A parameter tuning method and device of a ROADM system and a storage medium

Abstract

A parameter tuning method, device and storage medium of a ROADM system, the method comprising: step 30, at the current time t, using a deep learning model to obtain the OSNR prediction value of each wavelength at T1+t; step 40, according to the OSNR prediction value of each wavelength, for each wavelength, the following operation is performed: for the current wavelength, if the OTU power of the current wavelength is less than the OTU power upper limit value when the OSNR prediction value of the current wavelength is less than the OSNR lower limit value, at least one of the constraint condition and the initial solution generation condition of the genetic algorithm is adjusted; step 50, based on the operation result of step 40, at n*T2+t, the iteration operation of the genetic algorithm is performed, and the parameter tuning operation is performed according to the optimal solution obtained by the nth iteration; wherein T1=N*T2, wherein n=1, 2, 3, …, N, N is a positive integer, and the parameter tuning of the ROADM system is realized.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of information processing, and in particular to a parameter tuning method and device of a ROADM system and a storage medium. BACKGROUND

[0002] Parameter tuning of a ROADM (Reconfigurable Optical Add-Drop Multiplexer) system is a highly complex and specialized work. The debugging process involves fine adjustment of multiple key parameters, such as launch power, OA (Optical Amplifier) gain, VOA (Variable Optical Attenuator) attenuation, and gain slope. Changes in each parameter will significantly affect the overall performance of the system, and there are complex interactions between these parameters. In addition, to ensure the consistency of signal-to-noise ratio and transmission performance, the gain slope difference between different wavelength channels needs to be precisely controlled. At the same time, external environmental factors (such as temperature fluctuations and voltage instability) will significantly affect the performance of the ROADM system, increasing the complexity of debugging. Therefore, technicians must have a deep theoretical foundation in optical communication and rich field experience to make accurate judgments and take appropriate measures under actual conditions.

[0003] Currently, the tuning of a ROADM system is usually done manually or semi-automatically. However, existing debugging is highly dependent on subjective experience: engineers need to accumulate experience data, use semi-automatic means such as tables or formulas to simulate and calculate site configuration data, and manually fine-tune according to actual instrument measurement results, which needs to be repeated multiple times to obtain better data. This manual debugging is inefficient and difficult to respond to dynamic changes. Changes in external environment and equipment characteristics often lead to changes in parameters, resulting in debugging results that are not suitable and may cause communication accidents such as business errors and interruptions. SUMMARY

[0004] The present disclosure provides a parameter tuning method and device of a ROADM system and a storage medium.

[0005] A parameter tuning method of a ROADM system comprises the following steps:

[0006] Step 30: At the current time t, use a deep learning model to obtain the OSNR prediction value of each wavelength at T1+t.

[0007] Step 40: According to the OSNR prediction value of each wavelength, perform the following operations on each wavelength:

[0008] If the OTU power of the current wavelength is less than the upper limit of the OTU power when the predicted value of the OSNR of the current wavelength is less than the lower limit of the OSNR, at least one of the constraint condition and the initial solution generation condition of the genetic algorithm is adjusted;

[0009] Step 50, based on the operation result of step 40, performing an iterative operation of the genetic algorithm at the time of n*T2+t, and performing a parameter tuning operation according to the optimal solution obtained in the n th iteration when the optimal solution obtained in the n th iteration is obtained;

[0010] Wherein, T1=N*T2, wherein n=1,2,3,……,N, N is a positive integer. A storage medium, the storage medium has a computer program stored therein, wherein the computer program is configured to execute the method described above when running.

[0011] A parameter tuning device of a ROADM system, comprising a memory and a processor, the memory has a computer program stored therein, and the processor is configured to run the computer program to execute the method described above.

[0012] The embodiments of the present application realize the rapid response of the parameter tuning of the ROADM system through the genetic algorithm dynamic constraint guided by the deep learning model, and provide strong support for the efficient and stable operation of the optical communication network.

[0013] Other features and advantages of the present application will be set forth in the subsequent description, and some will become apparent from the description, or will be learned through practice of the present application. Other advantages of the present application can be achieved and obtained through the schemes described in the specification and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0014] The accompanying drawings are used to provide an understanding of the technical solutions of the present application, and constitute a part of the specification, and are used to explain the technical solutions of the present application together with the embodiments of the present application, and do not constitute a limitation on the technical solutions of the present application.

[0015] Figure 1 The flowchart of the parameter tuning method of the ROADM system provided by the embodiments of the present application is shown in the figure;

[0016] Figure 2 For Figure 1 Another flowchart of the method shown in the figure. DETAILED DESCRIPTION

[0017] The present application describes a plurality of embodiments, but the description is exemplary rather than limiting, and it will be apparent to those of ordinary skill in the art that more embodiments and implementations can be possible within the scope of the embodiments described in the present application. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are possible. Unless specifically intended to be limited, any feature or element of any embodiment can be used with any other feature or element of any other embodiment, or in any other embodiment.

[0018] The present application includes and contemplates combinations of features and elements known to those of ordinary skill in the art. The embodiments, features, and elements disclosed in the present application can also be combined with any conventional features or elements to form unique inventive solutions. Any feature or element of any embodiment can also be combined with features or elements from other inventive solutions to form another unique inventive solution. Therefore, it should be understood that any feature shown and / or discussed in the present application can be implemented alone or in any appropriate combination. Accordingly, the embodiments are not to be restricted, except as by the appended claims and their equivalents. Furthermore, various modifications and changes can be made within the scope of the claims.

