A method and system for real-time performance monitoring and intelligent optimization of optical transmission systems
By combining graph attention networks and deep deterministic policy gradient algorithms, real-time performance monitoring and intelligent optimization of optical transmission systems are achieved, solving the problem of lack of accurate monitoring and adaptive optimization in existing technologies, and improving the stability and reliability of the link.
Patent Information
- Application Number
- CN202611200879.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing optical transmission systems lack real-time and accurate performance monitoring and rapid adaptive optimization capabilities, making it difficult to effectively identify different types of performance degradation and perform targeted compensation. Traditional methods rely on single-threshold alarms and lack early warning capabilities.
A link quality assessment model based on graph attention networks is adopted. By collecting multi-dimensional performance data in parallel, quality scores and degradation feature description vectors are generated. Combined with a deep deterministic policy gradient algorithm, the modulation format switching and compensation parameter adjustment are dynamically decided to realize real-time performance monitoring and intelligent optimization of the optical transmission system.
It enables real-time and accurate monitoring and rapid adaptive optimization of optical transmission system link quality, can identify different types of degradation and perform targeted compensation, thereby improving the stability and reliability of the link.
Smart Images

Figure CN123123764A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical transmission technology, and in particular to a method and system for real-time performance monitoring and intelligent optimization of optical transmission systems. Background Technology
[0002] With the continuous growth in bandwidth demands for data center interconnects and backbone transmission networks, optical transmission systems are evolving towards higher speeds and denser wavelength division multiplexing (WDM). Link quality is frequently affected by multiple factors, including dispersion, nonlinear effects, temperature drift, and device aging, resulting in frequent performance fluctuations. Traditional optical performance monitoring methods primarily rely on single-threshold alarms such as optical signal-to-noise ratio (SNR) and bit error rate (BER). By the time an alarm is triggered, the link is often already severely degraded or on the verge of failure, lacking early warning and proactive optimization capabilities. In recent years, the application of digital signal processing technology in optical transceiver modules has matured, making modulation format adaptation and online adjustment of compensation parameters possible. However, existing solutions often use lookup tables or empirical formulas for parameter switching, making it difficult to fully utilize the correlation information between multi-dimensional performance parameters for accurate decision-making. Furthermore, the optimal compensation strategies required for different types of performance degradation differ significantly. For example, dispersion-dominated degradation and nonlinear-dominated degradation have drastically different requirements for the direction of compensation parameter adjustment, and existing methods generally lack the ability to finely identify degradation types and generate targeted strategies. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for real-time performance monitoring and intelligent optimization of optical transmission systems, in order to overcome the shortcomings of existing technologies and achieve real-time and accurate monitoring and rapid adaptive optimization of the link quality of optical transmission systems.
[0004] One embodiment of this application provides a method for real-time performance monitoring and intelligent optimization of an optical transmission system, the method comprising: The input optical power, bias current, temperature, voltage, and dispersion compensation forward error correction code rate of the optical transceiver module in the optical transmission system are collected in parallel with a fixed sampling period to generate a multi-dimensional performance time series dataset. The multidimensional performance time series dataset is input into the link quality assessment model built on graph attention network, the coupling features between each performance parameter are extracted, and the quality score and degradation feature description vector reflecting the current link health are generated by combining the historical decay trend. Based on the degradation type and severity in the degradation feature description vector, the modulation format switching level and the digital signal processing compensation parameter adjustment amount are dynamically decided by the deep deterministic strategy gradient algorithm, generating an optimized action set containing modulation format instructions and compensation parameter instructions; The optimized action set is sent to the digital signal processing unit of the optical transceiver module to perform online switching of modulation format and dynamic updating of dispersion compensation and nonlinear compensation parameters, thereby realizing real-time performance monitoring and intelligent optimization of the optical transmission system.
[0005] Another embodiment of this application provides a real-time performance monitoring and intelligent optimization system for an optical transmission system, the system comprising: The acquisition module is used to collect the input optical power, bias current, temperature, voltage and dispersion compensation forward error correction code rate of the optical transceiver module in the optical transmission system in parallel at a fixed sampling period, and generate a multi-dimensional performance time series dataset. The extraction module is used to input the multidimensional performance time series dataset into the link quality assessment model built based on graph attention network, extract the coupling features between various performance parameters, and generate a quality score and degradation feature description vector reflecting the current link health by combining historical decay trends. The decision module is used to dynamically decide the modulation format switching level and the digital signal processing compensation parameter adjustment amount based on the degradation type and severity in the degradation feature description vector, and generate an optimized action set containing modulation format instructions and compensation parameter instructions. The execution module is used to send the optimized action set to the digital signal processing unit of the optical transceiver module, perform online switching of modulation format and dynamic updating of dispersion compensation and nonlinear compensation parameters, and realize real-time performance monitoring and intelligent optimization of the optical transmission system.
[0006] One or more technical solutions proposed in this application have at least the following technical effects: they can realize real-time and accurate monitoring and rapid adaptive optimization of the link quality of optical transmission systems. Attached Figure Description
[0007] Figure 1 Hardware structure block diagram of a computer terminal for a real-time performance monitoring and intelligent optimization method for an optical transmission system provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a real-time performance monitoring and intelligent optimization method for an optical transmission system provided in an embodiment of the present invention; Figure 3 A flowchart illustrating another method for real-time performance monitoring and intelligent optimization of an optical transmission system provided in an embodiment of the present invention; Figure 4 A flowchart illustrating another method for real-time performance monitoring and intelligent optimization of an optical transmission system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a real-time performance monitoring and intelligent optimization system for an optical transmission system provided in an embodiment of the present invention. Detailed Implementation
[0008] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0009] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.
[0010] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.
[0011] This invention first provides a method for real-time performance monitoring and intelligent optimization of an optical transmission system. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.
[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a real-time performance monitoring and intelligent optimization method for an optical transmission system provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.
[0013] See Figures 2-4 The present invention provides a method for real-time performance monitoring and intelligent optimization of an optical transmission system, which may include the following steps: S201 uses a fixed sampling period to collect the input optical power, bias current, temperature, voltage, and dispersion compensation forward error correction code rate of the optical transceiver module in the optical transmission system in parallel, generating a multi-dimensional performance time series dataset. Specifically, a sampling clock can be configured in the digital diagnostic monitoring interface of the optical transceiver module, a fixed sampling period can be set, and read requests for four analog quantities—input optical power, bias current, temperature, and voltage—can be sent to the module's internal register simultaneously to generate a set of raw photoelectric analog readings. The core of this step is to use the built-in digital diagnostic monitoring interface of the optical transceiver module to collect four key photoelectric analog quantities in parallel with a unified time base. This ensures that each parameter is sampled at the same time, providing a time-consistent raw data foundation for subsequent multi-dimensional performance analysis. The specific implementation method is as follows: The digital diagnostic monitoring interface is a standardized monitoring interface integrated within the optical transceiver module. It is used to read various analog parameters during module operation in real time. Internally, it integrates multiple analog-to-digital conversion channels, converting physical quantities such as optical power, current, temperature, and voltage into digital register values for external controllers to read. The sampling clock is generated by a programmable timer within the interface. A precise sampling period can be set by configuring the clock divider register. The selection of the sampling period must balance real-time performance and system load; a period that is too short will increase the burden on bus communication, while a period that is too long will fail to capture transient performance fluctuations.
[0014] In the example, the sampling period T_s is set to 100ms, meaning 10 sets of data are collected per second. This sampling frequency is sufficient to capture performance fluctuations in the optical link at the millisecond level or higher, without putting excessive pressure on the module's internal bus. After the sampling clock is configured, at the beginning of each sampling period, the controller simultaneously initiates read requests to four register addresses within the module via the serial bus. These four registers correspond to the input optical power, laser bias current, module internal temperature, and supply voltage, respectively. The input optical power is obtained by sampling the photodiode at the module's receiving end, which is converted into a voltage signal by a transimpedance amplifier, reflecting the intensity level of the received optical signal. The bias current is the DC bias value output by the laser driver circuit, directly related to the laser's luminous power and aging state. The temperature is obtained by sampling the thermistor inside the module, reflecting the operating environment temperature of the optical device; temperature changes directly affect the stability of the laser's wavelength and optical power. The voltage is the voltage value of the module's power supply pins, reflecting the power supply quality; voltage fluctuations may cause the laser's operating point to shift.
[0015] The four analog quantities are read using a parallel triggering and serial feedback method, with strict alignment of trigger times and time deviation controlled within 10μs to ensure that the four parameters correspond to the same physical moment. The raw data obtained is a 16-bit unsigned integer, stored in four registers respectively, forming a set of raw photoelectric analog quantity readings. For example, in a certain sampling, the raw reading of input optical power is 0x2710, the raw reading of bias current is 0x1A80, the raw reading of temperature is 0x0FA0, and the raw reading of voltage is 0x3333. These raw values have not yet undergone range conversion and are only the digital encoding results of analog-to-digital conversion.
