A data processing method for marine optical cable communication

Abstract

The application discloses a data processing method for marine optical cable communication, comprising the following steps: preprocessing the input marine optical cable communication signal; performing feature extraction to obtain a feature sequence, calculating an interpolation weight and a relative position offset, inputting the weighted feature and the offset information into a feature promotion module to generate an initial feature; inputting the initial feature into a feature fusion module based on a linear attention mechanism, and introducing a depth separable convolution operation; inputting the feature into a feedforward network to perform nonlinear transformation and channel mapping, and then generating a prediction residual through a projection layer; calculating the interpolation of the marine optical cable communication signal at a target position, and then adding the prediction residual to obtain the final marine optical cable communication signal. The application considers high precision, low complexity and strong robustness under complex marine environment, and provides a new technical path for high-quality transmission of the marine optical cable communication system.

Description

Technical Field

[0001] This invention relates to a data processing method for marine optical cable communication. Background Technology

[0002] In submarine optical cable communication, long-distance signal transmission is often affected by various factors, such as signal attenuation, noise pollution, and complex environmental changes, leading to signal distortion and reduced bandwidth utilization. Traditional data processing methods mainly rely on fixed filters and averaging algorithms. While these methods can be effective in simple communication environments, they exhibit significant limitations when dealing with the complex and ever-changing marine environment, especially in high-frequency signal recovery. Therefore, how to effectively improve the accuracy of signal recovery and transmission efficiency has become an urgent problem to be solved.

[0003] In recent years, the rapid development of deep learning technology has provided new ideas for solving this problem. In particular, Super-Resolution Networks (SRNs) have demonstrated excellent performance in image processing, improving image detail and clarity by recovering high-resolution (HR) images from low-resolution (LR) images. However, despite the significant achievements of deep neural networks (DNNs) in image processing, their direct application to one-dimensional signal processing still faces some technical challenges. First, the computational complexity is high, especially in long-distance signal processing, where performance bottlenecks are obvious. Second, deep learning models often struggle to fully capture the global correlation of signals, particularly in the recovery of high-frequency components.

[0004] To address these issues, novel methods such as Implicit Neural Functions (INFs) have emerged in recent years. These methods represent signals in a continuous space, thus avoiding the limitations of traditional discrete grid methods. By parameterizing continuous functions using Multilayer Perceptrons (MLPs), these methods can flexibly handle signal recovery at different resolutions. However, due to the local behavior of MLPs, it remains difficult to effectively recover high-frequency components during decoding. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a simple and efficient data processing method for submarine optical cable communication.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is: a data processing method for submarine optical cable communication, comprising the following steps:

[0007] S1: Preprocess the input submarine optical cable communication signal to obtain a signal sequence;

[0008] S2: Use the feature encoder to extract features from the preprocessed marine optical cable communication signal to obtain a feature sequence. Based on one-dimensional linear interpolation, determine two adjacent sampling points corresponding to the target position, calculate the interpolation weight and relative position offset, and input the weighted features and offset information into the feature enhancement module to generate initial features.

[0009] S3: Input the initial features into the feature fusion module based on a linear attention mechanism, and then query... ,key ,value The computation enables global information interaction; depthwise separable convolution operations are introduced into the output of the linear attention mechanism to enrich feature diversity;

[0010] S4: Input the fused and convolutional features into the feedforward network for nonlinear transformation and channel mapping, and then generate the prediction residual through the projection layer;

[0011] S5: Calculate the interpolation of the submarine optical cable communication signal at the target location using a one-dimensional linear interpolation function, and then add it to the prediction residual to obtain the final submarine optical cable communication signal.

[0012] In the data processing method for submarine optical cable communication described above, step S1 includes preprocessing such as missing value interpolation filling, convolutional smoothing and denoising, and normalization processing.

[0013] Assume the input submarine optical cable communication signal is The pre-processed submarine optical cable communication signal is ;but:

[0014] ;

[0015] in, Indicates the preprocessing function; This represents an interpolation function used to fill in missing values; The mean of the sample; Standard deviation; This indicates a smoothing filter used for smoothing and denoising.

[0016] In the data processing method for submarine optical cable communication described above, step S2, the process of feature extraction of the preprocessed submarine optical cable communication signal, is as follows:

[0017] ;

[0018] in, Indicates the location Features obtained through extraction; , Represents the sampling space of the signal; It is a learnable weight matrix used to map the preprocessed submarine optical cable communication signals to a unified feature space;

[0019] Obtain the feature sequence as follows:

[0020] ;

[0021] in, Indicates the length of the feature sequence.

