Compensation method for nonlinear distortion of radio frequency PON system

Abstract

The invention relates to the technical field of radio frequency passive optical networks, in particular to a compensation method for nonlinear distortion of a radio frequency PON system, which is used for solving the problems that nodes cannot be converted into uniform state parameter feature vectors, data consistency and computational feasibility cannot be ensured, link similarity and influence intensity cannot be quantized, and the system performance cannot be influenced in the prior art. The importance of neighbor nodes cannot be dynamically allocated by using attention weight, and the capturing capability of a complex non-linear relationship and the robustness and efficiency of feature propagation are reduced; according to the method, entities are mapped into nodes, edges are established according to optical fiber connection, and each node is represented by using normalized state parameter feature vectors, so that data consistency and calculation feasibility are ensured, link similarity and influence intensity are quantified, importance of neighbor nodes is dynamically allocated by using attention weight, and the method is suitable for the dynamic distribution of the neighbor nodes. And the capturing capability of the complex non-linear relationship and the robustness and efficiency of feature propagation are effectively enhanced.

Description

Technical Field

[0001] This invention relates to the field of radio frequency passive optical network technology, and more specifically, to a method for compensating for nonlinear distortion in a radio frequency PON system. Background Technology

[0002] With the development of communication technology, radio frequency PON systems are widely used in 5G deployment, Internet of Things and other fields due to their advantages of high bandwidth and low latency. However, the high power of nonlinear components such as optical amplifiers in the system is prone to saturation, and nonlinear transmission media and multipath effects can cause nonlinear distortion, resulting in a decline in signal quality, an increase in bit error rate and a shorter communication distance. Among the existing compensation methods, the power back-off method is not very efficient, the negative feedback and feedforward methods have high system design and debugging requirements and are difficult to process broadband signals, and the digital predistortion technology has stringent requirements on the sampling rate of ADC / DAC, which increases the difficulty and cost of implementation.

[0003] Reference patent application CN116318210A discloses a compensation method, compensator, and system for nonlinear distortion of pulse field sources. The compensation method includes: constructing a Volterra series model for nonlinear distortion of an analog-to-digital receiver; loading the signal to be compensated received by the digital receiver into the Volterra series model to obtain a distorted signal, wherein the distorted signal carries a nonlinear distortion amount; constructing a compensation model to characterize the nonlinear distortion amount; using the compensation model to eliminate the nonlinear distortion amount in the distorted signal to obtain a compensation output; and updating the compensation kernel vector of the compensation model using the least squares method based on the compensation output, so as to eliminate the nonlinear distortion amount in the distorted signal in real time using the compensation model updated with the compensation kernel vector. This invention solves the problem of not being able to directly compensate for receiver nonlinearity without adding an additional analog-to-digital converter (ADC) for nonlinear distortion of pulse field sources. However, while the aforementioned reference patent analyzes the frequency distribution of nonlinear components in the digital receiver output signal, constructs filters to extract nonlinear quantities, and establishes a compensation model with nonlinear distortion energy as the cost function, effectively identifying and updating parameters without the need for an additional ADC to acquire the original input signal, significantly improving the spurious-free dynamic range performance of the system, it cannot transform each node into a unified state parameter feature vector, cannot guarantee data consistency and computational feasibility, cannot quantify link similarity and influence intensity, and cannot dynamically allocate the importance of neighboring nodes using attention weights, thus reducing the ability to capture complex nonlinear relationships and the robustness and efficiency of feature propagation. Furthermore, it cannot configure independent reinforcement learning agents for each type of nonlinear distortion, cannot achieve differentiated and refined compensation, cannot comprehensively consider distortion suppression effects, system performance gains, and action costs, cannot introduce Nash equilibrium mechanisms to coordinate multi-agent policy updates, and cannot achieve efficient, accurate, and scalable dynamic compensation for nonlinear distortion.

[0004] To address the aforementioned issues, we propose a method for compensating for nonlinear distortion in radio frequency PON systems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for compensating nonlinear distortion in a radio frequency PON system. This method addresses the problems of existing technologies, which cannot transform each node into a unified state parameter feature vector, cannot guarantee data consistency and computational feasibility, cannot quantify link similarity and influence intensity, cannot dynamically allocate the importance of neighboring nodes using attention weights, thus reducing the ability to capture complex nonlinear relationships and the robustness and efficiency of feature propagation. Furthermore, these technologies cannot configure independent reinforcement learning agents for each type of nonlinear distortion, cannot achieve differentiated and refined compensation, cannot comprehensively consider distortion suppression effects, system performance gains, and action costs, cannot introduce Nash equilibrium mechanisms to coordinate multi-agent policy updates, and ultimately cannot achieve efficient, accurate, and scalable dynamic compensation for nonlinear distortion.