[0016] In the digital signal processing unit of the optical transceiver module, the bit error rate count before error correction is read from the forward error correction decoder with the same sampling period, and the bit error rate count is divided by the total number of error correction codewords to obtain the dispersion-compensated forward error correction bit rate, generating a set of original bit error rate readings. The core of this step is to obtain the bit error statistics of the forward error correction decoder before error correction within the same sampling period as the photoelectric analog quantity, and calculate the bit error rate before dispersion compensation. This parameter directly reflects the transmission quality of the optical link and is a core indicator for evaluating the health status of the link. The specific implementation method is as follows: The digital signal processing unit (DSP) is the core unit responsible for signal processing in the optical transceiver module, integrating functions such as dispersion compensation, nonlinearity compensation, and forward error correction (FEC) encoding / decoding. The FEC decoder is located at the back end of the receiving link. Its working principle is to add redundant check bits to the information data at the transmitting end according to specific encoding rules, and then correct the bit errors generated during transmission at the receiving end using a decoding algorithm, thereby improving the system's fault tolerance. The bit error rate before correction refers to the original bit error rate before the data enters the decoder. This metric accurately reflects the transmission quality of the optical link's physical layer, is unaffected by the masking effect of error correction algorithms, and is of great significance for early warning of link degradation.
[0017] The bit error rate (BER) count is accumulated in real time by an BER statistics counter inside the forward error correction decoder. The counter is reset to zero at the beginning of each sampling period and latched at the end of the period. The sampling period is consistent with the aforementioned photoelectric analog quantity sampling period, which is also 100ms, ensuring strict alignment of the BER data and the photoelectric analog quantity data on the time axis. At the end of each sampling period, the controller reads the BER count register and the total codeword count register from the decoder. The BER count register records the total number of erroneous bits detected in this period, and the total codeword count register records the total number of bits processed in this period.
[0018] The formula for calculating the forward error rate (BER) with dispersion compensation is BER_pre = N_err / N_total, where BER_pre represents the bit error rate before dispersion compensation, N_err represents the number of erroneous bits detected in one sampling period, and N_total represents the total number of bits transmitted in the same period. In the example, assuming the optical transmission system has a line rate of 100Gbps, the total number of bits N_total in a 100ms sampling period is approximately 1 × 10^10 bits. If the number of erroneous bits detected in this period, N_err, is 2 × 10^4 bits, then the calculated BER_pre with dispersion compensation is 2 × 10^-6. This BER value is within the correctable range of forward error correction, but it is already higher than the 1 × 10^-8 level during normal system operation, indicating a certain degree of link degradation. The calculated BER value, along with the corresponding sampling timestamp, is stored in the raw BER reading set for subsequent data alignment and processing.
[0019] Align the four parameter values in the original set of photoelectric analog quantity readings with the original set of bit error rate readings according to the sampling timestamp, and perform range conversion and unit conversion on each parameter value to convert it into standard physical dimensions, generating a dimensionally normalized five-dimensional sampling record; The core of this step is to precisely align the five types of data from different acquisition paths according to their timestamps, and then convert the raw register values into standard dimensions with actual physical meaning using a range conversion formula, forming a unified format of five-dimensional sampling records. This provides standardized data for subsequent model input. The specific implementation method is as follows: Timestamp alignment is a crucial step in ensuring the validity of multidimensional data. Since photoelectric analog quantities and bit error rates originate from different acquisition modules, although the sampling period is the same, there may be slight time shifts. Therefore, alignment is required based on the timestamp of the sampling moment. Alignment employs the nearest neighbor matching principle: for each photoelectric analog quantity record, the bit error rate record with the closest timestamp is found. When the time difference between the two is less than 5% of the sampling period (5ms), they are considered data from the same sampling moment and merged into one record. If the time difference exceeds the threshold, the data is discarded and the process waits for the next sampling, avoiding parameter association errors caused by time misalignment.
[0020] Range conversion is the process of converting the raw integer value of a register into a standard physical quantity. Each parameter has a corresponding conversion formula and range coefficient. The conversion formula for input optical power is P_in = 10 × log10(RAW_p × k_p / 1000), where P_in represents the input optical power in dBm; RAW_p represents the raw reading of the input optical power register; and k_p is the power conversion coefficient in μW / LSB. In this example, k_p is 0.1, meaning each least significant bit corresponds to 0.1 μW of optical power. The conversion formula for bias current is I_bias = RAW_i × k_i, where I_bias represents the bias current in mA; RAW_i represents the raw reading of the bias current register; and k_i is the current conversion coefficient. In this example, k_i is 0.02 mA / LSB. The temperature conversion formula is T = RAW_t × k_t - 273.15, where T represents the module temperature in degrees Celsius; RAW_t represents the raw reading of the temperature register; and k_t is the temperature conversion coefficient, which is 0.00390625 in this example. The voltage conversion formula is V = RAW_v × k_v, where V represents the supply voltage in volts (V); RAW_v represents the raw reading of the voltage register; and k_v is the voltage conversion coefficient, which is 0.0001V / LSB in this example.
[0021] In the example, after range conversion, the input optical power P_in is -8.2dBm, the bias current I_bias is 52mA, the temperature T is 42 degrees Celsius, the voltage V is 3.3V, and the bit error rate BER_pre is 2×10^-6. After converting all five parameters to standard physical dimensions, they are combined in a fixed order of input optical power, bias current, temperature, voltage, and bit error rate to form a five-dimensional sampling record. Each record also includes a precise sampling timestamp with millisecond-level accuracy, ensuring the time accuracy of subsequent time-series analysis.
[0022] The dimensionally normalized five-dimensional sampling records are written into a circular buffer in the order of sampling time. When the buffer is filled with a preset number of records, they are read out as a whole in the form of data blocks and start and end timestamps are attached to generate a multi-dimensional performance time series dataset.
[0023] The core of this step is to use a circular buffer mechanism to batch buffer and package the continuously sampled five-dimensional records to form a time-series dataset with a fixed time window length. This ensures both the continuity and integrity of the data and facilitates batch processing by the subsequent link quality assessment model. The specific implementation method is as follows: A circular buffer is a fixed-length circular storage structure consisting of a contiguous block of memory and read / write pointers. The write pointer moves forward as new data is written, and automatically wraps back to the beginning when it reaches the end of the buffer, overwriting the oldest written data. The advantage of using a circular buffer is that it always retains the latest N sampled records, forming a sliding time window. This eliminates the need for frequent memory allocation and deallocation, resulting in high efficiency and ensuring the temporal continuity of data.
[0024] In the example, the preset capacity N of the circular buffer is set to 100 records, corresponding to a 10-second sampling time window. This window length is sufficient to cover the characteristic information of the slowly deteriorating process in the optical link, while avoiding increasing the computational burden of subsequent models due to excessive data volume. After each five-dimensional sampling record is generated, it is written to the circular buffer sequentially according to the sampling time. The write pointer moves forward one position after each record is written. When the buffer is filled with 100 records for the first time, a data block read operation is triggered. At this time, the buffer contains consecutive sampling records from the 1st to the 100th record, corresponding to 10 seconds of complete monitoring data.
[0025] Data block reading employs a batch read approach, retrieving all 100 records from the buffer at once to avoid the system overhead of reading records one by one. Each read data block is appended with a start timestamp and an end timestamp. The start timestamp is the sampling time of the first record in the buffer, and the end timestamp is the sampling time of the last record. These two timestamps together define the time interval corresponding to the data block. Subsequently, the buffer continues to operate in a sliding window manner. Each new record written overwrites the oldest record. After a cumulative total of M new records written, batch reading is triggered again, where M is the sliding step size. In this example, M is set to 10, meaning a new data block is output every 1 second. Each data block contains the latest 100 records, with 90 records overlapping between adjacent data blocks to ensure the continuity of the time window and the traceability of degradation trends.
[0026] The resulting multidimensional performance time-series dataset consists of a 100×5 two-dimensional matrix for each data block, where 100 represents the time step and 5 represents the performance parameter dimension. The start and end timestamps attached to each data block precisely identify its corresponding time interval. This dataset serves as input to the subsequent link quality assessment model, providing continuous, well-organized, and dimensionally consistent multidimensional time-series data to support the model's accurate analysis and judgment of link status.
[0027] S202, the multidimensional performance time series dataset is input into the link quality assessment model constructed based on graph attention network, the coupling features between each performance parameter are extracted, and a quality score and degradation feature description vector reflecting the current link health are generated by combining historical decay trends. Specifically, it can include: S2021, normalizing the five performance parameter sequences in the multidimensional performance time series dataset, and forming a five-dimensional feature vector from the five normalized values at the same time, arranging them into a feature matrix in chronological order to generate a standardized time series feature matrix; The core of this step is to unify the dimensions and standardize the scale of the original multidimensional performance data, eliminating the numerical magnitude differences between different physical quantities. This avoids the problem of large numerical parameters dominating gradient updates during subsequent network training and inference, ensuring that the contribution weights of each dimension feature to the model output are within a comparable range. The specific implementation method is as follows: The multidimensional performance time series dataset contains five performance parameter sequences: input optical power, bias current, temperature, voltage, and dispersion-compensated forward error correction code rate. The physical dimensions and numerical ranges of each parameter are significantly different. The typical range of input optical power is -20 dBmW to 0 dBmW, the typical range of bias current is 20 mA to 120 mA, the typical range of temperature is -5°C to 75°C, the typical range of voltage is 3.0 volts to 3.6 volts, and the typical range of forward error correction code rate is 1e-6 to 1e-2. The numerical range of the five parameters spans multiple orders of magnitude. If directly input into the subsequent network model, the parameters with larger values will dominate the feature calculation and weight update, causing the model to be unable to effectively learn the true coupling relationship between the parameters. Therefore, normalization processing must be performed first.