[0022] In the data processing method for submarine optical cable communication described above, step S2, the process of generating initial features, is as follows:

[0023] For target signal coordinates First, obtain its two nearest neighbor sampling points in feature sequence A. and through the feature encoder Extract the feature representations of the two neighboring sampling points; calculate the weight of each neighboring sampling point based on linear interpolation. and weighted features relative position offset The concatenation yields intermediate features that integrate neighborhood features and relative positions. Then a local linear mapping function is used. Map the concatenated intermediate features to the initial features. ;

[0024] ;

[0025] ;

[0026] in, This represents concatenating vectors along the feature dimension; are nearby sampling points eigenvalues, ,in For adjacent indexes, ; This is the scaling factor.

[0027] In the data processing method for submarine optical cable communication described above, step S3 involves using a dual-path structure to enhance the input features within the linear attention mechanism; on one hand, the query generated through linear mapping... ,key ,value Perform attention operations to build global context information; on the other hand, values The branch enhances the local feature representation through depthwise separable convolution; finally, the results from both aspects are fused and added to the original input residual to obtain the updated feature representation; the specific process is as follows:

[0028] S31, Signal Generation: , , It is generated from the input signal by position-by-position projection, with each position... , , They all originate from the same input feature, and , , The dimension is consistent with the feature space of the input signal;

[0029] The sequence of input signals is , ,in It is the first Feature vectors at each position Indicates the position index. Indicates the total number of positions;

[0030] Input features are mapped using different projection matrices, and the projection matrix queried is used to... right Perform a linear transformation to obtain the query vector. , ; Projection matrix of the key right Perform a linear transformation to obtain the query vector. , ; through the projection matrix of values right Perform a linear transformation to obtain the query vector. , , , For the real number field, Indicates hidden dimensions;

[0031] S32, Calculation and Aggregation: and Aggregate into a key-value mapping matrix , superscript This is the transpose of the matrix. It is a positive feature mapping function, ensuring that the output is non-negative; Each column contains all This is the aggregation of information in the attention mechanism;

[0032] S33, Calculate the output of attention. ; after obtaining Then, use Calculation and all The similarity, and and Perform matching to obtain the attention output. :

[0033] ;

[0034] S34, to Normalization is performed to obtain global context information. :

[0035] ;

[0036] in, It is a constant. The position index in the sequence;

[0037] S35, to The branch performs local enhancement to obtain an output that includes both local temporal information and fused channel information. ;

[0038] ;

[0039] ;

[0040] ;

[0041] in, For depthwise separable convolution, For pointwise convolution, For independent one-dimensional convolutions, λ is the fusion coefficient. This is the feature representation obtained by fusing the enhanced values ​​from the convolution and the original values;

[0042] S36, merges two information streams and performs residual connection;

[0043] ;

[0044] ;

[0045] ;

[0046] in, For learnable linear mappings, To integrate weights, This represents the output after attention calculation and fusion with local enhancement information. To pass The output is projected back onto the original input features. Indicates the first Layer Input features at each location, Indicates the updated number Layer Input features at each location.

[0047] In the data processing method for submarine optical cable communication described above, in step S4, the updated feature representation is input into the feedforward network, first processed... Normalization is performed, followed by two layers of linear transformation and activation function to obtain... :

[0048] ;

[0049] in, and All are learnable weight matrices. , , is the dimension of the intermediate hidden layer of the feedforward network; All are biased; For activation functions; This indicates normalization.

[0050] In the data processing method for submarine optical cable communication described above, in step S4, it is assumed that the input features of the last layer have been obtained after reaching the last layer of iteration. ,Will Input projection function Hidden dimensions Mapped to target output dimension The predicted residuals are obtained. ;

[0051] ;

[0052] in, This is the weight matrix. ; For bias, .

[0053] In the data processing method for submarine optical cable communication described above, step S5 involves proportionally weighting and fusing the prediction residual with the interpolation of the submarine optical cable communication signal at the target location to obtain the final submarine optical cable communication signal. ;

[0054] ;

[0055] in, This indicates the communication signal of the submarine optical cable at the [number]th [number]. The result obtained by linear interpolation at each position. , These represent the sampled values ​​of two adjacent sampling points. This is the normalized distance between two adjacent sampling points.

