A turbulence scintillation coefficient data completion method and system for an underwater wireless optical turbulence channel

Abstract

The application discloses a kind of turbulent flicker coefficient data completion methods and systems for underwater wireless optical turbulence channel.First, for the characteristics of underwater environment, construct double condition information, use known data as condition information to assist denoising process, and construct environmental condition information using the temperature, salinity and propagation distance corresponding to the flicker coefficient data.Subsequently, a conditional diffusion model based on double condition information is constructed.The forward process is used to simulate the distribution change of the flicker coefficient data under gradually increasing noise.The reverse process uses double condition information to gradually denoise the noise data based on a noise learning function, achieving flicker coefficient completion.To extract the coupling features of three-dimensional flicker coefficient data, a multi-dimensional data coupling feature extraction neural network based on double condition information is constructed as the noise learning function.The application can complete the flicker coefficient data in various data missing scenarios, laying the foundation for building underwater wireless optical channel model and optimizing the performance of communication system.

Description

Technical Field

[0001] This invention relates to a turbulent scintillation coefficient data completion technology for underwater wireless optical turbulent channels, belonging to the field of underwater wireless optical communication technology. Background Technology

[0002] Underwater wireless optical communication, as a communication method with low latency, high bandwidth, and high security, plays an important role in many fields such as nearshore exploration, pollution monitoring, climate change detection, and oceanographic research. This communication system relies on the propagation of light signals in seawater to achieve data transmission, but its performance is susceptible to the influence of the complex marine environment. Factors such as suspended particles, temperature changes, and salinity variations in natural seawater can cause light scattering and refraction during propagation, resulting in fluctuations in the optical signal. The scintillation coefficient is an important parameter for measuring the intensity of optical signal fluctuations and is closely related to the bit error rate and received power stability of the communication link. Obtaining an accurate scintillation coefficient is crucial for optimizing the performance of underwater wireless optical communication systems.

[0003] In their paper "Scintillation index and BER performance for optical wave propagation in anisotropic underwater turbulence under the effect of eddy diffusivity ratio" published in Applied Optics, Xu G et al. studied the scintillation coefficients of plane waves and spherical waves and simulated the different effects of temperature and salinity disturbances on the scintillation coefficients. However, this method relies on idealized turbulence assumptions, which differs from the complex variations of the real underwater environment. In their paper "Bubbles-induced turbulence channel prediction mechanism based on machine vision in underwater wireless optical communication" published in OpticsExpress, Zhixin Dong et al. proposed a machine vision-based scintillation prediction method. This method analyzes bubble density through bubble images and combines the relationship between bubble density and scintillation coefficient to achieve bubble image-based scintillation prediction. However, this method is based on experimental conditions and does not consider the complexity of the real underwater environment, and obtaining scintillation coefficient data in real-world scenarios is difficult.

[0004] Existing research primarily focuses on theoretical modeling or mapping prediction of scintillation coefficients, neglecting the challenges of data acquisition during actual underwater observations. Due to limitations in sampling equipment, environmental disturbances, or observation difficulties under extreme ocean conditions, scintillation coefficient data often suffers from missing or uneven distribution. This leads to uneven data sample coverage, affecting the model's feature learning ability under different environmental conditions. Therefore, a method is needed to complete the missing parts of the scintillation coefficient data by combining environmental condition information with known data. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels, so as to solve the problem of difficulty in collecting and missing underwater turbulent scintillation coefficient data, and the completion of turbulent scintillation coefficient data is more comprehensive and targeted.

[0006] Technical Solution: To achieve the above-mentioned objectives, this invention provides a method for completing turbulent scintillation coefficient data for underwater wireless optical communication, comprising the following steps:

[0007] To address the characteristics of the underwater environment, dual-condition information is constructed. A conditional masking mechanism is used to distinguish between known data and data to be completed. Known data is used as conditional information to assist the denoising process. At the same time, environmental condition information is constructed using the temperature, salinity, and propagation distance corresponding to the scintillation coefficient data.