[0028] Normalization employs the min-max normalization method, expressed as x_norm = (x - x_min) / (x_max - x_min), where x represents the original sampled value of a parameter, x_min represents the minimum value of the parameter across the entire range, and x_max represents the maximum value of the parameter across the entire range. Both x_min and x_max are determined based on the factory specifications of the optical transceiver module and historical operational data statistics, ensuring coverage of the entire value range within both normal operating and abnormal degradation intervals. x_norm represents the normalized feature value, ranging from 0 to 1, where 0 corresponds to the parameter being at the lower limit of the range, 1 corresponds to the parameter being at the upper limit of the range, and 0.5 corresponds to the parameter being at the middle value of the range. After normalization, the numerical ranges of the five parameters are uniformly mapped to the same 0-1 interval, maintaining a consistent numerical scale for each dimension of the feature. This allows the subsequent model to fairly learn the impact of each parameter on link quality.
[0029] In the example, the sampling period is set to 1 second, and the number of consecutive sampling records stored in the circular buffer is 300, corresponding to a 5-minute historical data window. After the five parameter sequences are independently normalized, the five normalized values corresponding to the same sampling time are combined into a five-dimensional feature vector in a fixed order: input optical power, bias current, temperature, voltage, and forward error correction code rate. Each time point corresponds to a five-dimensional feature vector. The five-dimensional feature vectors from the 300 time points are stacked vertically in chronological order to form a feature matrix with 300 rows and 5 columns. The row dimension corresponds to the time dimension, and the column dimension corresponds to the parameter dimension. The resulting standardized temporal feature matrix retains the temporal trend information of each parameter while eliminating numerical bias caused by differences in units, and can be directly used as input data for the subsequent graph attention network.
[0030] S2022 uses the five-dimensional feature vector of each time step in the standardized temporal feature matrix as the node feature of the graph attention network, and defines the adjacency edge of the graph structure based on the physical relationship between each performance parameter. It calculates the mutual attention weight between nodes through a multi-head attention mechanism to generate a parameter-coupled attention coefficient matrix. The core of this step is to construct a graph structure representing the physical relationships between parameters. It then utilizes the multi-head attention mechanism of a graph attention network to automatically learn the dynamic coupling strength between performance parameters, quantifying the weight differences in the impact of different parameters on link quality. This overcomes the limitation of traditional fixed-weight methods, which cannot adapt to dynamically degraded scenarios. The specific implementation is as follows: Graph Attention Networks (GANs) are deep learning models based on graph structures. Their core feature is the ability to adaptively learn the association weights between nodes in a graph through an attention mechanism, eliminating the need for pre-setting influence coefficients between parameters. They can dynamically adjust the coupling strength between parameters based on real-time data, making them ideal for handling complex scenarios with multi-parameter nonlinear coupling in optical transmission systems. In the graph structure constructed in this step, five performance parameters correspond to five graph nodes. The initial feature vector of each node is the normalized temporal sequence of that parameter over the entire time window, and the node feature dimension is equal to the time window length, i.e., 300 dimensions.
[0031] The adjacency edges of the graph structure are initialized based on the physical relationships between performance parameters. There is a direct physical relationship between input optical power and bias current; the laser's output optical power is driven by the bias current, and the two are strongly coupled. There is also a physical relationship between temperature and bias current; temperature changes cause drift in the laser's threshold current and slope efficiency, leading to adjustments in the bias current. An indirect relationship exists between temperature and input optical power; temperature drift causes optical power fluctuations. A relationship exists between voltage and bias current; fluctuations in the supply voltage affect the output current of the drive circuit. The forward error correction bit rate is directly related to input optical power, dispersion, and nonlinear effects; changes in input optical power directly affect the bit error rate level. Based on these physical relationships, the initialized adjacency matrix is a 5x5 symmetric matrix. The adjacency values for two physically related parameters are set to 1, and the adjacency values for parameters without direct physical relationships are set to 0. This adjacency matrix defines the range of node pairs allowed for attention calculation in the graph attention network.
[0032] Multi-head attention is the core computational module of graph attention networks. By using multiple independent attention heads to compute attention weights between nodes in parallel, and then concatenating and fusing the results, it can capture the coupling relationships between parameters in different dimensions, thereby improving the richness and robustness of feature extraction. The attention coefficients of each attention head are calculated according to the following expression: e_ij=LeakyReLU(a^T[Wh_i||Wh_j]), where h_i and h_j represent the initial feature vectors of node i and node j, respectively, W represents a learnable linear transformation matrix used to map the original features to a higher-dimensional feature space, a represents the attention weight vector used to calculate the correlation score between the features of two nodes, || represents the vector concatenation operation, and LeakyReLU represents the linear unit activation function with leakage correction, used to introduce nonlinear transformation. e_ij represents the original attention score of node i to node j. Then, the scores of all neighboring nodes are normalized by the softmax function to obtain the final attention coefficient α_ij. The value of α_ij ranges from 0 to 1. The sum of the attention coefficients of all neighboring nodes is equal to 1. The larger the value of α_ij, the stronger the influence of node j on node i.
[0033] In the example, the multi-head attention mechanism is set up with four independent attention heads, each independently calculating a 5x5 attention coefficient matrix. The four matrices are then concatenated along the feature dimension and fused through a linear transformation to form the final parameter-coupled attention coefficient matrix. Each element in this matrix corresponds to the dynamic coupling weight between two performance parameters, reflecting in real-time the strength of the interaction between parameters under the current link state. For example, when the link experiences optical power drift degradation due to increased temperature, the attention coefficient of the temperature node to the input optical power node will significantly increase, and the model can automatically identify temperature as the dominant factor affecting link quality. The final parameter-coupled attention coefficient matrix fully characterizes the dynamic coupling relationship between the five performance parameters, providing a core basis for subsequent feature enhancement and quality score calculation.
[0034] S2023, the parameter coupling attention coefficient matrix is weighted and fused with the original feature matrix to obtain the enhanced features of each parameter after being affected by other parameters at each time step. Then, the enhanced features are weighted and pooled along the time dimension to generate the overall quality score of the current link. The core of this step is to use attention coefficients to weight and enhance the original features, so that the features of each parameter incorporate the influence information of other related parameters, fully representing the comprehensive effect of parameter coupling on link quality. Then, historical trend information is aggregated through weighted pooling in the time dimension to output a quantitative score reflecting the overall health level of the current link. The specific implementation method is as follows: The parameter coupling attention coefficient matrix represents the influence weights between various parameters. Weighted fusion of this matrix with the original feature matrix essentially involves aggregating the feature information of each node with the feature information of its neighboring nodes, distributing the aggregation ratio according to the attention weights. This results in enhanced features that incorporate the parameter coupling effect. The weighted fusion is calculated as follows: for each node i, the feature vectors of all neighboring nodes j are multiplied by their corresponding attention coefficients α_ij, summed, and then added to node i's own original feature vector to obtain node i's enhanced feature vector. After weighted fusion, the features of each parameter are no longer isolated single-parameter time series, but rather comprehensive features that include the influence of other related parameters, more accurately reflecting the link state under parameter coupling.
[0035] In the example, the enhanced features of the input optical power node not only include its own temporal variation information but also incorporate the influence information of related parameters such as bias current and temperature after being weighted by attention weights. When the temperature rises and causes the optical power to decrease, the enhanced features of the input optical power can simultaneously reflect the contribution of temperature changes, thus more accurately characterizing the degree of impact of this degradation mode on link quality. After feature enhancement is completed for each of the five parameter nodes, an enhanced feature matrix with the same dimensions as the original feature matrix is obtained. The number of rows is still 300 corresponding to the time dimension, and the number of columns is still 5 corresponding to the parameter dimension, but each element has incorporated the coupling effect between parameters.
[0036] The purpose of time-dimensional weighted pooling is to aggregate the enhanced features from 300 time points into a single quality score, assigning different weights to different time points—higher weights to recent time points and lower weights to older time points. This ensures the quality score better reflects the current link status while also taking into account historical trends. The time weights are calculated using an exponential decay function, expressed as w_t=exp(-(Tt) / τ), where t represents a sampling time, T represents the current latest time, τ represents the time decay constant, and w_t represents the weight value at time t. The value of τ determines the rate at which historical information is forgotten; a larger τ results in higher weights for historical information and a smoother quality score, while a smaller τ results in higher weights for recent information and a more sensitive quality score to changes. In this example, τ is set to 60 seconds, corresponding to a 1-minute time constant, which smooths out short-term random fluctuations while promptly responding to continuous degradation trends.
