A CNN-LSTM-based phase calibration method for mode division multiplexing optical fiber communication
By using a hybrid neural network architecture based on CNN-LSTM, the problems of phase noise compensation and signal processing under strong coupling environment in optical fiber communication are solved. It achieves efficient phase noise suppression and signal quality improvement, with strong adaptability, high computational efficiency, and meets the requirements of real-time processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING JIAOTONG UNIV
- Filing Date
- 2025-06-04
- Publication Date
- 2026-06-26
AI Technical Summary
In existing optical fiber communication technologies, traditional algorithms suffer from limitations in phase noise compensation capabilities, insufficient utilization of spatiotemporal correlation, poor convergence performance under strong coupling conditions, lack of adaptive phase calibration mechanisms, and difficulty in balancing computational complexity and performance during long-distance transmission.
A hybrid neural network architecture based on CNN-LSTM is adopted. By constructing a dual-input network and combining the spatial feature extraction capability of convolutional neural networks with the temporal modeling capability of long short-term memory networks, phase noise compensation for modular division multiplexing optical fiber communication is achieved. Iterative training and phase feedback mechanisms are used for adaptive estimation and compensation, supporting multi-mode collaborative processing and real-time processing.
It significantly improves adaptability and convergence performance in strongly coupled environments, effectively suppresses phase noise, reduces bit error rate, improves signal quality, enhances system robustness, and optimizes computational efficiency to meet real-time processing requirements.
Smart Images
Figure CN120498555B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of optical fiber communication technology and relates to a phase calibration method for mode division multiplexing optical fiber communication based on CNN-LSTM. Background Technology
[0002] With the rapid development of applications such as 5G, IoT, and cloud computing, the demand for data transmission is exploding. Mode division multiplexing (MDM) technology has become one of the key technologies for improving the capacity of optical fiber communication. However, in long-distance transmission, factors such as mode coupling, phase noise, and dispersion seriously affect signal quality, and traditional linear equalization algorithms have limitations in handling strong coupling and nonlinear effects.
[0003] Commonly used equalization algorithms in the prior art include:
[0004] 1. Least Mean Square Algorithm: An adaptive filtering algorithm based on gradient descent. It has a slow convergence speed, limited performance under strong coupling conditions, and difficulty in handling rapidly changing channel conditions.
[0005] 2. Normative Algorithm: A blind equalization algorithm that does not require training sequences, but is sensitive to phase noise and is prone to mode aliasing in multi-mode systems;
[0006] 3. Neural network equalization: It can handle nonlinear effects, but it lacks effective extraction of temporal features and has high training complexity.
[0007] The shortcomings of existing technology are:
[0008] Limited phase noise compensation capability: Traditional algorithms mainly focus on amplitude equalization and lack effective compensation mechanisms for laser phase noise and nonlinear phase noise in optical fibers; Insufficient utilization of spatiotemporal correlation: Failure to fully utilize the correlation of mode-division multiplexed signals in time and space dimensions leads to limited equalization performance; Poor convergence performance under strong coupling conditions: Under strong mode coupling conditions in long-distance transmission, traditional algorithms are prone to getting trapped in local optima, resulting in slow convergence speed and poor stability; Lack of adaptive phase calibration mechanism: Existing methods are difficult to track and compensate for rapidly changing phase noise in real time in high-speed transmission systems; Difficulty in balancing computational complexity and performance: High-performance algorithms are often accompanied by high computational complexity, making it difficult to meet real-time processing requirements. Summary of the Invention
[0009] This invention addresses the problems of existing technologies by providing a phase calibration method for mode division multiplexing optical fiber communication based on CNN-LSTM.
[0010] A phase calibration method for mode-division multiplexing (MDD) fiber optic communication based on CNN-LSTM is proposed. This method constructs a dual-input CNN-LSTM hybrid neural network architecture to acquire the MIMO input signal and the original target signal of the MMDD fiber optic communication. Data preprocessing is performed, and the spatial feature extraction capability of convolutional neural networks and the temporal modeling capability of long short-term memory networks are combined to achieve effective compensation for phase noise in long-distance, strongly coupled MMDD fiber optic communication. By separating the complex MIMO signal into real and imaginary parts and generating sliding window training samples, a dual-branch network is constructed, including an image input layer, convolutional layers, LSTM layers, deep connection layers, and a custom phase calibration layer. Adaptive estimation and compensation of phase noise are achieved through iterative training and a phase feedback mechanism, and the test results are optimized using decision feedback.