[0006] The objective of this invention is achieved through the following technical solution: A method for compensating for nonlinear distortion in a radio frequency PON system includes the following steps: Step 1: Collect system status characteristic parameters during PON link transmission, including optical power, bit error rate, error vector amplitude, link delay and reflection loss. Perform outlier processing, filtering, time alignment and normalization preprocessing on the collected data. Step 2: Based on the preprocessed system state characteristic parameters, construct the PON link unit graph structure, determine the edges and weights according to the physical connection relationship and state differences, propagate the features through the graph attention network, and introduce a self-modulation mechanism to dynamically adjust the propagation weights. Step 3: Based on the link unit node features output by the graph attention network, a deep classification model is used to identify the nonlinear distortion type, and a spatiotemporal frequency heat map of its intensity distribution is generated by combining time series, signal frequency points and link location. Step 4: Based on the nonlinear distortion identification results and their spatiotemporal frequency distribution characteristics, configure an independent reinforcement learning agent for each type of distortion, construct an observation space based on distortion characteristics and system state, design a reward function by combining local error changes and system-level performance indicators, and update the compensation strategy of each agent using the Nash equilibrium mechanism.

[0007] The process of constructing the PON link unit graph structure and determining edges and weights based on physical connection relationships and state differences in step two includes: The physical connections between optical line terminals, optical splitters, and optical network units are obtained to form a tree-like topology. Each link unit is treated as a node in the graph, and a node set V is constructed, where each node v is a node in the graph. i∈V represents a link unit. According to the physical connection relationship, if there is a direct optical fiber connection between two link units, an undirected edge is added between their corresponding nodes to form an edge set E, forming a graph structure G=(V,E). For each node v i The system state characteristic parameters are collected, including optical power, bit error rate, error vector amplitude, link delay and reflection loss. The five state parameters are organized into a five-dimensional real vector in a fixed order. Normalization is performed on each component and mapped to the interval [0,1] as the feature vector of the node. Each edge e ij The weight of an edge is determined by the difference in the state vectors of the two nodes it connects to. The calculated weights are assigned to the corresponding edges to obtain a weighted undirected graph G=(V,E,W), where W is the set of edge weights.

[0008] In a preferred embodiment of the present invention, the process of performing feature propagation through a graph attention network and introducing a self-modulation mechanism to dynamically adjust the propagation weights in step two includes: Get each node v i The normalized state vector is used to obtain the connection relationship between nodes, forming a graph structure G=(V,E). The preset hidden layer dimension is obtained, and the trainable parameters are obtained, including the trainable weight matrix, attention vector and gating parameter vector. For each node v i The normalized state vector is multiplied by the weight matrix to obtain the initial feature representation; For each node v i and its neighboring node v j Perform the following operations: [The command is missing here, likely due to an error in the original text.] i with h j Concatenate them into a joint vector; Calculate the dot product of the joint vector and the attention vector to obtain a scalar value. Apply the LeakyReLU function to the scalar value to obtain the unnormalized attention score. For v i The unnormalized attention scores of all neighboring nodes are normalized to a set of non-negative weights by applying the Softmax function. For each edge e ij Generate a gating factor g ij The generation process is as follows: Calculate X i -X j , and dot product with the trainable weight vector θ; Obtain the edge direction feature vector d ij , and dot product with the trainable weight vector φ; Add the results of the two dot products together, input them into the Sigmoid function, and the output is g.ij ; Each node v i Its representation is updated based on the characteristics of its neighboring nodes and the composite propagation coefficient, which is the product of the attention weight and the gating factor.

[0009] In a preferred embodiment of the present invention, the process of identifying nonlinear distortion types using a deep classification model based on the link unit node features output by the graph attention network in step three includes: Obtain the original feature vector of each link unit node, which contains four components: transmit power value, modulation error vector amplitude value, spectrum envelope value, and nonlinear disturbance index value. The connection relationship between nodes in the link system constitutes a directed graph G=(V,E), where each directed edge represents the signal transmission direction. All features are normalized before input, mapped to the interval [0,1], and a normalized feature vector is generated; Graph attention networks are used to classify link nonlinear distortion. The graph attention network model updates node representations through the following steps: T1: Perform a linear transformation on the features of each node and its neighboring nodes to generate an intermediate representation; T2: For each target node, concatenate it with the intermediate representations of all its incoming neighbors, input a trainable linear function, output a scalar value, process it through the LeakyReLU activation function, and obtain the unnormalized attention score; T3: Normalize the attention scores of all incoming neighbors of the target node so that the sum of all weights is 1; T4: The intermediate representations of neighboring nodes are weighted and summed using normalized weights. The result is then processed by an activation function to generate a new feature representation for the target node. The graph attention network model employs a multi-head attention mechanism, where the outputs of multiple heads are concatenated along the feature dimension to form the final output. Multiple graph attention network layers are connected sequentially, with the output of the previous layer serving as the input of the next layer, and the final layer output serving as the embedding representation of each node. For the embedded representation of each node, a classification score is calculated using a fully connected layer. The score is converted into a probability distribution through a softmax function. The category with the highest probability is taken as the nonlinear distortion type of the node, and the nonlinear distortion type determination result of each link unit node is output.

