A deep learning channel estimation method and device suitable for underwater acoustic sensor networks

Abstract

The application relates to a deep learning channel estimation method and device suitable for an underwater acoustic sensor network. The method utilizes OFDM pilot data to establish a sparse signal model, and obtains a real-value sparse model by performing real-value transformation on the sparse signal model. An approximate message passing (AMP) estimation framework is introduced to realize preliminary recovery of channel information. Further, a model-driven deep learning framework is proposed. An ST-LAMP network based on AMP without considering sparse prior and a GGM-LAMP network considering sparse prior are respectively established. Network training is respectively performed according to a predetermined strategy. Optimal matching parameters are learned through data learning and are updated and fixed. Correct channel estimation values can be output through input of new measurement values. The problem that traditional empirical parameter setting cannot be applied to high complexity of an underwater acoustic channel, leading to deviation of channel estimation from an actual channel, is solved. The precision of channel estimation is improved. The channel estimation method has the characteristics of low complexity and strong self-adaptability.

Description

Technical Field

[0001] This application relates to the field of channel estimation technology, and mainly to a deep learning channel estimation method and apparatus suitable for underwater acoustic sensor networks. Background Technology

[0002] Underwater acoustic channels are extremely complex, and this complexity is even more pronounced in underwater acoustic sensor networks. The communication links in underwater acoustic sensor networks place higher demands on underwater acoustic channel estimation methods. Traditional channel estimation methods typically make certain assumptions about the state model and sparse structure of the underwater acoustic channel. However, it is difficult to simulate the actual channel in underwater acoustic sensor networks using a unified model. The actual channel usually deviates from the preset channel model or sparse structure. When model deviations occur, traditional methods cannot be applied to channel estimation in the more complex underwater acoustic sensor networks.

[0003] Time-varying channel estimation has always been a hot and challenging research topic in underwater acoustic communication. Currently, representative methods include model-based methods and compressed sensing-based methods. Model-based methods track the channel evolution state model and employ techniques such as Kalman filtering to achieve dynamic channel estimation. Compressed sensing-based methods, through assumptions such as co-sparseness among different channels, can achieve joint estimation of dynamically sparse channels. However, the unpredictability of actual channel models leads to model mismatch issues in practical applications, and these methods also have high computational complexity. For the high real-time and low-latency signal processing requirements of underwater acoustic networks, the performance of these traditional methods is limited. Therefore, designing a low-complexity, highly adaptive underwater acoustic channel estimation method has become an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems in the prior art, the inventors have proposed a deep learning channel estimation method and apparatus suitable for underwater acoustic sensor networks.

[0005] According to one aspect of the present invention, a deep learning channel estimation method suitable for underwater acoustic sensor networks is proposed, comprising the following steps:

[0006] S1: Use OFDM pilot signals to establish a sparse signal model, perform a real-valued transformation on the sparse signal model, and establish a real-valued sparse model;

[0007] S2: Introduce the Approximate Message Passing Algorithm (AMP), establish a deep learning neural network model based on AMP, use the training dataset, and train the deep learning neural network model using a model-based method, and configure and update the learnable parameters of the deep learning neural network model.

[0008] S3: Deploy the trained deep learning neural network model online, input new observations into the network, and obtain the estimated parameters of the sparse underwater acoustic channel based on the trained parameters.

[0009] Traditional channel estimation methods use empirical parameters in the Approximate Message Passing Algorithm (AMP) to obtain channel estimates for underwater acoustic channels. However, fixed empirical parameter settings cannot adapt to the high complexity of underwater acoustic channels, and model mismatch may lead to significant errors between the estimated channel values ​​and the actual channel. Building a deep learning neural network model based on AMP, and replacing the fixed empirical parameters with the optimal parameters obtained from network training, can output underwater acoustic channel estimates that are closer to reality, improving the adaptive optimization capability of traditional algorithms in complex environments. The establishment of a real-valued sparse model solves the problem that sparse signal models are complex numbers, making them unsuitable for deep learning extensions.

[0010] Specifically, in S1, the pilot data in the OFDM data block As an observed signal, a sparse signal model is constructed. , where y is the observed signal, h is the channel impulse response information with sparse characteristics, X is the diagonal matrix composed of pilot sequences, F is the Fourier transform subarray corresponding to the unknown pilot and channel, w is Gaussian noise, and T is the total number of network layers.