[0011] The beneficial effects of this invention include:
[0012] 1. Strong coupling adaptability: The CNN-LSTM architecture can effectively extract the spatiotemporal features of the signal and adapt to long-distance strong coupling environments. Through the 3×3 kernel function of the convolutional layer and the 256 hidden units of the LSTM, it can handle complex inter-mode coupling relationships. Compared with traditional linear equalizers, its adaptability to strong coupling environments is significantly improved.
[0013] 2. Phase noise compensation capability: Real-time estimation and compensation of phase noise are achieved through a phase calibration layer and an iterative feedback mechanism; the phase estimation algorithm based on block mean is adopted, which can effectively suppress phase jitter caused by laser phase noise and fiber nonlinear effects, thus effectively improving the phase noise suppression capability.
[0014] 3. Adaptive convergence performance: Decision feedback and multi-round iterative training (configurable number of iterations) are adopted to improve the system's adaptability and convergence; the Adam optimizer and gradient pruning techniques ensure training stability and convergence speed is faster than the traditional recursive least squares algorithm.
[0015] 4. Multi-mode collaborative processing: Fully utilize the correlation information between multiple modes through deep connection layers; Supports 6-mode MIMO processing, and can simultaneously process multiple spatially multiplexed optical field modes.
[0016] 5. Real-time processing capability: The "last" output mode of LSTM ensures real-time processing capability; the processing latency of a single sample is less than 10μs, which meets the real-time requirements of high-speed optical communication systems and supports data transmission rates of over 100Gbps.
[0017] 6. Signal quality improvement: Significantly improve signal quality through sliding window mechanism and timing modeling; the bit error rate (BER) is reduced by 1-2 orders of magnitude compared with traditional DSP algorithms, and the convergence of the signal constellation diagram is significantly improved.
[0018] 7. System robustness: The system adopts a dual-input architecture and phase feedback mechanism to enhance its robustness to channel changes; it can still maintain stable demodulation performance under deteriorating channel conditions, and its adaptability is stronger than that of traditional fixed-parameter algorithms.
[0019] 8. Optimized computational efficiency: Improved computational efficiency through network structure optimization and batch processing techniques; compared with traditional deep learning methods, the number of parameters is reduced by 40% and training time is shortened by 50%, making it easier to implement in hardware and deploy in practice.
[0020] 9. Engineering Applicability: The algorithm has good engineering applicability and scalability; it supports the configuration of different modes and channel conditions, and can flexibly adapt to the needs of different optical fiber communication systems, providing an effective technical solution for the industrial application of mode division multiplexing technology.
[0021] 10. Innovative technology integration: For the first time, CNN spatial feature extraction, LSTM temporal modeling and custom phase calibration layer are organically combined to form a complete end-to-end optimization scheme; compared with single technology methods, the overall performance is significantly improved, opening up a new technical path for intelligent signal processing in the field of optical fiber communication. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. As shown in the figures:
[0023] Figure 1 This is a schematic diagram of the overall structure of the CNN-LSTM network architecture of the present invention, showing the complete process of dual input branches, feature extraction, fusion processing and output.
[0024] Figure 2 This diagram illustrates the data preprocessing for a sliding window, showcasing the method for generating time-series windows and the boundary handling strategy.
[0025] Figure 3 The diagram shows a detailed structure of the phase calibration layer, illustrating the specific implementation of phase noise estimation and compensation.
[0026] Figure 4 The flowchart illustrates the iterative training process, including adaptive estimation of phase noise and updating of network parameters.
[0027] Figure 5 This is a convergence graph of the method of the present invention, which reflects the effects of the algorithm's convergence speed, convergence error, etc.