[0008] A conditional diffusion model based on dual-condition information is constructed. In the forward process, noise is gradually added to the original scintillation coefficient data to make the data distribution approximate a Gaussian distribution. In the reverse process, the dual-condition information is used as the conditional input, and the noise data is gradually denoised based on a noise learning function. The noise learning function is constructed by a multi-dimensional data coupling feature extraction network based on dual-condition information, which is used to calculate the noise component to be removed at each step in the reverse process. The network takes noise data, dual-condition information, and diffusion time step as input. First, it extracts the local spatial features of noise data, known data condition information, and diffusion time step through three-dimensional convolution. Then, it extracts global dependencies dimension by dimension through temperature attention layer, salinity attention layer, and propagation distance attention layer. Then, it fuses the data features extracted locally and globally with environmental condition information, and performs multi-layer feature aggregation through gated activation units and residual connection structures. Finally, it outputs the noise component gradually removed in the reverse process.

[0009] The noise learning function is used to perform stepwise denoising on the noise data to complete the flicker coefficient data.

[0010] Furthermore, the conditional masking mechanism utilizes a conditional mask matrix to distinguish known data within the flicker coefficient sample data. Data to be completed The conditional masking mechanism is defined as follows:

[0011]

[0012]

[0013] In the formula, These are scintillation coefficient sample data, where L, W, and H represent the temperature, salinity, and propagation distance dimensions of the data, respectively. It is a conditional mask matrix. For Hadama product; when When, it means The flicker coefficient at that location is the data value to be completed. When, it means The flicker coefficient at that location is known data.

[0014] Furthermore, environmental condition information is defined as:

[0015]

[0016] In the formula, This is environmental condition information; 4 represents the channel dimension size. It is the result of normalizing the temperature, salinity, and propagation distance data corresponding to the scintillation coefficient. This indicates that the data is spliced ​​along the channel dimension.

[0017] Furthermore, the forward process of the conditional diffusion model based on dual-conditional information follows a Markov process. Noise is gradually added to the scintillation coefficient data to approximate its distribution as a Gaussian distribution, which can be expressed as:

[0018]

[0019] In the formula, This represents the conditional probability distribution of the forward process. The mean is variance is The distribution is Gaussian, where t represents the current diffusion step. It is a hyperparameter that controls the level of added noise. This is the noisy flicker coefficient data for the current diffusion step. These are the noisy scintillation coefficients from the t-1 diffusion step, where I is the identity matrix, after T steps. It approximates a standard normal distribution of Gaussian noise.

[0020] Furthermore, the inverse process of the conditional diffusion model based on dual-conditional information takes dual-conditional information as input, including known data and environmental condition information, and gradually denoises the noisy data to obtain... , can be represented as:

[0021]

[0022] In the formula, This represents the conditional probability distribution of the reverse process. This is known data; En_info contains environmental condition information. For the mean term, This is the variance term;

[0023] Defined as: ;

[0024] In the formula, , , It is a noise learning function; Defined as: .

[0025] Furthermore, in the multidimensional data coupling feature extraction network based on dual-condition information, local spatial features of the data are extracted through three-dimensional convolution, including:

[0026] The known data and the data to be completed, distinguished by the conditional mask matrix, are concatenated and fed into the input projection layer, thereby using the known data to assist the denoising process.

[0027] The output of the input projection layer is concatenated with the embedding code of the diffusion step to obtain the input of the three-dimensional convolutional layer. Through three-dimensional convolution, the local spatial features of the three-dimensional scintillation coefficient data are extracted.

[0028] Furthermore, in the multidimensional data coupling feature extraction network based on dual-condition information, global dependencies are extracted dimension by dimension through a temperature attention layer, a salinity attention layer, and a propagation distance attention layer, including:

[0029] The output of the 3D convolutional layer is transformed by dimension and then input into the temperature attention layer to obtain the global dependencies in the temperature dimension.