[0010] The process of generating a spatiotemporal frequency heatmap of intensity distribution by combining time series, signal frequency points, and link location in step three includes: Obtain the embedding representation of each node at each time step and the corresponding nonlinear distortion type classification score; obtain the number of nodes in the link system, the total number of observation time steps, and the total number of discrete frequency points. A three-dimensional data structure is constructed based on the number of link nodes, the number of observation time steps, and the total number of discrete frequency points. This structure contains three dimensions: time dimension, frequency dimension, and spatial dimension. For the embedding representation of each node at each time step, the classification score of the target distortion type is extracted as the original intensity value. The dominant frequency of each node at each time step is determined. The dominant frequency corresponds to a fixed index on the frequency dimension. The intensity value of the node at that time step is written into the corresponding time, frequency, and node position in the three-dimensional structure. For each time step, calculate the arithmetic mean and standard deviation of all node strength values, and update the standardized values ​​to the 3D structure; Two-dimensional slices were extracted from the three-dimensional structure to generate heat maps, namely time-space heat map, space-frequency heat map and time-frequency heat map.

[0011] Step four involves configuring an independent reinforcement learning agent for each type of distortion and constructing an observation space based on distortion features and system state, including: To obtain the linear distortion identification results and their spatiotemporal frequency distribution characteristics, in a system with multiple types of nonlinear distortion, an independent reinforcement learning agent is configured for each type of distortion. The system is defined to contain M types of nonlinear distortion, denoted as […]. Establish M reinforcement learning agents, denoted as follows: ; Each agent The policy function performs a compensation task for the m-th type of distortion, taking the current observation state as input and outputting the probability distribution of the compensation action. Each agent receives an observation state at time step t as a fixed-dimensional vector, which is composed of three parts concatenated in sequence: the local feature vector of the m-th type of distortion, the system-level state feature vector, and the localization information vector. The three vectors are concatenated to form a complete observation state vector, which serves as the input of the m-th agent at time step t.

[0012] In a preferred embodiment of the present invention, the process of designing the reward function in step four, which combines local error changes and system-level performance indicators, includes: Obtain the intensity index of the m-th type of distortion before and after the compensation action, obtain the changes of the four system performance indicators before and after compensation and their respective weight coefficients, and obtain the cost information introduced by the compensation action. The immediate reward of the m-th agent at time step t consists of a linear combination of three parts: Distortion reduction value: Calculate the difference between the intensity index of the m-th type of distortion before and after the compensation action; System performance improvement value: Calculate the changes in the four system performance indicators before and after compensation, multiply each change by the corresponding preset weight coefficient, and sum the weighted results; Action cost: Calculate the cost introduced by the compensation action. The cost consists of the order of the compensation filter, the additional signal overhead, and the computational burden. The sum of these three items is used to calculate the cost. The three results are multiplied by non-negative weighting coefficients and then added together to form the final reward value. The instantaneous reward value R of the m-th agent at time step t is output.

[0013] In a preferred embodiment of the present invention, the process of updating the compensation strategy of each agent using the Nash equilibrium mechanism in step four includes: The following data are used as input data: the instantaneous reward sequence obtained by each agent in all time steps from the initial time to the task end time, the current policy parameters, discount factor, global state trajectory, parameter update step size and policy network structure configuration of each agent; For each agent, iterate through its instantaneous rewards at all time steps in chronological order, multiply the reward at each time step by the corresponding time power of the discount factor, and then sum all the results to obtain the cumulative reward value of the agent over the entire cycle. For each agent, the global state is received using its policy network, the action selection probability is output, the log probability value of the selected action is calculated, and the log probability value is multiplied by the cumulative reward value to obtain the adjustment direction vector of the policy parameters. For each agent, multiply its current policy parameters by the adjustment direction vector, multiply the product by the parameter update step size, add the resulting vector to the original parameters to obtain the updated policy parameters, and write the updated parameters into the policy network. For each agent, keep the policy parameters of all other agents unchanged, and compare the cumulative reward value of the agent when using the updated parameters with any other combination of parameters. If the cumulative reward value obtained by using the updated parameters is not less than the value obtained by using any other combination of parameters, then the agent is marked as policy stable. Check if all agents are marked as having a stable policy. If all are in a stable state, terminate the update process and output the updated policy parameters for all agents. Otherwise, continue to the next round of updates.

[0014] Compared with the prior art, the advantages of this invention are: (1) In this invention, by mapping entities to nodes and establishing edges based on fiber optic connections, and using normalized state parameter feature vectors to represent each node, data consistency and computational feasibility are guaranteed. Edge weights are determined based on the state differences between nodes, quantifying link similarity and influence intensity. Graph attention networks are used for feature propagation, and attention weights are used to dynamically allocate the importance of neighboring nodes. The information propagation intensity is adaptively adjusted through gating factors, effectively enhancing the ability to capture complex nonlinear relationships and the robustness and efficiency of feature propagation. (2) In this invention, by configuring an independent reinforcement learning agent for each type of nonlinear distortion, differentiated and fine compensation is achieved. Each agent constructs a high-dimensional observation space based on local distortion features, system state and positioning information to fully perceive the operating environment. The reward function integrates the distortion suppression effect, system performance gain and action cost to balance compensation efficiency and resource consumption. The Nash equilibrium mechanism is introduced to coordinate the policy update of multiple agents, ensure optimization stability, avoid policy conflicts, improve system convergence and robustness, and finally achieve efficient, accurate and scalable nonlinear distortion dynamic compensation. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the steps for updating node representation in Embodiment 1 of the present invention. Detailed Implementation