[0028] Figure 6 A comparison of constellation diagrams for a 6-mode MDM system demonstrates the balancing effect of the method of this invention in different modes. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Example 1: As Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown, a phase calibration method for mode division multiplexing optical fiber communication based on CNN-LSTM is presented. This method, based on convolutional neural networks and long short-term memory networks, is suitable for signal equalization and phase noise compensation in long-distance strongly coupled mode division multiplexing optical fiber communication systems. It is a novel method that can effectively handle phase noise in long-distance strongly coupled mode division multiplexing optical fiber communication, ensuring high-performance signal equalization while meeting the requirements of real-time processing.
[0031] A phase calibration method for mode-division multiplexing optical fiber communication based on CNN-LSTM is proposed. By combining the feature extraction capability of convolutional neural networks and the temporal modeling capability of long short-term memory networks, this method can effectively compensate for phase noise in long-distance strongly coupled mode-division multiplexing optical fiber communication, thereby improving the system's equalization performance and transmission quality.
[0032] A phase calibration method for mode-division multiplexing optical fiber communication based on CNN-LSTM includes the following steps:
[0033] Step 1: Data Preprocessing
[0034] The MIMO input signal and the original signal of the modulo-division multiplexing optical fiber communication are acquired, and the complex signal is separated into the real part and the imaginary part.
[0035] Set the sliding window parameters, including the tap length and the sample block size for phase calculation;
[0036] Training samples are generated using a sliding window technique, with a window size of 2 × window length + 1.
[0037] Step 2: Network Architecture Construction
[0038] Constructing a two-input CNN-LSTM hybrid neural network includes:
[0039] First input branch: The image input layer receives data with dimensions [number of patterns × 2, window size, 1], performs feature extraction through a convolutional layer, and then outputs the data through a flattening layer;
[0040] LSTM feature extraction layer: An LSTM layer containing 256 hidden units, using the "last" output mode, with an output dimension of 12;
[0041] The second input branch receives a phase noise compensation signal with dimensions [12,1,1].
[0042] Deep connection layer: concatenates the outputs of two branches;
[0043] Phase calibration layer: A custom phase calibration layer enables real-time compensation of phase noise;
[0044] Fully connected output layer: outputs 12-dimensional signals, corresponding to the real and imaginary parts of 6 modes;
[0045] Step 3: Iterative Training and Phase Estimation
[0046] The initial phase noise compensation vector PN is [1,1,1,1,1,1,1,0,0,0,0,0,0];
[0047] Perform the first network training to obtain initial prediction results;
[0048] Calculate the phase noise based on the predicted complex signal and the training target signal;
[0049] Update the phase noise compensation vector;
[0050] The training process is repeated to achieve adaptive estimation and compensation of phase noise;
[0051] Step 4: Testing and Decision Feedback
[0052] Predict the test data to obtain the initial equilibrium result;
[0053] Apply phase calibration and calculate the calibrated complex signal;
[0054] The training objective is updated by obtaining decision signals through hard decision-making.
[0055] Perform multiple rounds of iterative optimization to improve the equalization accuracy;
[0056] Step 5: Output Processing
[0057] The 12-dimensional real vector output by the network is converted into a complex signal output with 6 modes.
[0058] like Figure 1The diagram illustrates the neural network structure with two input branches. The feature extraction path consists of input layer 1 (12×(tap×2+1)×1 dimension), a 3×3 convolutional layer (12 filters), a flattening layer, an LSTM layer (256 hidden units, last output mode), and fully connected layer 2 (12 neurons). The phase compensation path consists of input layer 2 (12×1×1 dimension) and a flattening layer. The two branches are merged through a deep connected layer (Concat), then passed through a custom phase calibration layer (phaseCalibrationLayer3), and finally through fully connected layer 3 (12 neurons) and the output layer to produce a 12-dimensional output.
[0059] like Figure 2 As shown, the top section illustrates the sliding window mechanism for the input data sequence, constructing a window centered on the current sample and encompassing samples from all preceding and following time points. The bottom section demonstrates three boundary case handling methods: zero-padding at the beginning of the sequence, a full window in the normal case, and zero-padding at the end of the sequence, ensuring that all samples can have a uniformly sized input window.