[0037] The weighted pooling calculation process is as follows: First, the five-dimensional enhanced features at each time step are subjected to one-dimensional global average pooling to obtain single-value features at each time step, resulting in 300 single-value features corresponding to 300 time steps. Then, these 300 single-value features are multiplied by their corresponding time weights w_t and summed. Finally, a fully connected layer and a sigmoid activation function are used to map the results to the 0-1 range, yielding the final overall link quality score. The quality score ranges from 0 to 1, where 1 represents a perfectly healthy link with all performance indicators at their optimal levels, and 0 represents a completely failed link. The closer the score is to 1, the higher the link health; the closer the score is to 0, the more severe the link degradation. In the example, when all link parameters are within the normal range, the quality score remains stable between 0.92 and 0.98; when slight degradation occurs, the score drops to between 0.75 and 0.9; and when severe degradation occurs, the score drops below 0.5.
[0038] S2024 concatenates the overall quality score with the quality score sequence of multiple historical periods, calculates the time gradient and second-order difference of the quality score, identifies degradation patterns such as continuous decline, abrupt jumps or periodic fluctuations, and finally generates a quality score and degradation feature description vector that reflects the current link health.
[0039] The core of this step is to identify the specific patterns and severity of link degradation based on the time-series variation characteristics of quality scores. This expands the single quality scalar into a feature vector containing rich degradation semantics, providing accurate criteria for subsequent optimization decisions regarding degradation type and severity. This avoids the optimization strategy mismatch problem caused by a single score failing to distinguish different degradation patterns. The specific implementation method is as follows: The overall quality score only reflects the overall health level of the link at the current moment and cannot reflect the dynamic changes in degradation. Different degradation modes correspond to completely different optimization strategies. Slow and continuous decline usually corresponds to device aging degradation, which requires gradual adjustment of compensation parameters. Abrupt changes usually correspond to sudden failure degradation, which requires rapid switching of modulation format. Periodic fluctuations usually correspond to environmental disturbance degradation, which requires targeted adjustment of temperature compensation parameters. Therefore, it is necessary to perform pattern recognition on the time series of quality scores to extract the type and degree features of degradation.
[0040] The historical quality score sequence refers to quality score data sampled at fixed time intervals of 5 minutes over the past 24 hours, totaling 288 historical score points. The latest quality score is concatenated with these 288 historical scores to form a quality score time series of length 289. The time gradient, or first-order difference, is calculated as ΔS_t = S_t - S_{t-1}, where S_t represents the quality score at time t, S_{t-1} represents the quality score at the previous time, and ΔS_t represents the time gradient of the quality score at time t, reflecting the rate of change. A negative ΔS_t indicates a decreasing score, and a larger absolute value indicates a faster rate of decrease. The second-order difference is calculated as Δ²S_t = ΔS_t - ΔS_{t-1}, reflecting the rate of change of the quality score, i.e., acceleration. A negative second-order difference indicates that the rate of decrease is accelerating, and the degradation is in an accelerated phase.
[0041] The identification rule for a continuous decline pattern is that the time gradient of 12 consecutive sampling points (i.e., within 1 hour) is negative, and the absolute value of the average gradient is greater than 0.005. This pattern is typically associated with slow degradation processes such as laser aging and increased fiber loss. The identification rule for abrupt jump patterns is that the absolute value of the time gradient of a single sampling point is greater than 0.1, and the difference between the scores before and after the jump exceeds 0.1. This pattern is typically associated with rapid degradation events such as fiber bending, loose connectors, and sudden interference. The identification of periodic fluctuation patterns is achieved through Fast Fourier Transform (FFT). Spectral analysis of the quality score sequence is performed; if a significant dominant frequency component exists with an amplitude exceeding 0.05, it is identified as a periodic fluctuation degradation pattern. This pattern typically corresponds to periodic disturbances such as periodic changes in ambient temperature and the start / stop of air conditioning.
[0042] The degradation feature description vector is a multi-dimensional vector containing the following features: the first dimension is the current link quality score, ranging from 0 to 1; the second dimension is the degradation type label, where 0 represents no degradation, 1 represents continuous degradation, 2 represents abrupt degradation, and 3 represents periodic fluctuation degradation; the third dimension is the degradation severity quantification value, ranging from 0 to 1, calculated based on the magnitude and rate of change of the quality score, with a larger value indicating more severe degradation; the fourth dimension is the average gradient of the quality score, reflecting the recent rate of change; the fifth dimension is the second difference of the quality score, reflecting the acceleration of change; and the sixth dimension is the dominant frequency period of the fluctuation, which is only valid in periodic fluctuation mode and is 0 in other modes.
[0043] In the example, when the link experiences continuous degradation due to slow laser aging, the quality score gradually decreases from 0.95 to 0.82, with an average gradient of -0.008 every 5 minutes. The second-order difference is a negative value close to 0, the degradation type label is 1, and the severity quantification value is 0.35, corresponding to moderate continuous degradation. The final generated quality score and degradation feature description vector contain both an overall quantitative assessment of the link's health and semantic information and severity quantification information of the degradation pattern. This provides comprehensive and accurate input for subsequent intelligent optimization decisions, ensuring that optimization actions accurately match the current degradation type and severity.
[0044] S203, based on the degradation type and severity in the degradation feature description vector, dynamically decide the modulation format switching level and the digital signal processing compensation parameter adjustment amount through the deep deterministic strategy gradient algorithm, and generate an optimized action set containing modulation format instructions and compensation parameter instructions; Specifically, it can include: S2031, extracting the degradation type label and severity quantification value from the degradation feature description vector, combining the two into the current environment state vector, and stacking the state vector with the historical state vector of the most recent N steps to form a state tensor, generating the input state tensor of the deep deterministic policy gradient algorithm; The degradation feature description vector is a multi-dimensional feature vector output by the link quality assessment model. It includes information such as the current link degradation type label, severity quantification value, quality score temporal gradient, second-order difference value, and the contribution weights of each performance parameter. The degradation type label is represented using one-hot encoding, corresponding to five typical degradation modes: continuous decay, abrupt change, periodic fluctuation, noise-dominated, and dispersion-dominated. Each mode corresponds to one dimension of the vector, with a value of 1 indicating the dominant degradation mode and 0 indicating a non-dominant mode. The severity quantification value is a continuous scalar ranging from 0 to 1, where 0 indicates a completely healthy link with no degradation, and 1 indicates that the link degradation has reached the critical alarm threshold. This value is calculated by normalizing the deviation of the quality score from the healthy baseline.
[0045] After extracting the degradation type label and severity quantization value from the degradation feature description vector, the two are concatenated to form the current environment state vector. The current environment state vector is 6-dimensional, where the first 5 dimensions correspond to the one-hot encoding of the degradation type label, and the 6th dimension corresponds to the severity quantization value. To enable the reinforcement learning algorithm to perceive the temporal evolution trend of link degradation, rather than relying solely on the instantaneous state at the current moment, the current environment state vector and the historical state vectors of the most recent N steps need to be stacked along the time dimension to form a two-dimensional state tensor. Here, N is set to 10, indicating that the state information of the most recent 10 sampling periods is used to characterize the dynamic evolution process of the current link. The state tensor is 10 rows and 6 columns, with each row corresponding to the environment state vector at a sampling moment. The smaller the row index, the closer the time is to the current moment, and the larger the row index, the further back in time it is.
[0046] The Deep Deterministic Policy Gradient Algorithm (DPRQ) is a continuous action space decision-making algorithm based on deep reinforcement learning. Its input is a state tensor representing the environment state, and its output is an optimized action in continuous value form. This algorithm learns a deterministic policy through an Actor network, directly outputting the continuous values of the actions. Simultaneously, a Critic network evaluates the value of state-action pairs to guide the parameter updates of the Actor network. The state tensor, as the algorithm's input, needs to undergo standardization preprocessing to map the values of each dimension to the interval between 0 and 1, avoiding the adverse effects of differences in the scale of different dimensions on network training. In the example, the degradation type detected at the current time step is dispersion-dominated, with a one-hot encoding of [0,0,0,0,1] and a severity quantization value of 0.62. Therefore, the current environment state vector is [0,0,0,0,1,0.62]. Stacking this vector with the state vectors of the previous 9 historical moments yields an input state tensor of dimension 10×6. This tensor fully reflects the evolution of link degradation types and the changing trend of severity within the last 10 sampling periods, providing sufficient state information support for subsequent policy network decisions.