[0016] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0017] Example 1: As Figure 1 and Figure 2 As shown, the present invention proposes a method for compensating for nonlinear distortion in a radio frequency PON system, comprising the following steps: Step 1: Collect system status characteristic parameters during PON link transmission, including optical power, bit error rate, error vector amplitude, link delay and reflection loss. Perform outlier processing, filtering, time alignment and normalization preprocessing on the collected data. This system collects system state characteristic parameters such as optical power, bit error rate, error vector amplitude, link delay, and reflection loss during PON link transmission. It then performs outlier processing, filtering, time alignment, and normalization preprocessing to comprehensively acquire key information reflecting the link's operational status, ensuring data integrity and reliability. Outlier processing eliminates measurement errors and transient interference, improving data quality. Filtering suppresses noise fluctuations, enhancing signal stability. Time alignment ensures synchronization of multi-source parameters under a unified time reference, avoiding timing misalignment. Normalization unifies the numerical scale of various features, eliminating the impact of dimensional differences on subsequent models. This overall preprocessing workflow provides an accurate, consistent, and computable data foundation for model training and state analysis, improving system convergence, stability, and environmental adaptability.

[0018] Step 2: Based on the preprocessed system state characteristic parameters, construct the PON link unit graph structure, determine the edges and weights according to the physical connection relationship and state differences, propagate the features through the graph attention network, and introduce a self-modulation mechanism to dynamically adjust the propagation weights. Step two involves constructing the PON link cell graph structure and determining edges and weights based on physical connection relationships and state differences. The physical connections between optical line terminals, optical splitters, and optical network units are obtained to form a tree-like topology. Each link unit is treated as a node in the graph, and a node set V is constructed, where each node v is a node in the graph. i ∈V represents a link unit. According to the physical connection relationship, if there is a direct optical fiber connection between two link units, an undirected edge is added between their corresponding nodes to form an edge set E, forming a graph structure G=(V,E). For each node v i The system state characteristic parameters are collected, including optical power, bit error rate, error vector amplitude, link delay and reflection loss. The five state parameters are organized into a five-dimensional real vector in a fixed order. Normalization is performed on each component and mapped to the interval [0,1] as the feature vector of the node. Each edge e ij The weight of a node is determined by the difference in the state vectors of the two nodes it connects to, and the calculation formula is as follows: ,in For edge e ij The weights are scalars, with values ​​ranging from (0,1]. They are nodes v i With v j Normalized state vector, Let be the Euclidean norm of the vector difference. This is a positive real scalar parameter, greater than 0, set during model initialization and not updated during training. This weight is used for the initial assignment of edges in the graph structure. The calculated weights are assigned to the corresponding edges to obtain a weighted undirected graph G=(V,E,W), where W is the set of edge weights, which is used as the input to the subsequent graph neural network. Step two involves using a graph attention network to propagate features and introducing a self-modulation mechanism to dynamically adjust the propagation weights. Get each node v i The normalized state vector is used to obtain the connection relationship between nodes, forming a graph structure G=(V,E). The preset hidden layer dimension is obtained, and the trainable parameters are obtained, including the trainable weight matrix, attention vector and gating parameter vector. For each node v i The normalized state vector is multiplied by the weight matrix to obtain the initial feature representation; For each node v i and its neighboring node v j Perform the following operations: [The command is missing here, likely due to an error in the original text.] i with h j Concatenate them into a joint vector; Calculate the dot product of the joint vector and the attention vector to obtain a scalar value. Apply the LeakyReLU function to the scalar value to obtain the unnormalized attention score. For v i The unnormalized attention scores of all neighboring nodes are normalized to a set of non-negative weights by applying the Softmax function, which sums to 1. For each edge e ij Generate a gating factor g ij It is used to adjust the intensity of information propagation, and the generation process is as follows: Calculate X i -X j , and dot product with the trainable weight vector θ; Obtain the edge direction feature vector d ij , and dot product with the trainable weight vector φ; Add the results of the two dot products together, input them into the Sigmoid function, and the output is g. ij ; Each node v i Its representation is updated based on the characteristics of its neighboring nodes and the composite propagation coefficient, which is the product of the attention weight and the gating factor. The update rule is as follows: ,in For node v i Updated feature representation, For the activation function, ReLU is used. For node v i The neighborhood group, To normalize attention weights, As the gating factor, For neighbor node v j The current feature representation, For composite propagation coefficients, this update operation is the only feature propagation method in graph neural networks; In PON, entities such as optical line terminals, splitters, and optical network units are mapped as nodes in a graph, and edges are established based on actual fiber connection relationships to construct a link unit graph structure, forming a tree diagram that reflects the real topology. Each node is represented by a feature vector consisting of five state parameters (optical power, bit error rate, error vector magnitude, link delay, and reflection loss), which is normalized to ensure data consistency and computability. The edge weights are determined based on the Euclidean distance between the state vectors of connected nodes, reflecting the degree of state difference between nodes and thus quantifying link similarity and influence strength. On this basis, a graph attention network is used for feature propagation. Attention scores are calculated through linear transformation and dot product operations, and converted into attention weights by the Softmax function, enabling each node to dynamically allocate the importance of its neighboring nodes. At the same time, a gating factor is generated through the Sigmoid function to achieve adaptive adjustment of information propagation strength. Finally, the node features are updated by combining the composite propagation coefficient formed by the attention weights and the gating factor, enhancing the model's ability to capture complex nonlinear relationships and its robustness, and improving the effectiveness of feature propagation.