[0060] like Figure 3 The diagram illustrates the internal working mechanism of the phase calibration layer. LSTM features (F_ lstm ) and phase input (P_ n As input, the phase angle and phase calibration factor are calculated by the phase estimation module, and finally the phase calibration is performed in the phase compensation module, outputting the calibrated signal F_ out .
[0061] like Figure 4 The diagram illustrates the complete training and testing process. It begins with initializing the PN (phase noise vector), then proceeds in a loop: training the network, predicting the output, calculating the phase noise, updating the PN vector, determining if the maximum number of iterations has been reached, and continuing the loop if the maximum number of iterations has been reached; otherwise, the final result is output.
[0062] like Figure 5 As shown, the mean squared error (MSE) changes with the number of iterations. The horizontal axis represents the number of iterations, and the vertical axis represents the MSE. The curve shows that the error decreases rapidly from an initial value of approximately 0.67, decreases sharply in the first 5 iterations, stabilizes around the 10th iteration, and finally converges to a level close to 0, demonstrating the algorithm's fast convergence.
[0063] like Figure 6As shown, the constellation diagrams for six patterns are compared on the training and test sets. The left side shows the training set results, and the right side shows the test set results. Each subplot displays the distribution of predicted and true values on the complex plane. The training set results show that the predicted points are highly concentrated and coincide well with the true values. Although the test set results are slightly dispersed, they still maintain good clustering effects, verifying the effectiveness and generalization ability of the algorithm.
[0064] Example 2: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown, a phase calibration method for mode-division multiplexing optical fiber communication based on CNN-LSTM is described in detail using a 6-mode mode-division multiplexing optical fiber communication system as an example.
[0065] Network parameter settings:
[0066] • Input dimension: 12 × (2 × window size + 1).
[0067] • Number of hidden units in LSTM: 256.
[0068] • Kernel size: 3×3, Number of output channels: 12.
[0069] • Learning rate: 1e-4.
[0070] Batch size: 128.
[0071] • Phase estimation block size: 20.
[0072] Data preprocessing:
[0073] Convert complex signals to real and imaginary part formats.
[0074] • Generate a sliding window.
[0075] Network training: The first stage uses the initial phase estimate for training, and subsequent stages use phase feedback for adaptive optimization. The phase noise estimate is updated after each epoch to achieve gradual convergence.
[0076] Phase calibration is achieved as follows:
[0077] 1. Receive feature and phase inputs from the LSTM layer.
[0078] 2. Calculate the phase compensation factor.
[0079] 3. Perform phase calibration on the signal.
[0080] 4. Output the characteristics of the calibrated signal.
[0081] Example 3: As Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown, a phase calibration method for mode division multiplexing (MDM) fiber communication based on CNN-LSTM can be adapted to MDM systems with different numbers of modes.
[0082] 4-mode system:
[0083] The input dimension is adjusted to 8 × (2 × window size + 1).
[0084] The output dimension has been adjusted to 8.
[0085] Other parameters remain unchanged.
[0086] 12-mode system:
[0087] The input dimension is adjusted to 24 × (2 × window size + 1).
[0088] The output dimension has been adjusted to 24.
[0089] The number of hidden units in an LSTM can be increased to 512 to improve processing power.
[0090] Example 4: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown, a phase calibration method for mode-division multiplexing optical fiber communication based on CNN-LSTM includes the following steps:
[0091] Step 1: Data preprocessing,
[0092] Acquire the MIMO input signal and the original target signal of the mode-division multiplexing optical fiber communication.
[0093] The complex signal is separated into real and imaginary parts, forming a data format in which real and imaginary parts are interleaved.
[0094] Configure the sliding window parameters, including the tap length and the sample block size for phase calculation.
[0095] Training samples are generated using a sliding window technique, with a window size of 2 × window length + 1, and the boundaries are padded with zeros.
[0096] Step 2: Construct a dual-input CNN-LSTM hybrid neural network architecture.
[0097] First input branch: The image input layer is set to receive sliding window data with dimensions of [number of patterns × 2, window size, 1]. Spatial features are extracted through the convolutional layer, and then a one-dimensional feature vector is output through the flattening layer.
[0098] LSTM Feature Extraction Layer: An LSTM layer containing 256 hidden units, employing the "last" output mode, to model temporal features.
