connection and layer normalization again. The final encoder output is:

[0105] Y (e) =LN(Z (e) +FCN(Z (e) )) (14)

[0106] For the decoder, since the channel state of the next L frames needs to be predicted, the decoder input needs to be zero-padded. In order to avoid unnecessary feature learning of the zero-padded part, the mask operation is introduced and defined as:

[0107] Mask(X)=X+tril(-inf) (15)

[0108] Where tril(-inf) represents a lower triangular matrix, and each lower triangular element is negative infinity.

[0109] Then the corresponding K, Q, and V matrices are calculated through the attention mechanism, and finally the output of the mask attention mechanism is obtained:

[0110] Z (d) =LN(H(d)+V (d) E (d) ) (16)

[0111] Among them E (d) is the mask attention matrix:

[0112]

[0113] In order to make full use of the features provided by the encoder to extract historical information, the full attention mechanism between the encoder and the decoder is adopted below. Specifically, the feature Y provided by the encoder is (e) The key matrix K and the value matrix V are calculated, and the query matrix Q is obtained by the decoder Z (d) Calculated, such a full attention mechanism can be established by multiplying the value matrix with the attention matrix to establish a feature Y (e) , taking into account the residual connection and layer normalization, we get the full attention mechanism output:

[0114]

[0115] Finally, through the fully connected neural network, the original decoder zero-filled input is overwritten with the predicted value, and the predicted output is finally obtained:

[0116]

[0117] Step 2.5: Perform channel equalization on the received data according to the predicted channel state, using zero-forcing equalization as follows:

[0118]

[0119] The predicted channel state is fed back to the transmitter for parameter adjustment in the next transmission cycle.

Claims

1. A method for underwater acoustic adaptive transmission based on channel prediction, comprising the following steps: Step 1: Design of the transmission link of the underwater acoustic adaptive transmission system based on channel prediction; Step 2: Design scheme of receiving link of underwater acoustic adaptive transmission system based on channel prediction; Step 1 includes: Step 1.1: Switch the transceiver switch to the transmit / receive mode of node A. Step 1.2: Constellation mapping is performed on the bits to be transmitted, and the transmission parameters are dynamically adjusted according to the predicted future channel state. Step 1.3: Map the modulation symbols onto subcarriers to form an affine Fourier domain transmit signal. Then, based on the predicted channel state, dynamically select parameters and construct the DAFT matrix A. Step 1.4: Use IDAFT matrix A -1 The affine Fourier domain transmit signal is transformed into the time domain, and a Chirp Cyclic Prefix (CPP) is added. The AFDM time domain signal is then transmitted on the transducer according to the traditional scheme. Step 2 includes: Step 2.1: Switch the transceiver switch to the receiving mode of Node B, and collect the noisy received signal y(t) at the receiving end of Node B; Step 2.2: Remove CPP and transform the signal into the affine domain using the DAFT matrix A. Step 2.3: Use the pilot signal to estimate the current effective channel matrix and store the CSI state in the channel state database; Step 2.4: Extract the historical CSI sequence from the database and input it into the neural network prediction model to output the channel state at the future moment; Step 2.5: Based on the channel prediction results, the received signal is equalized and the bit stream is demodulated. At the same time, the predicted channel state is fed back to the transmitter for parameter adjustment in the next transmission cycle.

2. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 1.2, based on the channel state information fed back by the receiver, the DAFT chirp parameters c1 and c2 are dynamically adjusted to match the delay-Doppler characteristics of the channel, and the appropriate modulation order is adaptively selected according to different signal-to-noise ratio conditions.

3. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 1.3, the modulation symbols are mapped to the subcarriers to form the affine Fourier domain data x, and then the DAFT matrix A is constructed based on the adjusted chirp parameters c1 and c2. Where F is the discrete Fourier transform matrix, Λ C Defined as 4. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 1.4, the IDAFT matrix A is used -1 Transform data in the affine Fourier domain Add the Chirp period prefix CPP to the transformed data, where CPP is: The above-organized AFDM data is sent as a signal on the corresponding transducer according to the traditional solution.

5. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 2.1, the transceiver switch is switched to the node B receiving mode, and the received signal of the data transmission link is: Where P is the number of paths, d p is the normalized delay, ν p is the Doppler shift, ω r [n] is complex white Gaussian noise.

6. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 2.2, the Chirp cyclic prefix of the received data is removed, and the received signal can be written as a matrix expression: r=Hs+w (7) in P is the number of paths, h i is the complex gain, It is the impact of CPP. Doppler shift diagonal array, is a time-delay diagonal matrix, W is Gaussian white noise; Use DAFT matrix A to transform it, and express the input-output relationship as:

7. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 2.3, the corresponding pilot prior information is used for embedded channel estimation. The full diversity characteristics of the AFDM system are used to estimate the delay Doppler and complex gain in turn, and the effective channel matrix is ​​restored based on these parameters.

8. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 2.4, the estimated effective channel matrix is ​​divided into real and imaginary parts. The historical channel state of the P frame is input into the encoder of the Transformer network prediction model. The encoder's attention mechanism generates the key matrix K, the query matrix Q, and the value matrix V through three linear transformations, where W is a trainable linear transformation matrix: Calculate K T The product of Q is the attention matrix E (e) : The attention matrix is ​​then multiplied by the value matrix to obtain the output of the attention mechanism. At the same time, in order to avoid the gradient disappearance problem, the encoder uses residual connections and layer normalization to obtain the output: Where LN stands for layer normalization; In order to further extract the channel features of each historical input, a fully connected neural network is used to obtain it, while considering the residual connection and layer normalization again. The final encoder output is: Y(e)=LN(Z (e) +FCN(Z (e) )) (14) For the decoder, since the channel state of the next L frames needs to be predicted, the decoder input needs to be zero-padded. In order to avoid unnecessary feature learning of the zero-padded part, the mask operation is introduced and defined as: Mask(X)=X+tril(-inf) (15) Where tril(-inf) represents the lower triangular matrix, and each lower triangular element is negative infinity; Then the corresponding K, Q, and V matrices are calculated through the attention mechanism, and finally the output of the mask-attention mechanism is obtained: Z (d) LN(H(d)+V (d) E (d) ) (16) Among them E (d) is the mask attention matrix: In order to make full use of the features provided by the encoder to extract historical information, the feature Y (e) The key matrix K and the value matrix V are calculated, and the query matrix Q is obtained by the decoder Z (d) Calculated, such a full attention mechanism can be established by multiplying the value matrix with the attention matrix to establish a feature Y (e) , taking into account the residual connection and layer normalization, we get the full attention mechanism output: Finally, through the fully connected neural network, the original decoder zero-filled input is overwritten with the predicted value, and the predicted output is finally obtained:

9. The underwater acoustic adaptive transmission method based on channel prediction according to claim 1, characterized in that: In step 2.5, channel equalization is performed on the received data according to the predicted channel state, where zero-forcing equalization is used as follows: The predicted channel state is fed back to the transmitter for parameter adjustment in the next transmission cycle.

Images