[0056] The beneficial effects of this invention are as follows: This invention combines coordinate feature enhancement, linear attention mechanism, and depth convolution operation for the recovery and optimization of marine optical cable communication signals, which can achieve global correlation modeling while maintaining computational efficiency; through the kernel decomposition characteristics of linear attention, the computational complexity of traditional self-attention is effectively reduced, which is suitable for the strict requirements of real-time performance and energy consumption in submarine communication; the introduction of coordinate information and weighting strategy in the feature enhancement stage reduces artifacts caused by interpolation and improves the continuity and smoothness of signal recovery; in the fusion and projection stage, the multi-layer iterative update mechanism helps to recover high-frequency details in low-resolution signals, thereby improving the transmission quality and stability of signals; the overall solution balances high precision, low complexity, and strong robustness in complex marine environments, providing a new technical path for high-quality transmission of marine optical cable communication systems. Attached Figure Description

[0057] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] like Figure 1 As shown, a data processing method for submarine optical cable communication includes the following steps:

[0060] S1: Preprocess the input low-resolution submarine optical cable communication signal to obtain a signal sequence.

[0061] The preprocessing includes missing value interpolation and filling, convolutional smoothing and denoising, and normalization.

[0062] Assume the input submarine optical cable communication signal is The pre-processed submarine optical cable communication signal is ;but:

[0063] ;

[0064] in, Indicates the preprocessing function; This represents an interpolation function used to fill in missing values; The mean of the sample; Standard deviation; This indicates a smoothing filter used for smoothing and denoising.

[0065] S2: Use the feature encoder to extract features from the preprocessed marine optical cable communication signal to obtain a feature sequence. Based on one-dimensional linear interpolation, determine two adjacent sampling points corresponding to the target position, calculate the interpolation weight and relative position offset, and input the weighted features and offset information into the feature enhancement module to generate initial features.

[0066] The process of feature extraction for preprocessed submarine optical cable communication signals is as follows:

[0067] ;

[0068] in, Indicates the location Features obtained through extraction; , Represents the sampling space of the signal; It is a learnable weight matrix used to map the preprocessed submarine optical cable communication signals to a unified feature space;

[0069] Obtain the feature sequence as follows:

[0070] ;

[0071] in, Indicates the length of the feature sequence.

[0072] The process of generating initial features is as follows:

[0073] For target signal coordinates First, obtain its two nearest neighbor sampling points in feature sequence A. and through the feature encoder Feature representations of these two neighboring sampling points are extracted; to reconstruct more continuous features in the high-resolution space, weights for each neighboring sampling point are calculated based on linear interpolation. and weighted features relative position offset The concatenation yields intermediate features that integrate neighborhood features and relative positions. Then a local linear mapping function is used. Map the concatenated intermediate features to the initial features. ;

[0074] ;

[0075] ;

[0076] in, This represents concatenating vectors along the feature dimension; are nearby sampling points eigenvalues, ,in For adjacent indexes, ; As a scaling factor, it provides global location information to ensure feature consistency across different coordinate scales; The context-enhanced feature vector, which integrates multi-scale features and location offset information, is used as subsequent input.

[0077] A linear weighted strategy based on two-neighbor interpolation is employed to smoothly transition the feature vectors in the coordinate space. Specifically, the position... The normalized position corresponding to the low-resolution signal space is :

[0078] ;

[0079] in This is the scaling factor between high and low resolution. , Indicates the length of the high-resolution signal; for obtaining location The two nearest sampling points and ,definition:

[0080] ;

[0081] in, This indicates the floor function. This indicates that the smaller of the two values ​​should be used to avoid index out-of-bounds errors. Then the position is calculated. Normalized distance between two adjacent sampling points :

[0082] ;

[0083] Based on this, two weighting coefficients are constructed. and , and These are the interpolation weights for the left and right adjacent sampling points, respectively, calculated using the following formula:

[0084] , .

[0085] S3: Input the initial features into the feature fusion module based on a linear attention mechanism, and then query... ,key ,value The computation enables global information interaction; depthwise separable convolution operations are introduced into the output of the linear attention mechanism to enrich feature diversity.

[0086] The linear attention mechanism employs a dual-path structure to enhance input features; on the one hand, the query generated through linear mapping... ,key ,value Perform attention operations to build global context information; on the other hand, values The branch enhances the local feature representation through depthwise separable convolution; finally, the results from both aspects are fused and added to the original input residual to obtain the updated feature representation; the specific process is as follows:

[0087] S31, Signal Generation: , , It is generated from the input signal by position-by-position projection, with each position... , , They all originate from the same input feature, and , , The dimension is consistent with the feature space of the input signal;

[0088] The sequence of input signals is , ,in It is the first Feature vectors at each position Indicates the position index. Indicates the total number of positions;