[0099] The second input branch receives a phase noise compensation signal with dimensions [12,1,1].
[0100] Deep connection layer: Deeply stitches together the LSTM output features and the phase compensation signal.
[0101] Phase calibration layer: A custom phase calibration layer enables real-time compensation of phase noise.
[0102] Fully connected output layer: Outputs a 12-dimensional signal, corresponding to the real and imaginary parts of 6 modes.
[0103] Step 3: Iterative training and adaptive phase estimation.
[0104] The initial phase noise compensation vector PN is [1,1,1,1,1,1,1,0,0,0,0,0,0] .
[0105] The first round of network training was conducted to obtain initial prediction results.
[0106] Calculate the phase noise based on the predicted complex signal and the target signal.
[0107] Update the phase noise compensation vector.
[0108] The training process is repeated to achieve adaptive estimation and compensation of phase noise.
[0109] Step 4: Testing and Decision Feedback Optimization
[0110] Initial predictions are made based on the test data to obtain the equilibrium results.
[0111] Apply phase calibration and calculate the calibrated complex signal.
[0112] The training objective is updated by obtaining decision signals through hard decision-making.
[0113] Multiple rounds of iterative optimization are performed to improve the equalization accuracy.
[0114] Step 5: Output complex signal.
[0115] The 12-dimensional real vector output by the network is converted into a complex signal output with 6 modes.
[0116] During the sliding window generation process, zero padding is performed on boundary samples: when the window length is insufficient, zeros are padded at the beginning or end of the sequence to ensure that the window size of each sample is consistent.
[0117] In the CNN-LSTM network architecture, the convolutional layers use 3×3 convolutional kernels, have 12 output channels, and use the same padding method; the LSTM layers have 256 hidden units, and the output mode is set to "last" to ensure real-time processing capability.
[0118] The phase noise estimation adopts a block processing method, which estimates the phase noise by calculating the phase difference between the predicted signal and the target signal within the sample block. The sample block size is a configurable parameter.
[0119] The iterative training process uses the Adam optimizer with a learning rate of 1e-4, a batch size of 128, and a gradient pruning threshold of 1. Through multiple iterations, the network parameters and phase estimation are jointly optimized.
[0120] In the decision feedback process, 4-QAM hard decision is used to obtain the decision signal, and the decision result is used as a new training target to achieve adaptive optimization in the testing phase.
[0121] The method is applicable to MDM systems with different numbers of modes.
[0122] For a 4-mode system, the input dimension is adjusted to 8 × (2 × window size + 1), and the output dimension is 8.
[0123] For the 12-mode system, the input dimension is adjusted to 24×(2×window size+1), the output dimension is 24, and the number of LSTM hidden units increases to 512.
[0124] The phase calibration layer receives the feature vector and phase compensation signal from the LSTM, performs real-time phase calibration on the signal using the phase compensation factor, and outputs the calibrated signal features.
[0125] The network training adopts a data storage combination method, which combines sliding window data, phase compensation signal and target signal to support batch parallel processing.
[0126] Example 5: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6As shown, a phase calibration method for mode-division multiplexing fiber optic communication based on CNN-LSTM is proposed. This method constructs a dual-input CNN-LSTM hybrid neural network architecture to acquire the MIMO input signal and the original target signal of the mode-division multiplexing fiber optic communication. Data preprocessing is performed, and the spatial feature extraction capability of convolutional neural networks and the temporal modeling capability of long short-term memory networks are combined to achieve effective compensation for phase noise in long-distance, strongly coupled mode-division multiplexing fiber optic communication. By separating the complex MIMO signal into real and imaginary parts and generating sliding window training samples, a dual-branch network is constructed, including an image input layer, convolutional layer, LSTM layer, deep connection layer, and a custom phase calibration layer. Adaptive estimation and compensation of phase noise are achieved through iterative training and a phase feedback mechanism, and the test results are optimized using decision feedback.
[0127] The data preprocessing process includes: separating the complex signal into real and imaginary parts to form a data format with alternating real and imaginary parts; setting sliding window parameters including tap length and phase calculation sample block size; generating training samples using sliding window technology, wherein the window size is 2 × window length + 1; and padding the boundaries with zeros to ensure that the window size of each sample is consistent.