[0019] Step 3: Based on the link unit node features output by the graph attention network, a deep classification model is used to identify the nonlinear distortion type, and a spatiotemporal frequency heat map of its intensity distribution is generated by combining time series, signal frequency points and link location. Step three involves using a deep classification model to identify nonlinear distortion types based on the link unit node features output by the graph attention network. Obtain the original feature vector of each link unit node, which contains four components: transmit power value, modulation error vector amplitude value, spectrum envelope value, and nonlinear disturbance index value. The connection relationship between nodes in the link system constitutes a directed graph G=(V,E), where each directed edge represents the signal transmission direction. All features are normalized before input, mapped to the interval [0,1], and a normalized feature vector is generated; Graph attention networks are used to classify link nonlinear distortion. The graph attention network model updates node representations through the following steps: T1: Perform a linear transformation on the features of each node and its neighboring nodes to generate an intermediate representation; T2: For each target node, concatenate it with the intermediate representations of all its incoming neighbors, input a trainable linear function, output a scalar value, process it through the LeakyReLU activation function, and obtain the unnormalized attention score; T3: Normalize the attention scores of all incoming neighbors of the target node so that the sum of all weights is 1; T4: The intermediate representations of neighboring nodes are weighted and summed using normalized weights. The result is then processed by an activation function to generate a new feature representation for the target node. The graph attention network model employs a multi-head attention mechanism. Each attention head uses independent trainable parameters to execute the above information propagation process in parallel. The outputs of multiple heads are concatenated along the feature dimension to form the final output. Multiple graph attention network layers are connected sequentially, with the output of the previous layer serving as the input of the next layer, and the output of the final layer serving as the embedding representation of each node. For the embedded representation of each node, a classification score is calculated using a fully connected layer. The score is converted into a probability distribution through a softmax function. The category with the highest probability is taken as the non-linear distortion type of the node. The non-linear distortion type determination result of each link unit node is output for subsequent compensation strategy selection. The model training adopts the cross-entropy loss function, and the optimization process minimizes the difference between the predicted distribution of all nodes and the true label. Step three, which combines time series data, signal frequency points, and link location to generate a spatiotemporal frequency heatmap of its intensity distribution, includes: Obtain the embedding representation of each node at each time step and the corresponding nonlinear distortion type classification score; obtain the number of nodes in the link system, the total number of observation time steps, and the total number of discrete frequency points. A three-dimensional data structure is constructed based on the number of link nodes, the number of observation time steps, and the total number of discrete frequency points to record the distribution of nonlinear distortion intensity in time, frequency, and spatial dimensions. This structure contains three dimensions: The time dimension represents the total number of time steps observed. Frequency dimension represents the total number of discrete frequency points; Spatial dimension, representing the total number of link nodes; Each element in the structure represents the nonlinear distortion intensity value at a specific time, frequency, and node. For the embedding representation of each node at each time step, the classification score of the target distortion type is extracted as the original intensity value. The dominant frequency of each node at each time step is determined. The dominant frequency corresponds to a fixed index on the frequency dimension. The intensity value of the node at that time step is written into the corresponding time, frequency, and node position in the three-dimensional structure. For each time step, calculate the arithmetic mean and standard deviation of all node intensity values, and standardize the intensity values ​​using the mean and standard deviation so that the intensity values ​​at that time step have a mean of 0 and a standard deviation of 1. Update the standardized values ​​to the 3D structure. Generating a heatmap by extracting two-dimensional slices from a three-dimensional structure: Time-space heatmap: Extract a two-dimensional matrix composed of time and space dimensions from a certain index of a fixed frequency dimension; Spatial-frequency heatmap: Extract a two-dimensional matrix consisting of spatial and frequency dimensions from a fixed time dimension index; Time-frequency heatmap: Extract a two-dimensional matrix composed of time and frequency dimensions from a fixed spatial dimension index; All heatmaps are displayed using continuous color levels, and bilinear interpolation is used to fill non-integer coordinates to ensure the continuity of the image in both spatial and frequency directions. This method utilizes deep learning to identify nonlinear distortion types in PON link units and generates a spatiotemporal frequency heatmap to display the distortion intensity distribution. Its advantages lie in using node transmit power, modulation error vector amplitude, spectral envelope, and nonlinear perturbation index as feature vectors to comprehensively describe node states. All features are normalized before being input into the model to ensure data consistency and algorithm stability. The method employs a graph attention network (GAT) combined with a multi-head attention mechanism to effectively capture complex nonlinear relationships between link units, enhancing the model's expressive power. Fully connected layers calculate classification scores and convert them into probability distributions using a softmax function, improving recognition accuracy. Furthermore, the method constructs a three-dimensional data structure to record and analyze the intensity distribution of nonlinear distortion. Standardizing the intensity values ​​at each time step facilitates cross-time period comparison and analysis. The final generated heatmap helps to intuitively understand the distribution patterns of distortion across different dimensions, and bilinear interpolation is used to improve image continuity and interpretability, thereby optimizing PON system performance.