[0089] Input features are mapped using different projection matrices, and the projection matrix queried is used to... right Perform a linear transformation to obtain the query vector. , ; Projection matrix of the key right Perform a linear transformation to obtain the query vector. , ; through the projection matrix of values right Perform a linear transformation to obtain the query vector. , , , For the real number field, Indicates hidden dimensions;

[0090] S32, Calculation and Aggregation: and Aggregate into a key-value mapping matrix , superscript This is the transpose of the matrix. It is a positive feature mapping function, ensuring that the output is non-negative; Each column contains all This is the aggregation of information in the attention mechanism;

[0091] S33, Calculate the output of attention. ; after obtaining Then, use Calculation and all The similarity, and and Perform matching to obtain the attention output. :

[0092] ;

[0093] Towards Query relevant information, that is, each According to Similarity, from Select the corresponding information;

[0094] S34, to Normalization is performed to obtain global context information. :

[0095] ;

[0096] in, It is a constant used to prevent errors caused by division by zero. This represents the position index in the sequence; the purpose of the normalization operation is to adjust the output at each position so that its numerical range remains within a stable range.

[0097] S35, to The branch performs local enhancement to obtain an output that includes both local temporal information and fused channel information. ;

[0098] ;

[0099] ;

[0100] ;

[0101] in, For depthwise separable convolution, For pointwise convolution, For independent one-dimensional convolutions, λ is the fusion coefficient, controlling... and The fusion ratio; This is the feature representation obtained by fusing the enhanced values ​​from the convolution and the original values;

[0102] S36, merges two information streams and performs residual connection;

[0103] ;

[0104] ;

[0105] ;

[0106] in, A learnable linear mapping maps the channel dimensions of the fusion path back to the original channel dimensions to facilitate residual summation and maintain consistency. If the attention output already has the same dimension as the input, no further mapping is needed; otherwise, a learnable linear mapping must be used. (Or 1×1 convolution) Align the dimensions before feeding them into the next layer; To integrate weights, This represents the output after attention calculation and fusion with local enhancement information. To pass The output is projected back onto the original input features. Indicates the first Layer Input features at each location, Indicates the updated number Layer Input features at each location.

[0107] S4: The fused and convolutional features are input into the feedforward network for nonlinear transformation and channel mapping, and then the prediction residual is generated through the projection layer.

[0108] The updated feature representation is input into the feedforward network, first... Normalization is performed, followed by two layers of linear transformation and activation function to obtain... :

[0109] ;

[0110] in, and All are learnable weight matrices. , , is the dimension of the intermediate hidden layer of the feedforward network; All are biased; For activation functions; This indicates normalization.

[0111] Assuming that the last layer of iterations has been reached and the input features of the last layer are obtained... ,Will Input projection function Hidden dimensions Mapped to target output dimension The predicted residuals are obtained. ;

[0112] ;

[0113] in, This is the weight matrix. ; For bias, .

[0114] S5: Calculate the interpolation of the submarine optical cable communication signal at the target location using a one-dimensional linear interpolation function, and then add it to the prediction residual to obtain the final high-resolution submarine optical cable communication signal.

[0115] The predicted residuals and the interpolated values ​​of the submarine optical cable communication signal at the target location are weighted and fused proportionally to obtain the final submarine optical cable communication signal. ;

[0116] ;

[0117] in, This indicates the communication signal of the submarine optical cable at the [number]th [number]. The result obtained by linear interpolation at each position. , These represent the sampled values ​​of two adjacent sampling points.

[0118] The interpolation results ensure the stability of the low-frequency components of the signal, while the prediction residuals compensate for high-frequency details, thereby significantly improving the recovery quality while reducing computational complexity.