[0047] S2032, the input state tensor is fed into the Actor online network of the deep deterministic policy gradient algorithm. Through the convolutional layers and fully connected layers in the network, a continuous two-dimensional action vector is output. The first dimension of the vector represents the continuous value of the modulation format switching level, and the second dimension represents the continuous value of the compensation parameter adjustment amount, thus generating the original continuous action vector. The Actor online network is the core network responsible for policy output in the deep deterministic policy gradient algorithm. Its function is to directly output deterministic continuous action values based on the input environment state tensor. The network employs a hybrid architecture combining convolutional layers and fully connected layers. The convolutional layers extract local correlation features of the state tensor in the time dimension, capturing short-term change patterns of link degradation. The fully connected layers perform nonlinear mapping on the extracted high-level features, ultimately outputting continuous action values. The network consists of two one-dimensional convolutional layers and three fully connected layers. The first convolutional layer has a kernel size of 3 and 32 output channels, used to extract fine-grained time-series features. The second convolutional layer has a kernel size of 3 and 64 output channels, used to extract more abstract temporal evolution patterns. After the convolutional layers, a global average pooling layer is connected to compress the two-dimensional feature map into a one-dimensional feature vector, followed by three fully connected layers with 128, 64, and 2 neurons respectively, ultimately outputting a continuous action vector of dimension 2. The hidden layers use a linear rectified function as the activation function to introduce nonlinear transformation capability and improve the feature representation capability of the network. The output layer uses a hyperbolic tangent activation function to constrain the output value to a continuous interval from -1 to 1.
[0048] The first dimension of the two-dimensional action vector represents a continuous value for the modulation format switching level. This value is a continuous scalar ranging from -1 to 1, where -1 corresponds to the lowest modulation format level and 1 corresponds to the highest modulation format level. The continuous values in between correspond to the transitions between different levels. A higher modulation format level results in higher spectral efficiency, but also places higher demands on the link's signal-to-noise ratio. When link degradation is severe, the modulation format level needs to be reduced to ensure transmission reliability. The second dimension of the two-dimensional action vector represents a continuous value for the compensation parameter adjustment, also a continuous scalar ranging from -1 to 1. Negative values indicate a decrease in compensation intensity, while positive values indicate an increase in compensation intensity. This value is used to guide the joint adjustment of dispersion compensation parameters and nonlinear compensation parameters.
[0049] After the input state tensor is fed into the Actor online network, it first passes through the first convolutional layer. The convolutional kernel slides along the time dimension, performing convolution operations on the state vectors at adjacent time points to extract local change features of the degradation trend, outputting a temporal feature map with 32 feature channels. Subsequently, the second convolutional layer further abstracts the features from the first layer, outputting a high-level feature map with 64 channels. A global average pooling layer averages the temporal features of each channel, resulting in a one-dimensional feature vector of length 64. This feature vector is then passed through three fully connected layers for nonlinear mapping. The first fully connected layer maps the 64-dimensional feature to 128-dimensional, the second layer maps the 128-dimensional feature to 64-dimensional, and the third layer maps the 64-dimensional feature to 2-dimensional. The output is then constrained to between -1 and 1 by a hyperbolic tangent activation function. In the example, after the input state tensor is forward-propagated through the network, the original continuous action vector output is [-0.35, 0.48]. The continuous value of the modulation format switching level is -0.35, which means that the strategy tends to adjust the current modulation format down by about 35% to cope with link degradation. The continuous value of the compensation parameter adjustment is 0.48, which means that the strategy tends to adjust the compensation parameter in the direction of enhancement by about 48% of the adjustable range, so as to offset the performance degradation caused by degradation by enhancing the compensation capability.
[0050] S2033, the continuous values of modulation format switching level in the original continuous action vector are discretized into the actual usable modulation format level number through nearest neighbor mapping, and the continuous values of compensation parameter adjustment amount are constrained within the adjustable range through the amplitude limiting function and quantized into an integer multiple of the step resolution of the digital signal processing unit to generate a discretized action vector. Since the modulation formats supported by the optical transceiver module are limited to a finite number of discrete levels rather than continuously adjustable ones, the continuous modulation format switching level values output by the Actor network need to be discretized into practically usable modulation format level numbers. Nearest neighbor mapping is a simple and efficient discretization method; its core idea is to map continuous values to the nearest discrete level. The system supports five modulation formats, with levels numbered from 0 to 4 corresponding to Quadrature Phase Shift Keying (QPSK), 8th-order Quadrature Amplitude Modulation (QAPM), 16th-order QAPM, 32nd-order QAPM, and 64th-order QAPM, respectively. The spectral efficiency increases progressively with each level, and the requirements for link signal-to-noise ratio also increase progressively. The continuous value interval from -1 to 1 is uniformly divided into five sub-intervals corresponding to the five discrete levels. The center values of each sub-interval are -0.8, -0.4, 0, 0.4, and 0.8, corresponding to levels 0 to 4, respectively. The distance between the continuous value and each center value is calculated, and the level number corresponding to the center value with the smallest distance is taken as the discretization result.
[0051] The continuous values of the compensation parameter adjustment also need to be constrained and quantized to adapt to the actual adjustable range and step resolution of the digital signal processing unit. The function of the limiting function is to constrain the continuous values within the physically adjustable range, preventing the output from exceeding the upper and lower limits supported by the hardware. The expression of the limiting function is y=min(max(x,x_min),x_max), where x is the continuous input value, x_min is the lower limit of the adjustable range, x_max is the upper limit of the adjustable range, and y is the output value after limiting. The dispersion compensation parameter and the nonlinear compensation parameter each have independent adjustable ranges and step resolutions. The adjustable range of the dispersion compensation parameter is 0 to 1000 picoseconds per nanometer, and the step resolution is 10 picoseconds per nanometer; the adjustable range of the nonlinear compensation parameter is 0 to 100, and the step resolution is 1. After the continuous values of the compensation parameter adjustment are limited, they need to be quantized according to the step resolution, that is, the integer multiple of the step closest to the continuous value is taken as the final parameter adjustment value, ensuring that the generated parameter value can be accurately executed by the digital signal processing unit.
[0052] In the example, the modulation format switching level in the original continuous action vector has a continuous value of -0.35. Calculating its distance from the center values of each level, the distance to the center value of level 1 (-0.4) is 0.05, and the distance to the center value of level 2 (0) is 0.35. Therefore, the nearest neighbor mapping result is level 1, corresponding to an 8th-order orthogonal amplitude modulation format. The continuous value of the compensation parameter adjustment is 0.48, which remains 0.48 after the amplitude limiting function constraint. Mapping this to the adjustable range of the dispersion compensation parameter, the corresponding adjustment is 480 picoseconds per nanometer. Quantized with a step resolution of 10 picoseconds per nanometer, it becomes 480 picoseconds per nanometer, exactly an integer multiple of the step value. The adjustment for the nonlinear compensation parameter is 48, with a step resolution of 1, and remains 48 after quantization. In the final generated discretized action vector, the modulation format level number is 1, the dispersion compensation parameter adjustment is 480 picoseconds per nanometer, and the nonlinear compensation parameter adjustment is 48. All parameters meet the actual hardware constraints and can be directly used for subsequent instruction generation.
[0053] S2034 maps the modulation format level number in the discretized action vector to the corresponding modulation format switching protocol instruction code, maps the compensation parameter adjustment amount to the setting instructions for dispersion compensation parameter value and nonlinear compensation parameter value, and finally generates an optimized action set containing modulation format instruction and compensation parameter instruction.
[0054] The mapping from modulation format level numbers to protocol command codes is accomplished through a predefined mapping table. Each modulation format level number corresponds to a unique protocol command code, which is a standard operation code that the digital signal processing unit can recognize and execute. In the mapping table, the command code corresponding to level 0 is 0x01, indicating a switch to quadrature phase shift keying (QPSK); the command code corresponding to level 1 is 0x02, indicating a switch to 8th-order quadrature amplitude modulation (QAM); the command code corresponding to level 2 is 0x03, indicating a switch to 16th-order QAM; the command code corresponding to level 3 is 0x04, indicating a switch to 32nd-order QAM; and the command code corresponding to level 4 is 0x05, indicating a switch to 64th-order QAM. The command codes use 8-bit unsigned integer encoding for easy transmission and parsing through the management interface. In addition to the command code itself, the modulation format command also includes auxiliary parameters such as the symbol rate, roll-off factor, and frame structure type of the target format, ensuring that the switched modulation format remains compatible with the peer optical transceiver module.
[0055] The mapping from compensation parameter adjustment to specific parameter values uses an incremental update method. That is, based on the current compensation parameter baseline value, the discretized compensation parameter adjustment is added to obtain the new compensation parameter setting value. The current baseline value of the dispersion compensation parameter is provided in real time by the dispersion estimation module in the digital signal processing unit, while the current baseline value of the nonlinear compensation parameter is determined by the current configuration value of the nonlinear compensation module. The instruction for setting the dispersion compensation parameter value includes three fields: parameter address, data length, and parameter value. The parameter address corresponds to the starting address of the dispersion compensation filter coefficient register, the data length is 4 bytes, and the parameter value is the quantized dispersion compensation amount. The instruction for setting the nonlinear compensation parameter value has a similar structure: the parameter address corresponds to the starting address of the nonlinear compensation coefficient register, and the parameter value is the quantized nonlinear compensation intensity value.