[0030] The output of the temperature attention layer is transformed by dimension and then input into the salinity attention layer to obtain the global dependency in the salinity dimension.

[0031] The output of the salinity attention layer is transformed by dimension and then input into the propagation distance attention layer to obtain the global dependency of the propagation distance dimension; the output of the propagation distance attention layer is restored by dimension to obtain the final output of the three attention layers.

[0032] Furthermore, in the multidimensional data coupling feature extraction network based on dual-condition information, the data features extracted locally and globally are fused with environmental condition information, and multi-layer feature aggregation is performed through gated activation units and residual connection structures, including:

[0033] The data features obtained by sequentially passing through the temperature attention layer, salinity attention layer, and propagation distance attention layer are concatenated with environmental condition information and input into the gating activation unit to obtain residual output and jump output;

[0034] By utilizing multi-layer residual connections, the residual output of this layer is combined with the input of this layer as the input of the next layer;

[0035] The jump outputs of each residual layer are aggregated using the output projection layer to obtain the noise components that need to be removed in the current diffusion step.

[0036] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels.

[0037] The present invention also provides a computer program product, including a computer program that, when loaded into a processor, implements the steps of the method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels.

[0038] Beneficial Effects: This invention addresses the difficulty and incompleteness of underwater turbulent scintillation coefficient data collection, constructing a conditional diffusion model based on dual-condition information to complete the scintillation coefficient data. By introducing dual-condition information during the inverse denoising process, it can fully utilize known data and environmental condition information to achieve high-precision completion of the scintillation coefficient. Compared with existing technologies, the significant advantages of this invention are:

[0039] This invention constructs a dual-condition information-assisted inverse process for a diffusion model, utilizing a conditional masking mechanism to separate known data from data to be completed, with the known data serving as conditional information; simultaneously, temperature, salinity, and propagation distance corresponding to the scintillation coefficient are used as environmental conditional information. During the inverse process, a noise learning function is used to progressively denoise the noise data, achieving completion of the scintillation coefficient data. To improve the learning ability of the noise learning function, a multidimensional data coupling feature extraction neural network based on dual-condition information, combining 3D convolution, temperature attention, salinity attention, and propagation distance attention, is constructed as the noise learning function. Experiments show that the multidimensional data coupling feature extraction neural network based on dual-condition information can effectively capture local spatial features and global dependencies during the inverse process of the diffusion model, thereby significantly improving the noise prediction accuracy of the noise learning function and enhancing the network's ability to complete the scintillation coefficient. In block-shaped and L-shaped missing scenarios, the proposed method outperforms the Conv3_Atten and Conv3_MLP methods in RMSE, MAE, and CRPS metrics, achieving higher completion accuracy. Attached Figure Description

[0040] Figure 1 A step-by-step diagram of a method for completing scintillation coefficient data for underwater wireless optical communication;

[0041] Figure 2 The network structure diagram for the noise learning function;

[0042] Figure 3 This is a scene diagram showing a missing flicker coefficient data.

[0043] Figure 4 This is a diagram showing the completion result of the method when a block of the scene is missing.

[0044] Figure 5 The image shows the completion result of the method in the case of L-shaped missing elements in scenario two. Detailed Implementation

[0045] The invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown in the figure, the scintillation coefficient data completion method for underwater wireless optical communication disclosed in this embodiment of the invention mainly includes the following steps:

[0047] Step 1: Construct dual-condition information for the complex and ever-changing underwater environment; use a conditional masking mechanism to distinguish between known data and data to be completed, use known data as conditional information to assist in the denoising process, and use environmental data such as temperature, salinity and propagation distance corresponding to the scintillation coefficient data to construct environmental condition information.