[0011] By performing a real-valued transformation on the sparse signal model, a real-valued sparse model is obtained. The real-valued sparse model is as follows: ,

[0012] Real-valued transformation of sparse signal models facilitates deep learning extensions, and the transformation dimension becomes... , , To simplify the definition , M is the number of pilots, and N is the channel length.

[0013] Specifically, based on the real-valued sparse model, sparse channel recovery can be modeled as a ridge regression problem (LASSO), i.e.

[0014] ;

[0015] By introducing the Approximate Message Passing Learning (AMP) algorithm and through empirical parameter settings, a preliminary recovery of sparse channel information in the underwater acoustic channel can be obtained.

[0016] Based on the AMP algorithm, a series of iterations are performed, where the steps of the t-th iteration are as follows:

[0017] ;

[0018] in Contraction function, For Onsager correction items, To estimate the noise, for Estimate the standard deviation of the noise.

[0019] Onsager correction terms can achieve measurement matrix The decoupling, and thus the estimation error for each vector component follows a Gaussian distribution, i.e. The shrinkage function is equivalent to a denoising operation, that is, from the noise observations Sparse channel recovery .

[0020] Specifically, initialization is required at the start of the iteration, with the initialization parameters being: .

[0021] Specifically, the AMP algorithm is extended to process batch data. Assuming that D batches of data are processed simultaneously, the system model is extended as follows: The iteration of AMP is

[0022] ;

[0023] in , , .

[0024] In a specific embodiment, the deep learning neural network model is an ST-LAMP network based on a soft threshold shrinkage function, and the soft threshold shrinkage function of the ST-LAMP network is:

[0025] ,

[0026] The soft threshold function is simplified as follows: Corresponding to the above function, we can obtain , For threshold control parameters, for The estimated standard deviation.

[0027] The set of learnable parameters for the ST-LAMP network is as follows:

[0028] ;

[0029] in To match the filter matrix, It is a contraction function. Let t be the total number of network layers and t be the corresponding number of network layers. The soft thresholding function has been proven to be the minimum maximum optimal MSE function, which can obtain the minimum maximum mean square error. It is applicable to any sparse underwater acoustic channel. Regardless of any prior values, the ST-LAMP network based on the soft thresholding shrinkage function can be applied to any underwater acoustic channel without considering sparse priors.

[0030] The ST-LAMP network is an expansion of the AMP algorithm, with each iteration corresponding to one layer of the network, combining the features of AMP. Replace with learnable matrix Shrinkage parameters As learnable parameters, each layer of the network is obtained through network training. and The optimal value is obtained by replacing the fixed parameters with the learned parameters, which solves the problem that traditional channel estimation methods use fixed parameters, while underwater acoustic channels are time-varying and highly complex, making it difficult to simulate them with a unified model.

[0031] In a specific embodiment, the deep learning neural network model is a GGM-LAMP network based on the MMSE shrinkage function. The specific method for constructing the GGM-LAMP network based on the MMSE shrinkage function is as follows:

[0032] Establish a channel prior model;

[0033] The MMSE contraction function is derived based on the channel prior model.

[0034] A GGM-LAMP network is established based on the MMSE shrinkage function.

[0035] By introducing prior channel information and observing channel data, the network can learn approximate channel probability parameters. After the parameters are learned, the receiver parameters can be configured in a specific working environment, which greatly reduces the amount of parameter learning and improves the network performance. It is suitable for sparse underwater acoustic channel estimation scenarios with insufficient data.

[0036] In a further embodiment, the deep learning neural network model is trained using a model-based method with a training dataset, wherein the training dataset is:

[0037] ;

[0038] D represents the total number of training data batches, and d represents the specific batch.

[0039] The initial dataset for the network is , Pilot observations and real sample data Similar to the labeled data and feature datasets in supervised learning frameworks, the network is trained using pilot observation signals and real channel data, and the network obtains fixed optimal parameters after training.

[0040] In a specific embodiment, the channel prior model is established by using a generalized Gaussian mixture distribution as the channel prior model.

[0041] The joint edge distribution of the channel vectors is represented as follows:

[0042] .

[0043] In a further embodiment, the set of learnable parameters for the GGM-LAMP network is as follows:

[0044] .

[0045] The set of learnable parameters includes all location parameters from the prior model, where, Indicates the sparsity of the channel. Let be the number of Gaussian components. These represent the weights, mean, and variance of the corresponding Gaussian components, respectively.