[0056] The optimized action set is a structured set of instructions, comprising three parts: modulation format switching instructions, dispersion compensation parameter update instructions, and nonlinear compensation parameter update instructions. These instructions are arranged according to execution priority to ensure that the digital signal processing unit (DSP) executes them in the correct order. The optimized action set also includes metadata such as action number, generation timestamp, and expected effect identifier for subsequent effect verification and experience playback. In the example, the modulation format level number of the discretized action vector is 1, and the mapped modulation format instruction code is 0x02, indicating a switch to an 8th-order quadrature amplitude modulation format with a symbol rate of 32 gigabits and a roll-off factor of 0.1. The current baseline value of the dispersion compensation parameter is 1200 picoseconds per nanometer (pN), and after adding an adjustment of 480 pC / pN, the new setting value is 1680 pC / pN. The current baseline value of the nonlinear compensation parameter is 30, and after adding an adjustment of 48, the new setting value is 78. The final generated optimized action set contains the above three instructions, with a total instruction length of 32 bytes, and can be sent to the DSP for execution in one go through the management data input / output interface. After the optimized action set is generated, the system records complete information about the action, including the state before the action is executed, the content of the action, and the performance feedback after the execution. This information is used for the experience playback and policy update of the deep deterministic policy gradient algorithm to continuously optimize the decision quality.
[0057] S204, the optimized action set is sent to the digital signal processing unit of the optical transceiver module to perform online switching of modulation format and dynamic updating of dispersion compensation and nonlinear compensation parameters, so as to realize real-time performance monitoring and intelligent optimization of the optical transmission system.
[0058] Specifically, the modulation format switching instructions in the optimized action set can be encoded into mode switching register write values that can be recognized by the digital signal processing unit, and the compensation parameter instructions can be encoded into filter coefficient write values at the corresponding address of the digital signal processing unit, thereby generating a register write instruction sequence that can be directly executed by the digital signal processing unit. The core of this step is to transform the abstract optimization actions output by the upper-level algorithm into register-level instructions that can be directly executed by the underlying hardware. This completes the semantic mapping from policy decision-making to hardware operation, ensuring that the digital signal processing unit can accurately identify and execute modulation format switching and compensation parameter updates. The specific implementation method is as follows: The optimization action set is a structured set of instructions output by the deep deterministic policy gradient algorithm. It contains two core categories: modulation format instructions and compensation parameter instructions. The modulation format instructions specify the target modulation format level, while the compensation parameter instructions specify the adjustment amounts for dispersion compensation and nonlinear compensation parameters. These two types of instructions exist at the algorithm level in the form of continuous numerical values or discrete numbers, and cannot be directly recognized and executed by the digital signal processing unit. They must be encoded and converted into corresponding write values in hardware registers before forming a sequence of instructions that can be directly issued.
[0059] The modulation format switching register is a dedicated control register within the digital signal processing unit (DSP) used to control modulation and demodulation modes. Its written values correspond to different modulation format operating modes. In this example, the address offset of the modulation format switching register is 0x08, and the register width is 16 bits. The high 8 bits are used to identify the modulation format type, and the low 8 bits are used to identify the signal rate level. During encoding, the modulation format level number in the discretized action vector is converted into the corresponding register write value according to a preset mapping table. For example, level 0 corresponds to binary polarization multiplexed quadrature phase shift keying (QPSK) format, with a write value of 0x0101; level 1 corresponds to polarization multiplexed 8th-order quadrature amplitude modulation (QAM) format, with a write value of 0x0201; level 2 corresponds to polarization multiplexed 16th-order QAM format, with a write value of 0x0301; and level 3 corresponds to polarization multiplexed 32nd-order QAM format, with a write value of 0x0401. The mapping relationship is pre-stored in a lookup table in the DSP control unit. During encoding, the corresponding write value is directly retrieved from the table, ensuring conversion efficiency and accuracy.
[0060] The encoding of compensation parameter instructions involves two types of parameters: dispersion compensation and nonlinear compensation, each corresponding to different filter coefficient register groups within the digital signal processing unit. The dispersion compensation parameter corresponds to the tap coefficient register group of the time-domain equalizer filter. In this example, the register group contains 256 32-bit floating-point coefficient registers, with addresses ranging from 0x100 to 0x4FF. Each register stores the coefficient value of one tap. During encoding, the dispersion compensation parameter adjustment is converted into the increment value of the corresponding tap coefficient and written sequentially to the register at the corresponding address according to the tap number, forming a continuous coefficient write sequence. The nonlinear compensation parameter corresponds to the nonlinear order and step size registers of the digital backpropagation algorithm. In this example, the nonlinear order register address is 0x500, and the step size register address is 0x504. During encoding, the nonlinear compensation parameter adjustment is converted into the corresponding integer value of the order and the floating-point value of the step size, and written to the corresponding registers.
[0061] The register write instruction sequence is organized using a unified instruction frame format. Each instruction contains four fields: opcode, register address, write data, and checksum. The opcode occupies 8 bits and identifies the type of register write operation; the register address occupies 16 bits and specifies the offset address of the target register; the write data occupies 32 bits and carries the value to be written; and the checksum occupies 8 bits and verifies bit errors during instruction transmission. The entire instruction sequence is arranged in the order of writing the modulation format register, then the dispersion compensation coefficient register, and finally the nonlinear compensation parameter register, ensuring that the digital signal processing unit can parse and execute them in sequence. In the example, a complete optimization action corresponds to a register write instruction sequence containing 259 single register write instructions, with a total data volume of approximately 1280 bytes, which can be transmitted in a single management transaction.
[0062] The register write instruction sequence is sent to the control register group of the digital signal processing unit through the management data input / output interface of the optical transceiver module or the integrated circuit bus, triggering the digital signal processing unit to enter the parameter update preparation state and generate a register write completion status flag. The core of this step is to reliably transmit the register instruction sequence to the digital signal processing unit inside the optical transceiver module through a standardized management interface, completing instruction issuance and status synchronization, and ensuring that parameter update actions can be correctly triggered by the hardware. The specific implementation method is as follows: The management data input / output interface is a common management interface standard in the optical module field. It adopts a two-wire serial bus architecture, with one wire transmitting the clock signal and the other transmitting the data signal. It supports a master-slave communication architecture, with the system control unit on the master side and the management chip and digital signal processing unit inside the optical transceiver module on the slave side. The standard communication rate of this interface is 400kbps, supporting a high-speed mode of up to 1Mbps, which can meet the bandwidth requirements for register configuration and status reading. The integrated circuit bus is another widely used two-wire serial management interface. Its communication protocol is similar to the management data input / output interface, also adopting a master-slave architecture, suitable for control communication between board-level chips. In this system, both interfaces can be used as management channels, and the appropriate interface can be selected based on the actual configuration of the hardware platform, while maintaining communication protocol compatibility.
[0063] The instruction delivery process employs a combination of batch writing and packet-by-packet acknowledgment to ensure reliable instruction transmission. The system control unit, acting as the master, first initiates a write operation to the slave address of the optical transceiver module, splitting the register write instruction sequence into multiple data packets of 16 bytes each, which are then sent sequentially via the management bus. After each data packet is sent, the master waits for an acknowledgment signal from the slave to confirm successful reception before sending the next data packet. If no acknowledgment is received three consecutive times, the transmission is considered a failure, triggering a retransmission mechanism. A maximum of five retransmissions are allowed; exceeding this limit results in a communication fault report. Throughout the entire delivery process, the control register group of the digital signal processing unit remains writable. Received instruction data is directly written to the corresponding register's buffer and does not take effect immediately, preventing partial writes from causing operational anomalies.
[0064] The parameter update preparation state is a transitional state for the digital signal processing unit (DSP) after receiving a complete instruction sequence but before executing parameter updates. In this state, the DSP maintains its current operating mode while simultaneously preloading and verifying the new parameters. After the last register write instruction is transmitted, the host writes a parameter update trigger word to the DSP's control register (in this example, 0xA5A5). Upon detecting this trigger word, the DSP first verifies all written register values. If the verification passes, it enters the parameter update preparation state, sets the ready flag in the status register to 1, and generates a register write completion flag. If the verification fails, the error flag is set to 1, and all written buffered data is discarded, awaiting a re-send from the host.
[0065] The register write completion status flag is stored in the status register of the digital signal processing unit, with an address offset of 0x00. Bit 0 is the ready flag, and bit 1 is the checksum error flag. The host can read this status register at any time through the management interface to query the readiness status of parameter updates. In the example, the total time from issuing the instruction to generating the completion status flag is approximately 25 milliseconds, of which data transmission takes approximately 12 milliseconds, and parameter preloading and verification takes approximately 13 milliseconds. The overall response speed meets the latency requirements for real-time optimization.