[0048] In this embodiment, underwater wireless optical communication offers significant advantages such as low latency, high bandwidth, and high security, but its performance is susceptible to the complex marine environment. Marine environmental factors such as temperature variations, salinity changes, and bubble disturbances affect the propagation of optical signals in water. The scintillation coefficient is typically used to quantify the intensity of optical signal fluctuations affected by the marine environment. Considering the significant relationship between the marine environment and the scintillation coefficient, corresponding environmental data is used to assist in completing the scintillation coefficient. Simultaneously, the correlation between adjacent scintillation coefficient data is also considered, utilizing known data to aid in the denoising and completion process.

[0049] In view of the above characteristics, this embodiment constructs dual-condition information to assist the denoising process. The dual-condition information consists of known data obtained by the conditional masking mechanism and environmental condition information.

[0050] The conditional masking mechanism is implemented using a conditional mask matrix to distinguish known data in the flicker coefficient sample data. Data to be completed The conditional masking mechanism is defined as follows:

[0051] ; (1)

[0052] In the formula, These are scintillation coefficient sample data, where L, W, and H represent the temperature, salinity, and propagation distance dimensions of the data, respectively. It is a conditional mask matrix. For Hadama product; when When, it means The flicker coefficient at that location is the data value to be completed. When, it means The flicker coefficient at that location is known data;

[0053] Environmental condition information is defined as:

[0054] (2)

[0055] In the formula, This refers to environmental condition information; 4 represents the channel dimension size. This is the result of normalizing the temperature, salinity, and propagation distance data corresponding to the scintillation coefficient. Taking temperature data as an example, the temperature values ​​are broadcast and extended along the salinity and propagation distance dimensions to construct a three-dimensional temperature matrix with dimensions L×W×H. The maximum and minimum values ​​of the temperature data are then normalized and mapped to the [0,1] interval. This indicates that the data is concatenated along the channel dimension, with c_mask concatenated into En_info. This allows the model to better distinguish between the environmental information of the known data and the environmental information of the data to be completed, thereby making full use of the environmental information to assist in completing the missing data.

[0056] Step 2: Construct a conditional diffusion model based on dual-condition information. In the forward process, noise is gradually added to the original scintillation coefficient data to make the data distribution approximate a Gaussian distribution. In the reverse process, the dual-condition information is used as the conditional input, and the noise data is gradually denoised based on the noise learning function.

[0057] Specifically, the diffusion model, as a generative model, continuously optimizes the noise learning function through a forward process of gradually adding noise to the data and a reverse process, thereby recovering the data from the noise. The conditional diffusion model, based on the diffusion model, adds additional conditional information during the reverse denoising process to enhance the learning ability of the noise learning function, thus improving the performance of the diffusion model.

[0058] This embodiment is based on the conditional diffusion model and constructs a conditional diffusion model with dual conditional information.

[0059] The conditional diffusion model of dual-conditional information includes a forward process and a backward process. The forward process follows a Markov process, gradually adding noise to the scintillation coefficient data to make its distribution approximate a Gaussian distribution, which can be expressed as:

[0060] (3)

[0061] In the formula, t represents the current diffusion step. It is a hyperparameter that controls the level of added noise. This is the noisy flicker coefficient data for the current diffusion step. These are the noisy scintillation coefficients from the t-1 diffusion step, where I is the identity matrix, after T steps. Gaussian noise that approximates a standard normal distribution;

[0062] The reverse process takes dual-condition information as input, including known data and environmental condition information, and obtains the desired result by progressively denoising the noisy data. , can be represented as:

[0063] (4)

[0064] In the formula, This is known data; En_info contains environmental condition information. This is the noisy scintillation coefficient data for the current diffusion step, where I is the identity matrix. For the mean term, This is the variance term;

[0065] Defined as:

[0066] (5)

[0067] In the formula, , , It is a hyperparameter that controls the level of added noise. It is a noise learning function;

[0068] Defined as:

[0069] (6)

[0070] Noise learning function With noise data Known real data The environmental condition information En_info and the diffusion time step t are used as inputs. The noise to be added to the data to be completed is predicted to achieve the purpose of data completion.