[0046] In a specific embodiment, the MMSE contraction function based on the prior channel model is derived as follows:

[0047] Based on the MMSE criterion, the posterior probability estimate is calculated as follows: This function operates on each vector element individually, i.e., the input... Output The MMSE estimate is expressed as:

[0048] ,

[0049] Based on the MMSE estimation function, the learnable optimal MMSE contraction function is expressed as:

[0050] .

[0051] Based on the above MMSE optimal contraction function, where

[0052] ,

[0053] Therefore, the learnable parameter set in the GGM-LAMP network is defined as follows:

[0054] , , , , . , .

[0055] In a further embodiment, the relevant Onsager correction term requires calculating the derivative of the contraction function, based on... First calculate the variance term. Furthermore, according to We can obtain:

[0056] .

[0057] In a specific embodiment, the learnable parameters of the deep learning neural network model are updated as follows: Gradient calculations are performed on the learnable parameters using the loss function, and the gradients are calculated as follows:

[0058] The gradient descent algorithm is used to update the parameters.

[0059] In a further embodiment, the loss function is:

[0060] The loss function is: .

[0061] In a specific embodiment, a model-based method is used to train the deep learning neural network model. The training strategy used in the network training is as follows: S21: Define untied network training and tied network training;

[0062] In untied network training, the set of learnable parameters is defined as follows:

[0063] ,

[0064] Each layer of the network needs to be trained, and the parameters obtained for each layer are different.

[0065] In training a tied network, the set of learnable parameters is defined as follows: ,

[0066] After completing the learning Fixed for all layers, Train each layer separately;

[0067] S22: Use a layer-by-layer training method in both untied and tied network training;

[0068] S23: Train the ST-LAMP network using an untied network and train the GGM-LAMP network using a tied network.

[0069] Untied network training and tied network training are distinguished by whether the parameter set is learned independently. Untied network training requires parameter setting to be adjusted in each layer of the network. Training is performed in the training of the tied network. It will not be updated in every layer; the training parameters... The dimensionality is high, especially for untied network training, where the number of training parameters needs to be multiplied by the number of network layers. Therefore, end-to-end network structures are prone to parameter overfitting. Thus, a training strategy between layer-by-layer and function-by-function is established. Untied network training requires more training and computational storage resources than tied network training. ST-LAMP networks, which have no prior information, can achieve better estimation results when trained using untied networks. For GGM-LAMP networks, when data volume is insufficient, tied network training yields optimal performance.

[0070] Secondly, this application also proposes a computer system for an electronic device, the electronic device comprising: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the deep learning channel estimation method for underwater acoustic sensor networks described in the first aspect.

[0071] Thirdly, this application also proposes a computer-readable storage medium having stored thereon one or more computer programs that, when executed by a computer processor, implement the method described in the first aspect of this application.

[0072] Fourthly, this application also proposes a deep learning channel estimation device suitable for underwater acoustic sensor networks, the device comprising the following systems:

[0073] Communication signal acquisition module: configured to transmit signals and receive observation signals through the time-varying underwater acoustic channel;

[0074] Data preprocessing module: Configured to build sparse signal models, perform real-value transformation on sparse signal models to obtain real-value recovery models, which are convenient for deep learning extensions;

[0075] Channel estimation module: Introduces the Approximate Message Passing Algorithm (APM), establishes a deep learning neural network model based on AMP, obtains the optimal solution of learnable parameters through network training, and distributes the configuration.

[0076] Input / output module: Configured to take in observed signals and output correct channel estimates.

[0077] In summary, this application includes at least the following beneficial technical effects:

[0078] 1. This invention improves upon traditional channel estimation methods based on OFDM pilots and the AMP algorithm. Traditional channel estimation methods obtain the estimated value of the underwater acoustic channel by observing signals through fixed empirical parameter settings. However, underwater acoustic channels are time-varying and unpredictable, and traditional channel estimation methods suffer from model deviation, making them unsuitable for channel estimation in complex underwater acoustic sensor networks. Based on the AMP algorithm, this invention designs ST-LAMP networks without sparse priors and GGM-LAMP networks considering sparse priors. The optimal solution of learnable parameters is obtained through network training and then distributed for configuration, solving the performance limitation problem caused by fixed parameters in traditional methods.

[0079] 2. Establish a prior channel model and derive the MMSE contraction function based on the prior channel model. Based on the MMSE contraction function and the set of learnable parameters, establish a GGM-LAMP network. The GGM-LAMP network can learn approximate channel probability parameters. After the parameters are learned, the receiver parameters can be configured in a specific working environment. It is applicable to scenarios with insufficient data, achieves learning of fewer parameters, and obtains a significant performance improvement. Attached Figure Description

[0080] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of this application. Other embodiments and many anticipated advantages of these embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0081] Figure 1 This is a flowchart illustrating a deep learning channel estimation method applicable to underwater acoustic sensor networks according to this application.