[0128] The dual-input CNN-LSTM hybrid neural network architecture includes: a first input branch sets the image input layer to receive sliding window data with dimensions [number of modes × 2, window size, 1], which is then processed by a convolutional layer with 3×3 kernels, 12 output channels, and the same padding method for spatial feature extraction, and then outputs a one-dimensional feature vector through a flattening layer; an LSTM feature extraction layer with 256 hidden units and the "last" output mode models the temporal features; a second input branch receives a phase noise compensation signal with dimensions [12, 1, 1]; a deep connection layer performs deep concatenation of the LSTM output features and the phase compensation signal; a custom phase calibration layer realizes real-time compensation of phase noise; and a fully connected output layer outputs a 12-dimensional signal, corresponding to the real and imaginary parts of 6 modes.
[0129] The iterative training and phase feedback mechanism includes: initializing the phase noise compensation vector PN to [1,1,1,1,1,1,0,0,0,0,0,0], performing the first round of network training to obtain the initial prediction result, calculating the phase noise based on the predicted complex signal and the target signal, updating the phase noise compensation vector, and repeating the training process to achieve adaptive estimation and compensation of phase noise. The phase noise estimation adopts a block processing method, estimating the phase noise by calculating the phase difference between the predicted signal and the target signal within the sample block.
[0130] The decision feedback optimization process includes: making initial predictions on the test data to obtain the equalization result, applying phase calibration to calculate the calibrated complex signal, obtaining the decision signal and updating the training target through 4-QAM hard decision, performing multiple rounds of iterative optimization to improve the equalization accuracy, realizing adaptive optimization in the testing phase, and finally converting the 12-dimensional real vector output by the network into complex signals with 6 modes for output.
[0131] The iterative training process uses the Adam optimizer with a learning rate of 1e-4, a batch size of 128, and a gradient clipping threshold of 1. Through multiple iterations, the network parameters and phase estimation are jointly optimized. The network training adopts a data storage combination method, which combines sliding window data, phase compensation signals, and target signals, and supports batch parallel processing.
[0132] A phase calibration method for mode division multiplexing (MDM) fiber communication based on CNN-LSTM is proposed, applicable to MDM systems with different numbers of modes: for a 4-mode system, the input dimension is adjusted to 8×(2×window size + 1), and the output dimension is 8; for a 12-mode system, the input dimension is adjusted to 24×(2×window size + 1), and the output dimension is 24. The number of LSTM hidden units is increased to 512. The phase calibration layer receives the feature vector and phase compensation signal from the LSTM, performs real-time phase calibration on the signal through the phase compensation factor, and outputs the calibrated signal features.
[0133] A CNN-LSTM-based mode division multiplexing fiber optic communication phase calibration system includes: a data preprocessing module for acquiring MIMO input signals and target signals, separating complex signals into real and imaginary parts, and generating sliding window training samples; a network construction module for constructing a dual-input CNN-LSTM hybrid neural network architecture; a training optimization module for performing iterative training and adaptive phase estimation; a test feedback module for performing testing and decision feedback optimization; and an output processing module for converting the network output into a complex signal output.
[0134] A CNN-LSTM-based mode division multiplexing fiber optic communication phase calibration system includes:
[0135] Data preprocessing module: This module provides data preprocessing functionality for the above-mentioned execution steps.
[0136] Network building module: Used to build a dual-input CNN-LSTM network architecture to perform the above steps.
[0137] Training and optimization module: used to perform iterative training and phase estimation of the above steps.
[0138] Test Feedback Module: Used for testing and decision feedback in the above steps.
[0139] Output processing module: Used to perform complex signal output processing in the above steps.