[0020] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that: like Figure 1 As shown, step four: Based on the nonlinear distortion identification results and their spatiotemporal frequency distribution characteristics, configure an independent reinforcement learning agent for each type of distortion, construct an observation space based on distortion characteristics and system state, design a reward function by combining local error changes and system-level performance indicators, and update the compensation strategy of each agent using the Nash equilibrium mechanism. Step four involves configuring an independent reinforcement learning agent for each type of distortion and constructing an observation space based on distortion features and system state. To obtain the linear distortion identification results and their spatiotemporal frequency distribution characteristics, in a system with multiple types of nonlinear distortion, an independent reinforcement learning agent is configured for each type of distortion. The system is defined to contain M types of nonlinear distortion, denoted as […]. Establish M reinforcement learning agents, denoted as follows: ; Each agent The policy function of the m-th type of distortion compensation task takes the current observation state as input and outputs the probability distribution of the compensation action. The manifestation mechanism of the m-th type of distortion is determined by its physical cause, the interference mode is determined by the system signal path, the scope of action is defined by the resource allocation strategy, and the policy function is modeled with independent parameters. The observation state received by each agent at time step t is a fixed-dimensional vector, which is composed of three parts concatenated in sequence: The local feature vector of the m-th type of distortion has a dimension of 5, and its components are: distortion intensity spectrum distribution value, time domain transient rate of change value, frequency response change slope value, nonlinear index value, and historical residual compensation error trend value. The system-level state feature vector has a dimension of 4, and its components are, in order: normalized bit error rate, normalized signal-to-noise ratio, normalized system processing delay, and normalized throughput. The positioning information vector has a dimension of 3, and its components are, in order: subcarrier index value, time slice number value, and antenna array element number value. The three vectors are concatenated to form a complete observation state vector with a dimension of 12, which serves as the input of the m-th agent at time step t. Step four, which involves designing the reward function by combining local error variations and system-level performance metrics, includes: Obtain the intensity index of the m-th type of distortion before and after the compensation action, obtain the changes of the four system performance indicators before and after compensation and their respective weight coefficients, and obtain the cost information introduced by the compensation action. The immediate reward of the m-th agent at time step t consists of a linear combination of three parts: Distortion reduction value: Calculate the difference between the intensity index of the m-th type of distortion before and after the compensation action; System performance improvement value: Calculate the changes in the four system performance indicators before and after compensation, multiply each change by the corresponding preset weight coefficient, with weight coefficients of 0.3, 0.3, 0.2 and 0.2 respectively, and sum the weighted results; Action cost: Calculate the cost introduced by the compensation action. The cost consists of the order of the compensation filter, the additional signal overhead, and the computational burden. The sum of these three items is used to calculate the cost. Each of the three results is multiplied by a non-negative weighting coefficient. β and γ are then added together to form the final reward value; The reward function expression is as follows: ,in The instantaneous reward for the m-th agent at time step t. The weighting coefficients for the distortion reduction term are... The intensity reduction of the m-th type of distortion, The weighting coefficients for system performance improvement items. This is a weighted improvement in system performance metrics. This represents the weighting coefficient for the action cost item. To compensate for the overall cost of the action; Output the instantaneous reward value R of the m-th agent at time step t, which is used for subsequent reinforcement learning training or policy adjustment; Step four involves updating the compensation strategies of each agent using the Nash equilibrium mechanism, which includes: Obtain the following data as input data: the sequence of instantaneous rewards obtained by each agent at all time steps from the initial time to the task end time; The policy parameters for each agent are initially set to preset random numbers. The discount factor has a value of 0.99. The global state trajectory contains the state information of all nodes at all time steps; The parameter update step size is set to 0.001. Policy network architecture configuration: It contains three fully connected layers, each with 128 neurons, and the activation function is a modified linear unit; For each agent, iterate through its instantaneous rewards at all time steps in chronological order, multiply the reward at each time step by the corresponding time power of the discount factor, and then sum all the results to obtain the cumulative reward value of the agent over the entire cycle. For each agent, the global state is received using its policy network, the action selection probability is output, the log probability value of the selected action is calculated, and the log probability value is multiplied by the cumulative reward value to obtain the adjustment direction vector of the policy parameters. For each agent, multiply its current policy parameters by the adjustment direction vector, multiply the product by the parameter update step size, add the resulting vector to the original parameters to obtain the updated policy parameters, and write the updated parameters into the policy network. For each agent, keep the policy parameters of all other agents unchanged, and compare the cumulative reward value of the agent when using the updated parameters with any other combination of parameters. If the cumulative reward value obtained by using the updated parameters is not less than the value obtained by using any other combination of parameters, then the agent is marked as policy stable. Check if all agents are marked as having a stable policy. If all are in a stable state, terminate the update process and output the updated policy parameters for all agents. Otherwise, continue to the next round of updates. By configuring independent reinforcement learning agents for each type of nonlinear distortion, refined compensation for different types of distortion is achieved. Each agent constructs a high-dimensional observation space based on local distortion features, system-level state, and location information to comprehensively reflect the operating environment. The reward function integrates three indicators: distortion reduction degree, system performance improvement, and action cost, balancing compensation effect and resource consumption. By introducing a Nash equilibrium mechanism, the stability of policy updates for each agent is ensured during multi-agent collaborative optimization, policy conflicts are avoided, and the convergence and robustness of the overall system are improved, ultimately achieving efficient, accurate, and scalable dynamic compensation for nonlinear distortion.