[0082] Figure 2 This is a schematic diagram comparing the channel estimation performance of a deep learning channel estimation method applicable to underwater acoustic sensor networks under a simulated channel.

[0083] Figure 3 This is a schematic diagram comparing the channel estimation performance of a deep learning channel estimation method applicable to underwater acoustic sensor networks under measured channel conditions.

[0084] Figure 4a A schematic diagram of the structure of a computer system 400 of an electronic device according to an embodiment of the present invention is shown.

[0085] Figure 4b This is a schematic diagram of the structure of a deep learning channel estimation device applicable to underwater acoustic sensor networks according to this application. Detailed Implementation

[0086] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0087] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0088] Figure 1 This is a flowchart illustrating a deep learning channel estimation method applicable to underwater acoustic sensor networks, as described in this application. (Refer to reference...) Figure 1 A deep learning channel estimation method suitable for underwater acoustic sensor networks includes the following steps:

[0089] S1: Use OFDM pilot signals to establish a sparse signal model, perform a real-valued transformation on the sparse signal model, and establish a real-valued sparse model;

[0090] S2: Introduce the Approximate Message Passing Algorithm (AMP), establish a deep learning neural network model based on AMP, use the training dataset, and train the deep learning neural network model using a model-based method, and configure and update the learnable parameters of the deep learning neural network model.

[0091] S3: Deploy the trained deep learning neural network model online, input new observations into the network, and obtain the estimated parameters of the sparse underwater acoustic channel based on the trained parameters.

[0092] Traditional channel estimation methods use empirical parameters in the Approximate Message Passing Algorithm (AMP) to obtain channel estimates for underwater acoustic channels. However, fixed empirical parameter settings cannot adapt to the high complexity of underwater acoustic channels, and model mismatch may lead to significant errors between the estimated channel values ​​and the actual channel. Building a deep learning neural network model based on AMP, and replacing the fixed empirical parameters with the optimal parameters obtained from network training, can output underwater acoustic channel estimates that are closer to reality, improving the adaptive optimization capability of traditional algorithms in complex environments. The establishment of a real-valued sparse model solves the problem that sparse signal models are complex numbers, making them unsuitable for deep learning extensions.

[0093] Specifically, establishing a sparse signal model using OFDM pilot signals involves the following: the signal transmitted by the transmitter of the underwater acoustic OFDM communication system is... After passing through the time-varying underwater acoustic channel The total number of paths is , The path amplitude and path delay are time-varying.

[0094] Received signal ;

[0095] Define the discrete vector related to the pilot as follows: ,

[0096] , For pilot number, This is the channel length.

[0097] [0045 Specifically, in S1, the pilot data in the OFDM data block As an observed signal, a sparse signal model is constructed. , where y is the observed signal, h is the channel impulse response information with sparse characteristics, X is the diagonal matrix composed of pilot sequences, F is the Fourier transform subarray corresponding to the unknown pilot and channel, w is Gaussian noise, and T is the total number of network layers.

[0098] By performing a real-valued transformation on the sparse signal model, a real-valued sparse model is obtained. The real-valued sparse model is as follows:

[0099] ,

[0100] Real-valued transformation of sparse signal models facilitates deep learning extensions, and the transformation dimension becomes...

[0101] , , To simplify the definition , M is the number of pilots, and N is the channel length.

[0102] Specifically, based on the real-valued sparse model, sparse channel recovery can be modeled as a ridge regression problem (LASSO), i.e.

[0103] ;

[0104] By introducing the Approximate Message Passing Learning (AMP) algorithm and through empirical parameter settings, a preliminary recovery of sparse channel information in the underwater acoustic channel can be obtained.

[0105] Based on the AMP algorithm, a series of iterations are performed, where the steps of the t-th iteration are as follows:

[0106] ;

[0107] in It is a contraction function. For Onsager correction items, To estimate the noise, for Estimate the standard deviation of the noise.

[0108] Onsager correction terms can achieve measurement matrix The decoupling, and thus the estimation error for each vector component follows a Gaussian distribution, i.e. The shrinkage function is equivalent to a denoising operation, that is, from the noise observations Sparse channel recovery .

[0109] Specifically, initialization is required at the start of the iteration, with the initialization parameters being: .