[0071] In this embodiment, the noise learning function is constructed from a multidimensional data coupling feature extraction neural network based on biconditional information. It is used to extract local features and global dependencies of the data. The network structure diagram of the noise learning function is shown below. Figure 2 As shown, the network takes noisy data, biconditional information, and a diffusion time step as input. First, it extracts local spatial features from the noisy data, known data conditional information, and diffusion time step through 3D convolution. Then, it sequentially extracts global dependencies dimension-wise through temperature attention layers, salinity attention layers, and propagation distance attention layers. Next, it concatenates and fuses the locally and globally extracted data features with environmental conditional information, and performs multi-layer feature aggregation through gated activation units and residual connection structures. Finally, it outputs the noise components gradually removed during the inverse process. The following provides a detailed description of the multi-dimensional data coupling feature extraction neural network model structure based on biconditional information in this embodiment.

[0072] Embedding encoding of diffusion time step:

[0073] (7)

[0074] In the formula, It is a diffusion time-step embedding function, consisting of two linear transformation layers and two SiLU activation functions. It is the result of the diffusion time step encoding, and E_dim is the size of the diffusion step embedding dimension.

[0075] Diffusion time step embedding results and ( , The inputs are fed into convolutional layers for dimensionality transformation and then concatenated to obtain the input for a 3D convolution. The process is defined as follows:

[0076] (8)

[0077] (9)

[0078] In the formula, It is the input to the noise learning function, and C is the channel dimension of the data. It is the input projection layer, consisting of one layer It consists of a layer of ReLU activation function. It is a convolution with a kernel size of 1. In the... and During the integration process, the data is first... The dimensions become (C, L×W×H), and then... go through The results combined It is a linearly changing layer. This indicates that the data is added together. Finally, the dimensions of the combined data are transformed into (C, L, W, H) to obtain... .

[0079] In this embodiment, in order to extract the local spatial features and global dependencies of the data, the input data... First, it goes through a 3D convolutional layer, then sequentially through a temperature attention layer, a salinity attention layer, and a propagation distance dimension attention layer.

[0080] First, the data After passing through a 3D convolutional layer, local spatial features are extracted; this process is defined as follows:

[0081] (10)

[0082] In the formula, Y conv3 It is a three-dimensional convolutional layer from X input Extracted feature matrix, It is a 3D convolution with a kernel size of 3×3×3. In order to keep the data dimension size unchanged, this embodiment sets stride=1 and padding=1.

[0083] Secondly, the data undergoes an attention mechanism to obtain data dependencies along the temperature dimension. Before entering the attention layer, the data is transformed to match the task. Reshape( This indicates changing the data dimension to... Then, the dimensionally transformed data is input into the temperature attention layer; this process can be defined as:

[0084] (11)

[0085] In the formula, This is a temperature attention layer, implemented using a linear attention structure. The number of attention heads is set to 8, and the channel dimension is the same as the input data's channel dimension. It calculates the correlation between temperatures sequentially along the temperature dimension while maintaining the sequence length, thus obtaining global dependencies along the temperature dimension. temp_in It is Y conv3 The result after dimensional transformation, Y temp This is the feature data output by the temperature attention layer. Finally, Y temp After dimension transformation operation , This indicates that the data will be transformed back to its original dimensions, Y. temp_out It is Y temp The result after dimensional transformation, which is also the final output of the temperature attention layer, will be input into the salinity attention layer.

[0086] Then, the output of the temperature attention layer The data is fed into the salinity attention layer to extract features along the salinity dimension; this process is consistent with the temperature attention layer process described above. Before entering the attention layer, the data undergoes a dimensionality transformation. Then, the data Y salt_in The input to the salinity attention layer can be defined as follows:

[0087] (12)

[0088] In the formula, It is the salinity attention layer, Y salt_in It is Y temp_out The result after dimensional transformation, Y salt It is the feature data output by the salinity attention layer. Finally, Y salt After dimension transformation operation Y salt_out It is Y salt The result after dimensional transformation is also the final output of the salinity attention layer.