Claims

1. A data processing method for submarine optical cable communication, characterized in that, Includes the following steps: S1: Preprocess the input submarine optical cable communication signal to obtain a signal sequence; S2: Use the feature encoder to extract features from the preprocessed marine optical cable communication signal to obtain a feature sequence. Based on one-dimensional linear interpolation, determine two adjacent sampling points corresponding to the target position, calculate the interpolation weight and relative position offset, and input the weighted features and offset information into the feature enhancement module to generate initial features. The process of feature extraction for preprocessed submarine optical cable communication signals is as follows: ； Among them, the pre-processed submarine optical cable communication signal is , Indicates the location Features obtained through extraction; , Represents the sampling space of the signal; It is a learnable weight matrix used to map the preprocessed submarine optical cable communication signals to a unified feature space; Obtain the feature sequence as follows: ； in, Indicates the length of the feature sequence; The process of generating initial features is as follows: For target signal coordinates First, obtain its two nearest neighbor sampling points in feature sequence A. and through the feature encoder Extract the feature representations of the two neighboring sampling points; calculate the weight of each neighboring sampling point based on linear interpolation. and weighted features relative position offset The concatenation yields intermediate features that integrate neighborhood features and relative positions. Then a local linear mapping function is used. Map the concatenated intermediate features to the initial features. ; ； ； in, This represents concatenating vectors along the feature dimension; are nearby sampling points eigenvalues, ,in For adjacent indexes, ; This is the scaling factor; S3: Input the initial features into the feature fusion module based on a linear attention mechanism, and then query... ,key ,value The computation enables global information interaction; depthwise separable convolution operations are introduced into the output of the linear attention mechanism to enrich feature diversity; The linear attention mechanism employs a dual-path structure to enhance input features; on the one hand, the query generated through linear mapping... ,key ,value Perform attention operations to build global context information; on the other hand, values The branch enhances the local feature representation through depthwise separable convolution; finally, the results from both aspects are fused and added to the original input residual to obtain the updated feature representation; the specific process is as follows: S31, Signal Generation: , , It is generated from the input signal by position-by-position projection, with each position... , , They all originate from the same input feature, and , , The dimension is consistent with the feature space of the input signal; The sequence of input signals is , ,in It is the first Feature vectors at each position Indicates the position index. Indicates the total number of positions; Input features are mapped using different projection matrices, and the projection matrix queried is used to... right Perform a linear transformation to obtain the query vector. , ; Projection matrix of the key right Perform a linear transformation to obtain the query vector. , ; Projection matrix of values right Perform a linear transformation to obtain the query vector. , , , For the real number field, Indicates hidden dimensions; S32, Calculation and Aggregation: and Aggregate into a key-value mapping matrix , superscript This is the transpose of the matrix. It is a positive feature mapping function, ensuring that the output is non-negative; Each column contains all This is the aggregation of information in the attention mechanism; S33, Calculate the output of attention. ; after obtaining After that, use Calculation and All The similarity, and and Perform matching to obtain the attention output. : ； S34, to Normalization is performed to obtain global context information. : ； in, It is a constant. The position index in the sequence; S35, to The branch performs local enhancement to obtain an output that includes both local temporal information and fused channel information. ; ； ； ； in, For depthwise separable convolution, For pointwise convolution, For independent one-dimensional convolutions, λ is the fusion coefficient. This is the feature representation obtained by fusing the enhanced values ​​from the convolution and the original values; S36, merges two information streams and performs residual connection; ； ； ； in, For learnable linear mappings, To integrate weights, This represents the output after attention calculation and fusion with local enhancement information. To pass The output is projected back onto the original input features. Indicates the first Layer Input features at each location, Indicates the updated number Layer Input features at each location; S4: Input the fused and convolutional features into the feedforward network for nonlinear transformation and channel mapping, and then generate the prediction residual through the projection layer; The updated feature representation is input into the feedforward network, first... Normalization is performed, followed by two layers of linear transformation and activation function to obtain the mapping result. : ； in, and All are learnable weight matrices. , , is the dimension of the intermediate hidden layer of the feedforward network; All are biased; For activation functions; Indicates normalization; Assuming that the last layer of iterations has been reached and the input features of the last layer are obtained... ,Will Input projection function Hidden dimensions Mapped to target output dimension The predicted residuals are obtained. ; ； in, This is the weight matrix. ; For bias, ; S5: Calculate the interpolation of the submarine optical cable communication signal at the target location using a one-dimensional linear interpolation function, and then add it to the prediction residual to obtain the final submarine optical cable communication signal.

2. The data processing method for submarine optical cable communication according to claim 1, characterized in that, In step S1, the preprocessing includes missing value interpolation and filling, convolutional smoothing and denoising, and normalization processing. Assume the input submarine optical cable communication signal is The pre-processed submarine optical cable communication signal is ;but: ； in, Indicates the preprocessing function; This represents an interpolation function used to fill in missing values; The mean of the sample; Standard deviation; This indicates a smoothing filter used for smoothing and denoising.

3. The data processing method for submarine optical cable communication according to claim 1, characterized in that, In step S5, the predicted residual and the interpolation of the submarine optical cable communication signal at the target location are weighted and fused proportionally to obtain the final submarine optical cable communication signal. ; ； in, This indicates the communication signal of the submarine optical cable at the 1st The result obtained by linear interpolation at each position. , These represent the sampled values ​​of two adjacent sampling points. This is the normalized distance between two adjacent sampling points.

Images