[0110] Specifically, the AMP algorithm is extended to process batch data. Assuming that D batches of data are processed simultaneously, the system model is extended as follows: The iteration of AMP is

[0111] ;

[0112] in , , .

[0113] In a specific embodiment, the deep learning neural network model is an ST-LAMP network based on a soft threshold shrinkage function, and the soft threshold shrinkage function of the ST-LAMP network is: ,

[0114] The soft threshold function is simplified as follows: Corresponding to the above function, we can obtain , , For threshold control parameters, for The estimated standard deviation.

[0115] The set of learnable parameters for the ST-LAMP network is as follows:

[0116] ;

[0117] in To match the filter matrix, Let T be the total number of network layers and t be the corresponding number of network layers. The soft threshold function has been proven to be the minimum maximum optimal MSE function, which can obtain the minimum maximum mean square error. It is applicable to any sparse underwater acoustic channel. Regardless of any prior values, the ST-LAMP network based on the soft threshold shrinkage function can be applied to any underwater acoustic channel without considering sparse priors.

[0118] The ST-LAMP network is an expansion of the AMP algorithm, with each iteration corresponding to one layer of the network, combining the features of AMP. Replace with learnable matrix Shrinkage parameters As learnable parameters, each layer of the network is obtained through network training. and The optimal value is obtained by replacing the fixed parameters with the learned parameters, which solves the problem that traditional channel estimation methods use fixed parameters, while underwater acoustic channels are time-varying and highly complex, making it difficult to simulate them with a unified model.

[0119] In a specific embodiment, the deep learning neural network model is a GGM-LAMP network based on the MMSE shrinkage function. The specific method for constructing the GGM-LAMP network based on the MMSE shrinkage function is as follows:

[0120] Establish a channel prior model;

[0121] The MMSE contraction function is derived based on the channel prior model.

[0122] A GGM-LAMP network is established based on the MMSE shrinkage function.

[0123] By introducing prior channel information and observing channel data, the network can learn approximate channel probability parameters. After the parameters are learned, the receiver parameters can be configured in a specific working environment, which greatly reduces the amount of parameter learning and improves the network performance. It is suitable for sparse underwater acoustic channel estimation scenarios with insufficient data.

[0124] In a further embodiment, the deep learning neural network model is trained using a model-based method with a training dataset, wherein the training dataset is:

[0125] ;

[0126] D represents the total number of training data batches, and d represents the specific batch.

[0127] The initial dataset for the network is Pilot observations and real sample data Similar to the labeled data and feature datasets in supervised learning frameworks, the network is trained using pilot observation signals and real channel data, and the network obtains fixed optimal parameters after training.

[0128] In a specific embodiment, the channel prior model is established by using a generalized Gaussian mixture distribution as the channel prior model.

[0129] The joint edge distribution of the channel vectors is represented as follows: .

[0130] In a further embodiment, the set of learnable parameters for the GGM-LAMP network is as follows:

[0131] .

[0132] The set of learnable parameters includes all location parameters from the prior model, where, Indicates the sparsity of the channel. Let be the number of Gaussian components. These represent the weights, mean, and variance of the corresponding Gaussian components, respectively.

[0133] By setting different parameters, it can be represented as different distributions, such as When the model degenerates into a two-state Gaussian mixture distribution, when At that time, the model degenerates into a Bernoulli Gaussian distribution.

[0134] In a specific embodiment, the MMSE contraction function based on the prior channel model is derived as follows:

[0135] Based on the MMSE criterion, the posterior probability estimate is calculated as follows: This function operates on each vector element individually, i.e., the input... Output The MMSE estimate is expressed as:

[0136] ,

[0137] Based on the MMSE estimation function, the learnable optimal MMSE contraction function is expressed as:

[0138] .

[0139] Based on the above MMSE optimal contraction function, where

[0140] ,

[0141] Therefore, the learnable parameter set in the GGM-LAMP network is defined as follows:

[0142] , , ,

[0143] , . , .

[0144] In a further embodiment, the relevant Onsager correction term requires calculating the derivative of the contraction function, based on... First calculate the variance term. Furthermore, according to We can obtain:

[0145] .

[0146] By incorporating the aforementioned MMSE optimal contraction function and the set of learnable parameters into the network, a new GGM-LAMP network can be established. This new network introduces new prior parameters, which are initially unknown. These parameters are updated through learning from training data. Unlike traditional AMP methods and ST-LAMP networks, this network introduces a prior model of the channel. By observing channel data, the network can learn approximate channel probability parameters. After parameter learning is complete, receiver parameter configuration can be achieved under a specific operating environment.