[0089] Finally, data Y salt_out The data will pass through the propagation distance attention layer to obtain its features along the propagation distance dimension. Before entering the propagation distance attention layer, the data undergoes dimensionality transformation. Then, the data Y dist_in The input propagation distance to the attention layer can be defined as follows:

[0090] (13)

[0091] In the formula, It is the attention layer of the propagation distance, Y dist_in It is Y salt_out The result after dimensional transformation, Y dist This is the feature data output by the attention layer, representing the propagation distance. Finally, Y dist After dimension transformation operation Y dist_out It is Y dist The result after dimensional transformation is also the final output of the three attention layers.

[0092] This embodiment will use data features Combined with the environmental condition information En_info, it is sent to the entry control mechanism for processing to obtain the residual output and jump output of this layer. This process can be defined as:

[0093] (14)

[0094] (15)

[0095] In the formula, It is the result of splicing feature extraction results and environmental condition information. It is a gating activation unit. It is residual output. It is a jump output.

[0096] This embodiment uses multi-layer residual connections, with each layer implementing the operations of formulas (9) to (15). The output of the i-th layer... Simultaneously, it will interact with the input of the i-th layer. Combination , as the input of the (i+1)th layer At the same time, the jump output for each level will be... Aggregation, to preserve information at different levels and improve the final completion result, is a process defined as follows:

[0097] (16)

[0098] In the formula, is the jump output of the i-th layer, and N is the number of residual connection layers. This is the output projection layer, which projects the aggregated results of the skipped outputs from each layer. It consists of two one-dimensional convolutional layers and one ReLU activation function layer, ultimately yielding the noise prediction value. .

[0099] Step 3: Optimize the noise learning function described above through the training phase. During training, each iteration samples the noise data and the diffusion step size, using known data... As conditional information, environmental conditional information En_info is constructed from real-world environmental data such as temperature, salinity, and propagation distance. The noise learning function is continuously optimized by minimizing the mean square error between the actual noise and the predicted noise value. Learning function for noise Conduct training.

[0100] First, noise is gradually added to the scintillation coefficient data through a forward diffusion process to obtain noise data at different diffusion step sizes. Then, in the reverse process, using biconditional information as input, a noise learning function predicts the noise component at each step. Finally, the noise is minimized by minimizing the actual noise added during the forward process. and noise prediction values The mean squared error between the two values ​​is used to optimize the noise learning function. This process is defined as follows:

[0101] (17)

[0102] Step 4: Use the trained noise learning function to complete the flicker coefficient data during the completion stage.

[0103] During the completion process, known real data will be used. Environmental condition information En_info and noise data As input to the trained noise learning function, the noise prediction value is obtained. Then, based on the reverse process, the data is gradually denoised to obtain the flicker coefficient data. The completion result .

[0104] The beneficial effects of the method of the present invention can be further illustrated by the following simulation.

[0105] I. Simulation Conditions

[0106] The missing data scenarios considered include block-shaped missing data and L-shaped missing data. For example... Figure 3 As shown, the blue part represents the known flicker coefficient, and the white part represents the missing part, that is, the part to be completed. Figure 3 In (a), the missing scenes are denoted as block missing. Figure 3 In (b), it is denoted as L-type deletion.

[0107] The specific parameter settings for the experiment are shown in Table 1.

[0108] Table 1 Experimental Parameter Table

[0109]

[0110] II. Simulation Content and Results

[0111] Based on given parameters, experiments were conducted using the PyTorch framework on a GPU RTX 4090D. This experiment used a non-overlapping sliding window method to divide the data into more than 20,000 6×6×6 block data samples, which were then divided into training, validation, and test sets in a 7:2:1 ratio.

[0112] This embodiment uses Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Continuous Ranked Probability Score (CRPS) to measure the experimental results. RMSE and MAE reflect the absolute error between the imputed value and the corresponding true value, focusing more on the average of the error results for each data point, while CRPS reflects the degree of consistency between the probability distribution of the imputed value and the corresponding true value.