[0066] After the digital signal processing unit detects the register write completion status flag, it synchronously performs modulation format switching and compensation parameter update actions at the current frame boundary to avoid service interruption caused by switching in the middle of the data packet and generates a parameter effective confirmation signal. The core of this step is to synchronously switch the modulation format and compensation parameters within the time window at the frame boundary, ensuring the continuity and integrity of service data transmission and avoiding bit errors or data loss during the switching process. The specific implementation method is as follows: Frame boundaries are the start or end points of data frames in an optical transmission system, and are the only switching points for modulation formats and parameters. Optical transmission systems organize data transmission according to a fixed frame structure. Each frame consists of three parts: a frame header, a payload area, and a frame trailer. The frame header carries frame synchronization signals and a frame type identifier. The receiving end achieves frame synchronization by identifying the frame header. Switching the modulation format or compensation parameters in the middle of a frame will cause demodulation and decoding errors in the current frame, resulting in service interruption. Therefore, switching must be performed at frame boundaries to ensure that each frame uses the same modulation format and the same set of compensation parameters.
[0067] The frame synchronization module inside the digital signal processing unit (DSP) continuously monitors the frame header position of the received data, maintaining a precise frame timing signal. When a register write completion flag is detected, the DSP does not immediately perform a switch, but waits for the next frame boundary. During this waiting period, the DSP continues to process data using the current modulation format and compensation parameters, while preloading new parameters into the spare register set in preparation for the switch. When the frame boundary arrives, the DSP completes the switch between the working register set and the spare register set within one clock cycle. The modulation format, dispersion compensation coefficients, and nonlinear compensation parameters are updated simultaneously. The switching process takes less than 10 nanoseconds and does not cause any delay in frame data processing.
[0068] Modulation format switching and compensation parameter updates employ a synchronous execution mechanism, taking effect simultaneously at the same frame boundary to avoid parameter mismatch issues caused by step-by-step switching. In the example, the current frame uses a polarization-multiplexed 16th-order orthogonal amplitude modulation format with corresponding dispersion compensation coefficients; the next frame switches to a polarization-multiplexed 8th-order orthogonal amplitude modulation format with updated dispersion compensation coefficients. The switching action is completed within the guard interval between the two frames. The receiver automatically uses the new modulation format parameters for demodulation when parsing the next frame header. The entire process is transparent to the service and does not result in packet loss or bit errors. To ensure synchronous switching at both ends, the system employs an in-band negotiation mechanism. Several frames before the switch, the transmitter carries switching warning information in the frame header overhead field to notify the receiver to prepare for the switch. Upon receiving the warning, the receiver synchronously waits for the frame boundary to execute the switch, ensuring parameter consistency at both ends.
[0069] The parameter activation confirmation signal is a status feedback signal returned by the digital signal processing unit after the handover is completed. It is used to notify the control unit that the parameters have been successfully activated. After the handover is complete, the digital signal processing unit sets the parameter activation flag in the status register to 1, and simultaneously writes the currently active modulation format number and parameter version number for the host to read and confirm. In the example, the parameter activation confirmation signal is generated in the third clock cycle after the frame boundary, and the host can read the status register through the management interface to confirm the handover result. If an abnormality occurs during the handover process, the digital signal processing unit will automatically revert to the parameter state before the handover and set the error flag to 1, triggering the fault handling procedure to ensure that the system does not remain in an abnormal working state for an extended period due to handover failure.
[0070] The forward error correction bit rate is monitored over several consecutive sampling periods after the parameters take effect. If the bit error rate continues to decrease and the quality score increases, the optimization is confirmed to be effective. The confirmation result and the current optimization action are fed back to the Critic network of the deep deterministic policy gradient algorithm for experience playback and network parameter update, ultimately realizing real-time performance monitoring and intelligent optimization of the optical transmission system.
[0071] The core of this step is to verify the actual effect of the optimization action and feed the effect back to the reinforcement learning algorithm for model iteration, forming a closed-loop optimization mechanism that allows the algorithm to continuously improve decision accuracy during continuous operation. The specific implementation method is as follows: After the parameters take effect, the system does not immediately determine that the optimization is successful. Instead, it initiates an effect verification process, continuously monitoring the changing trends of key performance indicators over multiple consecutive sampling periods to comprehensively evaluate the actual effect of the optimization actions. The sampling period is consistent with the initial data acquisition period, which is 100 milliseconds in this example. The effect verification window is set to 20 consecutive sampling periods, with a total duration of 2 seconds. This duration is sufficient for the system to reach a new steady state without being too long, which would cause optimization response lag. The core indicators monitored include the dispersion compensation forward error correction bit rate and the link quality score. The bit error rate is the most direct indicator reflecting transmission quality, while the quality score is an overall health indicator obtained by integrating multiple dimensions of parameters. The combination of the two can comprehensively evaluate the optimization effect.
[0072] The effectiveness assessment employs a combination of trend analysis and threshold judgment to avoid misjudgments caused by fluctuations in single sampling. The system calculates the moving average and slope of the bit error rate (BER) within the verification window. Optimization is considered effective when the BER shows a continuous downward trend, and the final steady-state value is lower than the initial value before optimization, while the quality score shows a continuous upward trend, and the final score is higher than the initial score before optimization. In the example, the initial BER before optimization was 1.2 × 10⁻³, and the quality score was 72. After optimization, at the end of the verification window, the steady-state BER dropped to 3.5 × 10⁻⁵, a decrease of over 97%, and the quality score rose to 94, an increase of over 30%, meeting the criteria for effective optimization. If the BER does not decrease or even increases within the verification window, or the quality score decreases, optimization is considered ineffective. The system automatically reverts to the parameter state before optimization and records the experience of this failure.
[0073] Experience replay is the core training mechanism of the deep deterministic policy gradient algorithm. By storing historical experience in an experience replay buffer, small batches of experience are randomly sampled during training to update network parameters, breaking the temporal correlation between samples and improving training stability. After each optimization action is executed, the system combines the current state tensor, the executed optimization action, the obtained reward value, and the next state tensor into an experience sample and stores it in the experience replay buffer. The reward value is calculated based on the optimization effect. In the example, the reward function is defined as the change in quality score multiplied by the weight coefficient, and then subtracted from the cost penalty of action switching. The reward value is positive when the optimization is effective and negative when the optimization is ineffective.
[0074] The Critic network is the value evaluation network in deep deterministic policy gradient algorithms. It is responsible for evaluating the expected reward of performing an action in the current state, and its output value is called the Q-value. Network parameter updates employ a temporal difference learning method, calculating the target Q-value using the immediate reward from experience samples and the Q-value of the next state. The network parameters are then updated by minimizing the mean squared error between the current Q-value and the target Q-value. In the example, the experience replay buffer has a capacity of 10,000 entries. Each training iteration randomly samples 64 experiences from the buffer to form a mini-batch. The learning rate is set to 0.0001, the discount factor to 0.99, and the soft update coefficient to 0.001. After each optimization, a network parameter update is performed. As the running time increases, the accumulated experience becomes richer, and the value evaluation of the Critic network becomes more accurate. This guides the Actor network to output better optimization strategies, forming a continuously evolving closed-loop optimization system, ultimately achieving real-time monitoring and intelligent adaptive optimization of the optical transmission system performance.
[0075] Another embodiment of the present invention provides a real-time performance monitoring and intelligent optimization system for an optical transmission system, see [link to relevant documentation]. Figure 5 The system may include: The acquisition module 501 is used to acquire the input optical power, bias current, temperature, voltage and dispersion compensation forward error correction code rate of the optical transceiver module in the optical transmission system in parallel at a fixed sampling period, and generate a multi-dimensional performance time series dataset. The extraction module 502 is used to input the multidimensional performance time series dataset into the link quality assessment model constructed based on graph attention network, extract the coupling features between various performance parameters, and generate a quality score and degradation feature description vector reflecting the current link health by combining historical decay trends. The decision module 503 is used to dynamically decide the modulation format switching level and the digital signal processing compensation parameter adjustment amount based on the degradation type and severity in the degradation feature description vector through a deep deterministic strategy gradient algorithm, and generate an optimized action set containing modulation format instructions and compensation parameter instructions. The execution module 504 is used to send the optimized action set to the digital signal processing unit of the optical transceiver module to perform online switching of modulation format and dynamic updating of dispersion compensation and nonlinear compensation parameters, so as to realize real-time performance monitoring and intelligent optimization of the optical transmission system.
[0076] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.
[0077] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.
[0078] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.
[0079] The above description is merely an exemplary embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention specification and drawings under the technical concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.
Claims
1. A method for real-time performance monitoring and intelligent optimization of an optical transmission system, characterized in that, The method includes: The input optical power, bias current, temperature, voltage, and dispersion compensation forward error correction code rate of the optical transceiver module in the optical transmission system are collected in parallel with a fixed sampling period to generate a multi-dimensional performance time series dataset. The multidimensional performance time series dataset is input into the link quality assessment model built on graph attention network, the coupling features between each performance parameter are extracted, and the quality score and degradation feature description vector reflecting the current link health are generated by combining the historical decay trend. Based on the degradation type and severity in the degradation feature description vector, the modulation format switching level and the digital signal processing compensation parameter adjustment amount are dynamically decided by the deep deterministic strategy gradient algorithm, generating an optimized action set containing modulation format instructions and compensation parameter instructions; The optimized action set is sent to the digital signal processing unit of the optical transceiver module to perform online switching of modulation format and dynamic updating of dispersion compensation and nonlinear compensation parameters, thereby realizing real-time performance monitoring and intelligent optimization of the optical transmission system.