[0147] Specifically, before network training, the parameters are initialized as follows: The weight is The mean and variance are both 0.

[0148] In a specific embodiment, the learnable parameters of the deep learning neural network model are updated in the following way:

[0149] The gradients of the learnable parameters are calculated using the loss function, and the gradients are calculated as follows: The gradient descent algorithm is used to update the parameters.

[0150] In a further embodiment, the loss function is:

[0151] The loss function is: .

[0152] In a specific embodiment, a model-based method is used to train the deep learning neural network model. The training strategy used in the network training is as follows: S21: Define untied network training and tied network training;

[0153] In untied network training, the set of learnable parameters is defined as follows: ,

[0154] Each layer of the network needs to be trained, and the parameters obtained for each layer are different.

[0155] In the training of a tied network, the set of learnable parameters is defined as follows:

[0156] ,

[0157] After completing the learning Fixed for all layers, Training is performed separately for each layer; S22: Layer-by-layer training is used in both untied and tied network training;

[0158] S23: Train the ST-LAMP network using an untied network and train the GGM-LAMP network using a tied network.

[0159] Untied network training and tied network training are distinguished by whether the parameter set is learned independently. Untied network training requires parameter setting to be adjusted in each layer of the network. Training is performed in the training of the tied network. It will not be updated in every layer; the training parameters... The dimensionality is high, especially for untied network training, where the number of training parameters needs to be multiplied by the number of network layers. Therefore, end-to-end network structures are prone to parameter overfitting. Thus, a training strategy between layer-by-layer and function-by-function is established. Untied network training requires more training and computational storage resources than tied network training. ST-LAMP networks, which have no prior information, can achieve better estimation results when trained using untied networks. For GGM-LAMP networks, when data volume is insufficient, tied network training yields optimal performance.

[0160] Specifically, this method employs a layer-by-layer training approach. Taking the training of the i-th layer network as an example, in... Based on the completion of layer training, A new training cycle begins. The entire subnetwork, including layer i, is trained simultaneously, and the learned parameters are globally adjusted. During the training of each layer, the parameters of each layer are simultaneously fine-tuned locally, first fixing the parameter set. Training parameters After training is completed individually, the parameter set is then... Conduct joint training. The next step is to establish a fixed location. Learn parameters based on After training the network separately, all parameters are trained simultaneously. For The training steps for the tied network are similar to those described above, but due to the reduction in training parameters, the training steps need to be simplified. At that time, train according to the above steps to obtain In subsequent training Fixed, no longer applicable Conduct individual training.

[0161] Figure 2This is a schematic diagram comparing the channel estimation performance of a deep learning channel estimation method applicable to underwater acoustic sensor networks under simulated channel conditions, as shown in this application. Figure 2 As shown, the proposed GGM-LAMP network achieves the best channel estimation performance, followed by the ST-LAMP network. This indicates that the introduction of a priori channel model can greatly improve the estimation performance. In addition, the powerful learning ability of deep learning technology can greatly improve the adaptive optimization ability of traditional algorithms in complex environments.

[0162] In a specific embodiment, the experiments compared the OMP estimation algorithm and the AMP estimation algorithm, with sparsity set in OMP. Setting empirical parameters in AMP To demonstrate the impulse suppression and channel estimation performance of the algorithm, initial shrinkage parameters are set for ST-LAMP. For GGM-LAMP, the number of Gaussian components is set to... Initialize sparsity Weight .

[0163] The simulation setup uses a QPSK modulated CP-OFDM system. The simulation channel adopts a multipath channel model with 15 discrete multipath channels. The arrival time follows an exponential distribution with an average delay of 15ms. The path amplitude attenuation follows a Rayleigh distribution. The average power gradually decreases with the delay, and the total power attenuation is 20dB.

[0164] In the simulated channel, a publicly available channel model was used to generate 10⁵ training samples and 20,000 test samples. The training data was generated under an SNR of 10 dB. After training, the network was validated online, with validation data covering an SNR range from 0 dB to 20 dB.

[0165] In the specific experiment, the simulated channel adopted a 15-path discrete multipath channel model. The arrival time followed an exponential distribution with an average delay of 15ms. The path amplitude attenuation followed a Rayleigh distribution, and the average power gradually decreased with the delay, with a total power attenuation of 20 dB. Sufficient training samples could be generated using the simulated channel.