[0113] The proposed method was evaluated under the two different data missing scenarios mentioned above. In addition, the following two comparison methods were designed to verify the effectiveness of this embodiment.

[0114] (1) Conv3_Atten: Unlike this embodiment, Conv3_Atten does not include a diffusion model, such as noise learning and time-step embedding. It is only used as a one-time forward prediction network and does not have the ability to generate backward diffusion. Its output is directly the completed value, and there is no longer a noise value prediction process. This method optimizes the loss function by minimizing the mean square error between the completed value and the true value output by the model. Conv3_Atten consists of only one layer of Conv3 and three layers of Atten. For fair comparison, the same network parameter settings are used. Because Conv3_Atten is a deterministic algorithm, it cannot calculate the CRPS metric.

[0115] (2) Conv3_MLP: Unlike this embodiment, the noise learning function structure of Conv3_MLP does not use the method of sequentially feeding the data into three attention layers and extracting global dependencies of the data dimension by dimension. Conv3_MLP feeds the data into three attention layers respectively, and concatenates the outputs of the three attention mechanisms in parallel, and feeds them into a standard two-layer MLP layer, which consists of two fully connected layers and one ReLU activation function.

[0116] like Figure 4 As shown, in the case of a block-shaped missing scene, the completion result of this embodiment has the lowest accuracy, with RMSE and MAE reduced by about 75% compared to Conv3_Atten; RMSE reduced by about 14% compared to Conv3_MLP, MAE reduced by about 19%, and CRPS reduced by about 21%.

[0117] like Figure 5 As shown, in scenario 2 with L-shaped missing, the completion result of this embodiment has the lowest accuracy, with RMSE and MAE reduced by about 67% compared to Conv3_Atten; RMSE reduced by about 14% compared to Conv3_MLP, MAE reduced by about 22%, and CRPS reduced by about 19%.

[0118] In summary, the conditional diffusion model using dual-conditional information exhibits good performance in completing scintillation coefficient data, and can complete scintillation coefficient data with high accuracy for two different missing scenarios.

[0119] Based on the same inventive concept, an embodiment of the present invention discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the turbulent scintillation coefficient data completion method for underwater wireless optical turbulence channels.

[0120] Based on the same inventive concept, the present invention discloses a computer program product, including a computer program, which, when executed by a processor, implements the steps of the turbulent scintillation coefficient data completion method for underwater wireless optical turbulence channels.

[0121] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

Claims

1. A method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels, characterized in that, The steps include the following: To address the characteristics of the underwater environment, dual-condition information is constructed. A conditional masking mechanism is used to distinguish between known data and data to be completed. The known data is used as conditional information to assist in the denoising process, while environmental condition information is constructed using the temperature, salinity, and propagation distance corresponding to the scintillation coefficient data. The conditional masking mechanism is implemented using a conditional masking matrix to distinguish known data in the scintillation coefficient sample data. Data to be completed The conditional masking mechanism is defined as follows: ； ； In the formula, These are scintillation coefficient sample data, where L, W, and H represent the temperature, salinity, and propagation distance dimensions of the data, respectively. It is a conditional mask matrix. For Hadama product; when When, it means The flicker coefficient at that location is the data value to be completed. When, it means The flicker coefficient at that location is known data; environmental condition information is defined as: ； It is the result of normalizing the temperature, salinity, and propagation distance data corresponding to the scintillation coefficient. This indicates that the data is concatenated along the channel dimension; A conditional diffusion model based on dual-conditional information is constructed. In the forward pass, noise is progressively added to the original scintillation coefficient data to approximate a Gaussian distribution. In the reverse pass, the dual-conditional information is used as the input, and the noise data is progressively denoised based on a noise learning function. The noise learning function is constructed by a multi-dimensional data coupling feature extraction neural network based on dual-conditional information, which is used to calculate the noise component to be removed at each step in the reverse pass. The network takes noise data, dual-conditional information, and diffusion time step as input. First, it extracts the local spatial features of the noise data, known data conditional information, and diffusion time step through three-dimensional convolution. Then, it sequentially extracts global dependencies dimension by dimension through temperature attention layer, salinity attention layer, and propagation distance attention layer. Finally, the data extracted locally and globally are combined. Features and environmental condition information are fused, and multi-layer feature aggregation is performed through gated activation units and residual connection structures. Finally, the noise components gradually removed during the inverse process are output. Specifically, global dependencies are extracted dimensionally through a temperature attention layer, a salinity attention layer, and a propagation distance attention layer. This includes: inputting the output of the 3D convolutional layer into the temperature attention layer after dimensional transformation to obtain global dependencies in the temperature dimension; inputting the output of the temperature attention layer into the salinity attention layer after dimensional transformation to obtain global dependencies in the salinity dimension; inputting the output of the salinity attention layer into the propagation distance attention layer after dimensional transformation to obtain global dependencies in the propagation distance dimension; and finally, the output of the propagation distance attention layer is restored to its dimension to obtain the final output of the three attention layers. By using a trained noise learning function to perform stepwise denoising on the noise data, the flicker coefficient data can be completed.