[0140] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A phase calibration method for mode-division multiplexing optical fiber communication based on CNN-LSTM, characterized in that, By constructing a dual-input CNN-LSTM hybrid neural network architecture, the MIMO input signal and the original target signal of mode-division multiplexing optical fiber communication are acquired. Data preprocessing is performed, and the spatial feature extraction capability of convolutional neural networks and the temporal modeling capability of long short-term memory networks are combined to achieve effective compensation for phase noise in long-distance, strongly coupled mode-division multiplexing optical fiber communication. By separating the complex MIMO signal into real and imaginary parts and generating sliding window training samples, a dual-branch network containing an image input layer, convolutional layers, LSTM layers, deep connection layers, and a custom phase calibration layer is constructed. Adaptive estimation and compensation of phase noise are achieved through iterative training and a phase feedback mechanism. Decision feedback is used to optimize the test results. The dual-input CNN-LSTM hybrid neural network architecture includes: The first input branch sets up an image input layer that receives sliding window data with dimensions [number of modes × 2, window size, 1]. Spatial features are extracted through convolutional layers with 3×3 kernels, 12 output channels, and the same padding method. A flattening layer then outputs a one-dimensional feature vector. An LSTM feature extraction layer with 256 hidden units and the "last" output mode models temporal features. The second input branch receives a phase noise compensation signal with dimensions [12, 1, 1]. A deep connection layer performs deep concatenation of the LSTM output features and the phase compensation signal. A custom phase calibration layer implements real-time phase noise compensation. A fully connected output layer outputs a 12-dimensional signal, corresponding to the real and imaginary parts of the six modes. The decision feedback optimization process includes: making initial predictions on the test data to obtain the equalization result, applying phase calibration to calculate the calibrated complex signal, obtaining the decision signal and updating the training target through 4-QAM hard decision, performing multiple rounds of iterative optimization to improve the equalization accuracy, realizing adaptive optimization in the testing phase, and finally converting the 12-dimensional real vector output by the network into complex signals with 6 modes for output.
2. The phase calibration method for CNN-LSTM-based mode-division multiplexing optical fiber communication according to claim 1, characterized in that, The data preprocessing process includes: separating the complex signal into real and imaginary parts to form a data format with alternating real and imaginary parts; setting sliding window parameters including tap length and phase to calculate the sample block size; generating training samples using sliding window technology, where the window size is 2 × window length + 1; and padding the boundaries with zeros to ensure that the window size of each sample is consistent.
3. The phase calibration method for CNN-LSTM-based mode-division multiplexing optical fiber communication according to claim 1, characterized in that, The iterative training and phase feedback mechanism includes: initializing the phase noise compensation vector PN to [1,1,1,1,1,1,0,0,0,0,0,0], performing the first round of network training to obtain the initial prediction result, calculating the phase noise based on the predicted complex signal and the target signal, updating the phase noise compensation vector, and repeating the training process to achieve adaptive estimation and compensation of phase noise. The phase noise estimation adopts a block processing method, estimating the phase noise by calculating the phase difference between the predicted signal and the target signal within the sample block.
4. The phase calibration method for CNN-LSTM-based mode-division multiplexing optical fiber communication according to claim 1, characterized in that, The iterative training process uses the Adam optimizer with a learning rate of 1e-4, a batch size of 128, and a gradient clipping threshold of 1. Through multiple iterations, the network parameters and phase estimation are jointly optimized. The network training adopts a data storage combination method, which combines sliding window data, phase compensation signals, and target signals, and supports batch parallel processing.
5. The phase calibration method for CNN-LSTM-based mode-division multiplexing optical fiber communication according to claim 1, characterized in that, For MDM systems with different numbers of modes: For a 4-mode system, the input dimension is adjusted to 8×(2×window size+1), and the output dimension is 8; for a 12-mode system, the input dimension is adjusted to 24×(2×window size+1), and the output dimension is 24. The number of LSTM hidden units is increased to 512. The phase calibration layer receives the feature vector and phase compensation signal from the LSTM, performs real-time phase calibration on the signal through the phase compensation factor, and outputs the calibrated signal features.
6. The phase calibration method for CNN-LSTM-based mode-division multiplexing optical fiber communication according to claim 1, characterized in that, include: The data preprocessing module is used to acquire MIMO input signals and target signals, separate complex signals into real and imaginary parts, and generate sliding window training samples. The network building block is used to construct a dual-input CNN-LSTM hybrid neural network architecture; The training optimization module is used to perform iterative training and adaptive phase estimation; the test feedback module is used to perform testing and decision feedback optimization; and the output processing module is used to convert the network output into a complex signal output.