[0166] The experimental channel uses the publicly available dataset WaterMark. This dataset is used for the design of the communication physical layer. The method is validated using the NOF1 channel. This dataset contains a total of 15,000 samples, of which 10,000 are used as training samples and 5,000 as test samples.

[0167] Figure 3 This is a schematic diagram comparing the channel estimation performance of a deep learning channel estimation method applicable to underwater acoustic sensor networks under measured channel conditions, as shown in the figure. Figure 3 As shown, the training samples in the measured channel are insufficient. However, the experimental results show that the GGM-LAMP network has better estimation performance, followed by the ST-LAMP network. The model-driven deep learning technology reduces the system's dependence on the amount of data and can obtain better estimation performance with fewer samples. This advantage reduces the requirements for data collection capabilities in the sensor network.

[0168] Depend on Figure 2 and Figure 3 As shown, the ST-LAMP network model and the GGM-LAMP network model outperform the channel estimation obtained by traditional algorithms, and have better robustness, reduced computational complexity, and improved estimation performance.

[0169] Figure 4a A schematic diagram of the structure of a computer system 400 of an electronic device according to an embodiment of the present invention is shown, such as... Figure 4a As shown, the computer system 400 includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 402 or programs loaded from storage section 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the system 400. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0170] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a liquid crystal display (LCD) and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card and a modem, etc. The communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0171] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit (CPU) 401, it performs the functions defined in the methods of this application. It should be noted that the computer-readable storage medium described in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0172] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0173] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0174] The modules described in the embodiments of this application can be implemented in software or hardware. The described units can also be located in a processor, and the names of these units do not necessarily limit the unit itself.

[0175] Embodiments of the present invention also relate to a computer-readable storage medium storing a computer program that, when executed by a computer processor, performs the methods described above. The computer program includes program code for performing the methods shown in the flowchart. It should be noted that the computer-readable medium of this application may be a computer-readable signal medium, a computer-readable medium, or any combination of the two.

[0176] As an implementation of the above-described method, this application provides an embodiment of a deep learning channel estimation device suitable for underwater acoustic sensor networks. This device embodiment corresponds to the method embodiment of the first aspect. Figure 4b This is a schematic diagram of the structure of a deep learning channel estimation device suitable for underwater acoustic sensor networks, as described in this application. Figure 4b As shown, the deep learning channel estimation device for underwater acoustic sensor networks proposed in this application includes the following modules:

[0177] Communication signal acquisition module 501: configured to transmit signals and receive observation signals through a time-varying underwater acoustic channel;

[0178] Data preprocessing module 502: configured to establish a sparse signal model, perform real-value transformation on the sparse signal model to obtain a real-value recovery model, which is convenient for deep learning extension;

[0179] Channel estimation module 503: Introduces the Approximate Message Passing Algorithm (APM), establishes a deep learning neural network model based on AMP, obtains the optimal solution of learnable parameters through network training, and distributes the configuration.

[0180] Input / output module 504: Configured to input observed signals and output correct channel estimates.

[0181] This application discloses a deep learning channel estimation method and apparatus suitable for underwater acoustic sensor networks. Based on traditional channel estimation methods, this method establishes an ST-LAMP network model without sparse priors and a GGM-LAMP network model considering sparse priors. By training learnable parameters, optimal parameters are obtained and distributed for configuration, thereby improving performance and obtaining channel estimates with smaller errors. The GGM-LAMP network is suitable for scenarios with insufficient data, reducing the number of parameter training steps and achieving good estimation performance even with limited data. The advantage of this invention lies in utilizing model-driven deep learning technology, which can fully utilize sensor data to enhance the adaptability of the channel estimator to more complex marine environments.

[0182] In the description of this application, it should be understood that the terms "upper," "lower," "inner," "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and for simplification, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The simple fact that certain measures are recited in mutually different dependent claims does not indicate that combinations of these measures cannot be used for improvement. Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A deep learning channel estimation method suitable for underwater acoustic sensor networks, characterized in that, Includes the following steps: S1: Use OFDM pilot signals to establish a sparse signal model, perform a real-valued transformation on the sparse signal model, and establish a real-valued sparse model; S2: Introduce the Approximate Message Passing Algorithm (AMP), establish a deep learning neural network model based on AMP, and use a model-based method to train the deep learning neural network model using a training dataset. Configure and update the learnable parameters of the deep learning neural network model. Specifically, the training strategy used in the model-based method for training the deep learning neural network model is as follows: S21: Define untied network training and tied network training; in untied network training, define the learnable parameter set as follows: , Represents the contraction function. Represents the matched filter matrix. This represents the total number of network layers, where t represents the corresponding network layer number. Each layer of the network needs to be trained, and the parameters obtained in each layer are different; in the training of the tilde network, the set of learnable parameters is defined as follows: After completing the learning Fixed for all layers, Train each layer separately; S22: Use a layer-by-layer training method in both untied and tied network training; S23: The ST-LAMP network is trained using an untied network, and the GGM-LAMP network is trained using a tied network; S3: Deploy the trained deep learning neural network model online, input new observations into the network, and obtain the estimated parameters of the sparse underwater acoustic channel based on the trained parameters.