2. The method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels according to claim 1, characterized in that, The forward process of the conditional diffusion model based on biconditional information follows a Markov process. Noise is gradually added to the scintillation coefficient data to make its distribution approximate a Gaussian distribution, as shown below: ； In the formula, This represents the conditional probability distribution of the forward process. The mean is variance is The distribution is Gaussian, where t represents the current diffusion step. It is a hyperparameter that controls the level of added noise. This is the noisy flicker coefficient data for the current diffusion step. These are noisy scintillation coefficient data for the t-1 diffusion step. It is an identity matrix, after T steps, It approximates a standard normal distribution of Gaussian noise.

3. The method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels according to claim 2, characterized in that, The inverse process of the conditional diffusion model based on dual-conditional information takes dual-conditional information as input, including known data and environmental condition information, and gradually denoises the noisy data to obtain... , represented as: ； In the formula, This represents the conditional probability distribution of the reverse process. This is known data; En_info contains environmental condition information. For the mean term, This is the variance term; Defined as: ; In the formula, , , It is a noise learning function; Defined as: .

4. The method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels according to claim 1, characterized in that, In the multidimensional data coupling feature extraction network based on dual-condition information, local spatial features of the data are extracted through three-dimensional convolution, including: The known data and the data to be completed, as distinguished by the conditional mask matrix, are concatenated and then fed into the input projection layer. The output of the input projection layer is concatenated with the embedding code of the diffusion step to obtain the input of the three-dimensional convolutional layer. Through three-dimensional convolution, the local spatial features of the three-dimensional scintillation coefficient data are extracted.

5. A method for completing turbulent scintillation coefficient data for underwater wireless optical turbulence channels according to claim 1, characterized in that, In the multidimensional data coupling feature extraction network based on dual-condition information, the data features extracted locally and globally are fused with environmental condition information, and multi-layer feature aggregation is performed through gated activation units and residual connection structures, including: The data features obtained by sequentially passing through the temperature attention layer, salinity attention layer, and propagation distance attention layer are concatenated with environmental condition information and input into the gating activation unit to obtain residual output and jump output; By utilizing multi-layer residual connections, the residual output of this layer is combined with the input of this layer as the input of the next layer; The jump outputs of each residual layer are aggregated using the output projection layer to obtain the noise components that need to be removed in the current diffusion step.

6. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the steps of the turbulent scintillation coefficient data completion method for underwater wireless optical turbulent channels according to any one of claims 1-5.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is loaded into the processor, it implements the steps of the turbulent scintillation coefficient data completion method for underwater wireless optical turbulent channels according to any one of claims 1-5.

Images