[0021] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for compensating for nonlinear distortion in a radio frequency PON system, characterized in that, Includes the following steps: Step 1: Collect system status characteristic parameters during PON link transmission, including optical power, bit error rate, error vector amplitude, link delay and reflection loss. Perform outlier processing, filtering, time alignment and normalization preprocessing on the collected data. Step 2: Based on the preprocessed system state characteristic parameters, construct the PON link unit graph structure, determine the edges and weights according to the physical connection relationship and state differences, propagate the features through the graph attention network, and introduce a self-modulation mechanism to dynamically adjust the propagation weights. Step 3: Based on the link unit node features output by the graph attention network, a deep classification model is used to identify the nonlinear distortion type, and a spatiotemporal frequency heat map of its intensity distribution is generated by combining time series, signal frequency points and link location. Step 4: Based on the nonlinear distortion identification results and their spatiotemporal frequency distribution characteristics, configure an independent reinforcement learning agent for each type of distortion, construct an observation space based on distortion characteristics and system state, design a reward function by combining local error changes and system-level performance indicators, and update the compensation strategy of each agent using the Nash equilibrium mechanism.

2. The method for compensating for nonlinear distortion in a radio frequency PON system according to claim 1, characterized in that, The process of constructing the PON link unit graph structure and determining edges and weights based on physical connection relationships and state differences in step two includes: The physical connections between optical line terminals, optical splitters, and optical network units are obtained to form a tree-like topology. Each link unit is treated as a node in the graph, and a node set V is constructed, where each node v is a node in the graph. i ∈V represents a link unit. According to the physical connection relationship, if there is a direct optical fiber connection between two link units, an undirected edge is added between their corresponding nodes to form an edge set E, forming a graph structure G=(V,E). For each node v i The system state characteristic parameters are collected, including optical power, bit error rate, error vector amplitude, link delay and reflection loss. The five state parameters are organized into a five-dimensional real vector in a fixed order. Normalization is performed on each component and mapped to the interval [0,1] as the feature vector of the node. Each edge e ij The weight of an edge is determined by the difference in the state vectors of the two nodes it connects to. The calculated weights are assigned to the corresponding edges to obtain a weighted undirected graph G=(V,E,W), where W is the set of edge weights.

3. The method for compensating for nonlinear distortion in a radio frequency PON system according to claim 2, characterized in that, The process of feature propagation through a graph attention network and dynamically adjusting the propagation weights by introducing a self-modulation mechanism in step two includes: Get each node v i The normalized state vector is used to obtain the connection relationship between nodes, forming a graph structure G=(V,E). The preset hidden layer dimension is obtained, and the trainable parameters are obtained, including the trainable weight matrix, attention vector and gating parameter vector. For each node v i The normalized state vector is multiplied by the weight matrix to obtain the initial feature representation; For each node v i and its neighboring node v j Perform the following operations: [The command is missing from the original text] i with h j Concatenate them into a joint vector; Calculate the dot product of the joint vector and the attention vector to obtain a scalar value. Apply the LeakyReLU function to the scalar value to obtain the unnormalized attention score. For v i The unnormalized attention scores of all neighboring nodes are normalized to a set of non-negative weights by applying the Softmax function. For each edge e ij Generate a gating factor g ij The generation process is as follows: Calculate X i -X j , and dot product with the trainable weight vector θ; Obtain the edge direction feature vector d ij , and dot product with the trainable weight vector φ; Add the results of the two dot products together, input them into the Sigmoid function, and the output is g. ij ; Each node v i Its representation is updated based on the characteristics of its neighboring nodes and the composite propagation coefficient, which is the product of the attention weight and the gating factor.

4. The method for compensating nonlinear distortion in a radio frequency PON system according to claim 1, characterized in that, The process of identifying nonlinear distortion types using a deep classification model based on the link unit node features output by the graph attention network in step three includes: Obtain the original feature vector of each link unit node, which contains four components: transmit power value, modulation error vector amplitude value, spectrum envelope value, and nonlinear disturbance index value. The connection relationship between nodes in the link system constitutes a directed graph G=(V,E), where each directed edge represents the signal transmission direction. All features are normalized before input, mapped to the interval [0,1], and a normalized feature vector is generated; Graph attention networks are used to classify link nonlinear distortion. The graph attention network model updates node representations through the following steps: T1: Perform a linear transformation on the features of each node and its neighboring nodes to generate an intermediate representation; T2: For each target node, concatenate it with the intermediate representations of all its incoming neighbors, input a trainable linear function, output a scalar value, process it through the LeakyReLU activation function, and obtain the unnormalized attention score; T3: Normalize the attention scores of all incoming neighbors of the target node; T4: The intermediate representations of neighboring nodes are weighted and summed using normalized weights. The result is then processed by an activation function to generate a new feature representation for the target node. The graph attention network model employs a multi-head attention mechanism, where the outputs of multiple heads are concatenated along the feature dimension to form the final output. Multiple graph attention network layers are connected sequentially, with the output of the previous layer serving as the input of the next layer, and the final layer output serving as the embedding representation of each node. For the embedded representation of each node, a classification score is calculated using a fully connected layer. The score is converted into a probability distribution through a softmax function. The category with the highest probability is taken as the nonlinear distortion type of the node, and the nonlinear distortion type determination result of each link unit node is output.