2. The deep learning channel estimation method for underwater acoustic sensor networks according to claim 1, characterized in that, The deep learning neural network model is an ST-LAMP network based on a soft threshold shrinkage function, and the soft threshold shrinkage function of the ST-LAMP network is: The soft threshold function is simplified as follows: Corresponding to the above function, we can obtain , , For threshold control parameters, for The estimated standard deviation, This represents the observed noise value.

3. The deep learning channel estimation method for underwater acoustic sensor networks according to claim 1, characterized in that, The deep learning neural network model is a GGM-LAMP network based on the MMSE shrinkage function. The specific method for constructing the GGM-LAMP network based on the MMSE shrinkage function is as follows: Establish a channel prior model; The MMSE contraction function is derived based on the channel prior model. A GGM-LAMP network is established based on the MMSE shrinkage function.

4. The deep learning channel estimation method for underwater acoustic sensor networks according to claim 3, characterized in that, The specific method for establishing the channel prior model is as follows: a generalized Gaussian mixture distribution is used as the prior model for the channel; Wherein, the joint edge distribution of channel vectors represents , Indicates the sparsity of the channel. This indicates the number of Gaussian components. For channel length, .

5. The deep learning channel estimation method for underwater acoustic sensor networks according to claim 4, characterized in that, The MMSE contraction function based on the prior channel model is derived as follows: Based on the MMSE criterion, the posterior probability estimate is calculated as follows: This function operates on each vector element individually, i.e., the input... Output The MMSE estimate is expressed as: Based on the MMSE estimation function, the learnable optimal MMSE contraction function is expressed as: ,in This indicates the weight of the corresponding Gaussian component. This represents the mean of the corresponding Gaussian component. This represents the variance of the corresponding Gaussian component. , , , , , .

6. The deep learning channel estimation method for underwater acoustic sensor networks according to claim 5, characterized in that, The learnable parameters of the deep learning neural network model are updated by: calculating the gradient of each learnable parameter using the loss function, and calculating the gradient for each parameter as follows. The gradient descent algorithm is used to update the parameters.

7. A computer system for an electronic device, the electronic device comprising: Processor, and memory storing computer program instructions; The processor reads and executes the computer program instructions to implement the deep learning channel estimation method for underwater acoustic sensor networks as described in any one of claims 1-6.

8. A computer-readable storage medium having one or more computer programs stored thereon, characterized in that, When the one or more computer programs are executed by a computer processor, they perform the method according to any one of claims 1 to 6.

9. A deep learning channel estimation device suitable for underwater acoustic sensor networks, characterized in that, The device includes: Communication signal acquisition module: configured to transmit signals and receive observation signals through the time-varying underwater acoustic channel; Data preprocessing module: Configured to build sparse signal models, perform real-value transformation on sparse signal models to obtain real-value recovery models, which are convenient for deep learning extensions; Channel estimation module: Introducing the Approximate Message Passing Algorithm (APM), a deep learning neural network model based on APM is established. Using a training dataset, a model-based method is employed to train the deep learning neural network model. The optimal solution for learnable parameters is obtained through network training, and the configuration is then distributed. Specifically, the training strategy used in the model-based method for training the deep learning neural network model is as follows: S21: Define untied network training and tied network training; in untied network training… Define the set of learnable parameters as follows: , Represents the contraction function. Represents the matched filter matrix. This represents the total number of network layers, where t represents the corresponding network layer number. Each layer of the network needs to be trained, and the parameters obtained in each layer are different; in the training of the tilde network, the set of learnable parameters is defined as follows: After completing the learning Fixed for all layers, S22: Train each layer separately; S23: Use layer-by-layer training in both untied and tied network training; S24: Use untied network training for ST-LAMP network and tied network training for GGM-LAMP network. Input / output module: Configured to take in observed signals and output correct channel estimates.

Images