2. The method of claim 1, wherein, The method involves parallel acquisition of input optical power, bias current, temperature, voltage, and dispersion-compensated forward error correction code rate of the optical transceiver module in the optical transmission system at a fixed sampling period to generate a multi-dimensional performance time-series dataset, including: Configure a sampling clock in the digital diagnostic monitoring interface of the optical transceiver module, set a fixed sampling period, and simultaneously send read requests for four analog quantities—input optical power, bias current, temperature, and voltage—to the module's internal register to generate a set of raw photoelectric analog readings. In the digital signal processing unit of the optical transceiver module, the bit error rate count before error correction is read from the forward error correction decoder with the same sampling period, and the bit error rate count is divided by the total number of error correction codewords to obtain the dispersion-compensated forward error correction bit rate, generating a set of original bit error rate readings. Align the four parameter values in the original set of photoelectric analog quantity readings with the original set of bit error rate readings according to the sampling timestamp, and perform range conversion and unit conversion on each parameter value to convert it into standard physical dimensions, generating a dimensionally normalized five-dimensional sampling record; The dimensionally normalized five-dimensional sampling records are written into a circular buffer in the order of sampling time. When the buffer is filled with a preset number of records, they are read out as a whole in the form of data blocks and start and end timestamps are attached to generate a multi-dimensional performance time series dataset.
3. The method of claim 2, wherein, The process involves inputting the multidimensional performance time-series dataset into a link quality assessment model constructed based on a graph attention network, extracting the coupling features between various performance parameters, and generating a quality score and degradation feature description vector reflecting the current link health by combining historical decay trends. This includes: The five performance parameter sequences in the multidimensional performance time series dataset are normalized respectively, and the five normalized values at the same time are combined into a five-dimensional feature vector, which is then arranged into a feature matrix in chronological order to generate a standardized time series feature matrix. The five-dimensional feature vector of each time step in the standardized temporal feature matrix is used as the node feature of the graph attention network, and the adjacency edge of the graph structure is defined by the physical relationship between each performance parameter. The mutual attention weight between nodes is calculated through the multi-head attention mechanism to generate the parameter coupling attention coefficient matrix. The parameter coupling attention coefficient matrix is weighted and fused with the original feature matrix to obtain the enhanced features of each parameter after being affected by other parameters at each time step. Then, the enhanced features are weighted and pooled along the time dimension to generate the overall quality score of the current link. The overall quality score is concatenated with the quality score sequences of multiple historical periods. The temporal gradient and second-order difference of the quality score are calculated to identify degradation patterns such as continuous decline, abrupt jumps, or periodic fluctuations. Finally, a quality score and degradation feature description vector reflecting the current link health are generated.
4. The method of claim 3, wherein, Based on the degradation type and severity in the degradation feature description vector, the deep deterministic strategy gradient algorithm dynamically determines the modulation format switching level and the digital signal processing compensation parameter adjustment amount, generating an optimized action set containing modulation format instructions and compensation parameter instructions, including: Extract the degradation type label and severity quantization value from the degradation feature description vector, combine the two into the current environment state vector, and stack this state vector with the historical state vector of the most recent N steps to form a state tensor, which generates the input state tensor of the deep deterministic policy gradient algorithm. The input state tensor is fed into the Actor online network of the deep deterministic policy gradient algorithm. Through the convolutional and fully connected layers in the network, a continuous two-dimensional action vector is output. The first dimension of the vector represents the continuous value of the modulation format switching level, and the second dimension represents the continuous value of the compensation parameter adjustment amount, thus generating the original continuous action vector. The continuous values of modulation format switching levels in the original continuous action vector are discretized into actual usable modulation format level numbers through nearest neighbor mapping. The continuous values of compensation parameter adjustment are constrained within the adjustable range by the amplitude limiting function and quantized into integer multiples of the step resolution of the digital signal processing unit to generate a discretized action vector. The modulation format level number in the discretized action vector is mapped to the corresponding modulation format switching protocol instruction code, and the compensation parameter adjustment amount is mapped to the setting instructions for dispersion compensation parameter value and nonlinear compensation parameter value. Finally, an optimized action set containing modulation format instruction and compensation parameter instruction is generated.
5. The method of claim 4, wherein, The digital signal processing unit that sends the optimized action set to the optical transceiver module performs online modulation format switching and dynamic updates of dispersion compensation and nonlinear compensation parameters to realize real-time performance monitoring and intelligent optimization of the optical transmission system, including: The modulation format switching instructions in the optimized action set are encoded into mode switching register write values that can be recognized by the digital signal processing unit, and the compensation parameter instructions are encoded into filter coefficient write values at the corresponding address of the digital signal processing unit, generating a register write instruction sequence that can be directly executed by the digital signal processing unit. The register write instruction sequence is sent to the control register group of the digital signal processing unit through the management data input / output interface of the optical transceiver module or the integrated circuit bus, triggering the digital signal processing unit to enter the parameter update preparation state and generate a register write completion status flag. After the digital signal processing unit detects the register write completion status flag, it synchronously performs modulation format switching and compensation parameter update actions at the current frame boundary to avoid service interruption caused by switching in the middle of the data packet and generates a parameter effective confirmation signal. The forward error correction bit rate is monitored over several consecutive sampling periods after the parameters take effect. If the bit error rate continues to decrease and the quality score increases, the optimization is confirmed to be effective. The confirmation result and the current optimization action are fed back to the Critic network of the deep deterministic policy gradient algorithm for experience playback and network parameter update, ultimately realizing real-time performance monitoring and intelligent optimization of the optical transmission system.
6. A real-time performance monitoring and intelligent optimization system for an optical transport system, characterized in that, The system includes: The acquisition module is used to collect the input optical power, bias current, temperature, voltage and dispersion compensation forward error correction code rate of the optical transceiver module in the optical transmission system in parallel at a fixed sampling period, and generate a multi-dimensional performance time series dataset. The extraction module is used to input the multidimensional performance time series dataset into the link quality assessment model built based on graph attention network, extract the coupling features between various performance parameters, and generate a quality score and degradation feature description vector reflecting the current link health by combining historical decay trends. The decision module is used to dynamically decide the modulation format switching level and the digital signal processing compensation parameter adjustment amount based on the degradation type and severity in the degradation feature description vector, and generate an optimized action set containing modulation format instructions and compensation parameter instructions. The execution module is used to send the optimized action set to the digital signal processing unit of the optical transceiver module, perform online switching of modulation format and dynamic updating of dispersion compensation and nonlinear compensation parameters, and realize real-time performance monitoring and intelligent optimization of the optical transmission system.
7. The system of claim 6, wherein, The acquisition module is specifically used for: Configure a sampling clock in the digital diagnostic monitoring interface of the optical transceiver module, set a fixed sampling period, and simultaneously send read requests for four analog quantities—input optical power, bias current, temperature, and voltage—to the module's internal register to generate a set of raw photoelectric analog readings. In the digital signal processing unit of the optical transceiver module, the bit error rate count before error correction is read from the forward error correction decoder with the same sampling period, and the bit error rate count is divided by the total number of error correction codewords to obtain the dispersion-compensated forward error correction bit rate, generating a set of original bit error rate readings. Align the four parameter values in the original set of photoelectric analog quantity readings with the original set of bit error rate readings according to the sampling timestamp, and perform range conversion and unit conversion on each parameter value to convert it into standard physical dimensions, generating a dimensionally normalized five-dimensional sampling record; The dimensionally normalized five-dimensional sampling records are written into a circular buffer in the order of sampling time. When the buffer is filled with a preset number of records, they are read out as a whole in the form of data blocks and start and end timestamps are attached to generate a multi-dimensional performance time series dataset.
8. The system of claim 7, wherein, The extraction module is specifically used for: The five performance parameter sequences in the multidimensional performance time series dataset are normalized respectively, and the five normalized values at the same time are combined into a five-dimensional feature vector, which is then arranged into a feature matrix in chronological order to generate a standardized time series feature matrix. The five-dimensional feature vector of each time step in the standardized temporal feature matrix is used as the node feature of the graph attention network, and the adjacency edge of the graph structure is defined by the physical relationship between each performance parameter. The mutual attention weight between nodes is calculated through the multi-head attention mechanism to generate the parameter coupling attention coefficient matrix. The parameter coupling attention coefficient matrix is weighted and fused with the original feature matrix to obtain the enhanced features of each parameter after being affected by other parameters at each time step. Then, the enhanced features are weighted and pooled along the time dimension to generate the overall quality score of the current link. The overall quality score is concatenated with the quality score sequences of multiple historical periods. The temporal gradient and second-order difference of the quality score are calculated to identify degradation patterns such as continuous decline, abrupt jumps, or periodic fluctuations. Finally, a quality score and degradation feature description vector reflecting the current link health are generated.
9. A storage medium, characterized by The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.
10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.