5. The method for compensating for nonlinear distortion in a radio frequency PON system according to claim 4, characterized in that, The process of generating a spatiotemporal frequency heatmap of intensity distribution by combining time series, signal frequency points, and link location in step three includes: Obtain the embedding representation of each node at each time step and the corresponding nonlinear distortion type classification score; obtain the number of nodes in the link system, the total number of observation time steps, and the total number of discrete frequency points. A three-dimensional data structure is constructed based on the number of link nodes, the number of observation time steps, and the total number of discrete frequency points. This structure contains three dimensions: time dimension, frequency dimension, and spatial dimension. For the embedding representation of each node at each time step, the classification score of the target distortion type is extracted as the original intensity value. The dominant frequency of each node at each time step is determined. The dominant frequency corresponds to a fixed index on the frequency dimension. The intensity value of the node at that time step is written into the corresponding time, frequency, and node position in the three-dimensional structure. For each time step, calculate the arithmetic mean and standard deviation of all node strength values, and update the standardized values ​​to the 3D structure; Two-dimensional slices were extracted from the three-dimensional structure to generate heat maps, namely time-space heat map, space-frequency heat map and time-frequency heat map.

6. The method for compensating for nonlinear distortion in a radio frequency PON system according to claim 1, characterized in that, Step four involves configuring an independent reinforcement learning agent for each type of distortion and constructing an observation space based on distortion features and system state, including: To obtain the linear distortion identification results and their spatiotemporal frequency distribution characteristics, in a system with multiple types of nonlinear distortion, an independent reinforcement learning agent is configured for each type of distortion. The system is defined to contain M types of nonlinear distortion, denoted as […]. Establish M reinforcement learning agents, denoted as follows: ; Each agent The policy function performs a compensation task for the m-th type of distortion, taking the current observation state as input and outputting the probability distribution of the compensation action. Each agent receives an observation state at time step t as a fixed-dimensional vector, which is composed of three parts concatenated in sequence: the local feature vector of the m-th type of distortion, the system-level state feature vector, and the localization information vector. The three vectors are concatenated to form a complete observation state vector, which serves as the input of the m-th agent at time step t.

7. The method for compensating for nonlinear distortion in a radio frequency PON system according to claim 6, characterized in that, The process of designing the reward function by combining local error changes and system-level performance indicators in step four includes: Obtain the intensity index of the m-th type of distortion before and after the compensation action, obtain the changes of the four system performance indicators before and after compensation and their respective weight coefficients, and obtain the cost information introduced by the compensation action. The immediate reward of the m-th agent at time step t consists of a linear combination of three parts: Distortion reduction value: Calculate the difference between the intensity index of the m-th type of distortion before and after the compensation action; System performance improvement value: Calculate the changes in the four system performance indicators before and after compensation, multiply each change by the corresponding preset weight coefficient, and sum the weighted results; Action cost: Calculate the cost introduced by the compensation action. The cost consists of the order of the compensation filter, the additional signal overhead, and the computational burden. The sum of these three items is used to calculate the cost. The three results are multiplied by non-negative weighting coefficients and then added together to form the final reward value. The instantaneous reward value R of the m-th agent at time step t is output.

8. A method for compensating for nonlinear distortion in a radio frequency PON system according to claim 7, characterized in that, The process of updating the compensation strategy of each agent using the Nash equilibrium mechanism in step four includes: The following data are used as input data: the instantaneous reward sequence obtained by each agent in all time steps from the initial time to the task end time, the current policy parameters, discount factor, global state trajectory, parameter update step size and policy network structure configuration of each agent; For each agent, iterate through its instantaneous rewards at all time steps in chronological order, multiply the reward at each time step by the corresponding time power of the discount factor, and then sum all the results to obtain the cumulative reward value of the agent over the entire cycle. For each agent, the global state is received using its policy network, the action selection probability is output, the log probability value of the selected action is calculated, and the log probability value is multiplied by the cumulative reward value to obtain the adjustment direction vector of the policy parameters. For each agent, multiply its current policy parameters by the adjustment direction vector, multiply the product by the parameter update step size, add the resulting vector to the original parameters to obtain the updated policy parameters, and write the updated parameters into the policy network. For each agent, keep the policy parameters of all other agents unchanged, and compare the cumulative reward value of the agent when using the updated parameters with any other combination of parameters. If the cumulative reward value obtained by using the updated parameters is not less than the value obtained by using any other combination of parameters, then the agent is marked as policy stable. Check if all agents are marked as having a stable policy. If all are in a stable state, terminate the update process and output the updated policy parameters for all agents. Otherwise, continue to the next round of updates.

Images