[0019] Furthermore, in describing representative embodiments, the specification can have presented the method and / or process as a particular sequence of steps. However, to the extent that the method or process depends on the particular order of steps, this description should not be construe as implying that the steps are necessarily performed in this order. Other steps can be performed in between them; they can be performed in other sequences; or steps can be performed at the same time. Accordingly, this description should not be construed as limiting the described processes to the particular orders noted in the claims. Further, the claims should not be limited to the steps of the processes recited in the description, as the method and / or process steps can be varied in many ways. Therefore, the specific order of steps in the claims should not be construed as a limitation on the scope of the claims.

[0020] The embodiments of the present application provide a ROADM system parameter intelligent tuning system based on a hierarchical closed-loop control architecture, to cope with the complexity and inefficiency problems in the traditional ROADM system parameter tuning process, realize the automatic configuration and optimization of the ROADM system, and improve the tuning efficiency and system performance.

[0021] The architecture of the system is divided into three layers, namely the data acquisition layer, the control / processing layer and the algorithm engine layer, and the layers cooperate through standardized interfaces (NETCONF protocol) to form an organic whole, realizing the closed-loop acquisition, processing and optimization of data.

[0022] The data collection layer is responsible for periodically collecting key performance indicators such as OSNR and BER from OTU business units and OCM units. These data are transmitted to the control / processing layer through the NETCONF protocol after encryption. The control / processing layer cleans and standardizes the collected data and stores them in the SQLite database, providing support for subsequent analysis. At the same time, this layer contains global control modules (deployed in OA and OTU controllers) and local control modules (deployed in VOA controllers) to receive instructions from the algorithm engine layer and convert them into specific configuration operations, which are sent to each site.

[0023] The algorithm engine layer is the core of the entire architecture, which uses genetic algorithms and deep learning models to deeply process and analyze data. Genetic algorithms optimize search for ROADM system parameters by simulating selection, crossover, and mutation operations in biological evolution, finding the optimal configuration combination under given constraints. Deep learning models train historical data to uncover key factors affecting OSNR and BER and their relationships, and predict future OSNR and BER trends to achieve early warning or pre-adjustment functions. This combination of genetic algorithms and deep learning models can fully leverage their strengths to achieve precise optimization of ROADM system parameters.

[0024] In the data collection layer, the system periodically collects OSNR and BER data from OTU business units in source and destination sites and OCM units in relay sites. These data are encrypted and transmitted to the control / processing layer through the NETCONF protocol to ensure data security and reliability. The control / processing layer cleans and standardizes the data and stores them in the SQLite database to provide a basis for subsequent analysis and optimization.

[0025] The algorithm engine layer optimizes and predicts ROADM system parameters through the synergy of genetic algorithms and deep learning models. Genetic algorithms efficiently search the solution space to find the optimal configuration combination that meets the constraints. Deep learning models learn and analyze historical data to predict future system performance trends, providing a basis for adjusting system parameters in advance to achieve early warning or pre-adjustment functions.

[0026] where the optimal solution where:

[0027] represents the transmission power of K wavelengths (unit: dBm), k=1,2,3, …, K, where K is a positive integer;

[0028] represents the attenuation value of M sites (unit: dB), M is a positive integer;

[0029] represents a gain value (global scalar, unit: dB) of the optical amplifier;

[0030] represents a gain slope (global scalar, unit: dB / nm) of the optical amplifier;

[0031] In summary, the present application realizes intelligent optimization of the ROADM system parameters by constructing a hierarchical closed-loop control architecture and combining a genetic algorithm and a deep learning model. The scheme not only improves the optimization efficiency, but also enhances the stability and reliability of the system.

[0032] Figure 1 A flowchart of a parameter optimization method of a ROADM system provided by an embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, the method comprises the following steps. Figure 1

[0033] Step 30: At the current time t, the deep learning model is used to obtain the OSNR prediction value of each wavelength at T1+t.

[0034] In the related art, it is difficult to predict the OSNR change in advance. This step realizes the prediction of the OSNR at the future time by the deep learning model, provides forward-looking guidance for the subsequent parameter optimization, and thus can identify risks in advance, such as predicting the OSNR degradation, to provide a basis for timely adjusting the parameters.

[0035] Step 40: According to the OSNR prediction value of each wavelength, the following operation is performed for each wavelength: if the OSNR prediction value of the current wavelength is less than the lower limit value of the OSNR, and the OTU power of the current wavelength is less than the upper limit value of the OTU power, at least one of the constraint condition and the initial solution generation condition in the genetic algorithm is adjusted.

[0036] The prior art lacks effective genetic algorithm adjustment strategies when the OSNR prediction value is lower than the lower limit. This step flexibly adjusts the key elements of the genetic algorithm through conditional judgment, so that the genetic algorithm can better adapt to the state change of the current system, thereby improving the adaptability and effectiveness of the genetic algorithm under different conditions, and helping to find a better solution.

[0037] Step 50: Based on the operation result of step 40, the genetic algorithm is iterated at nT2+t, and when the optimal solution of the nth iteration is obtained, the parameter optimization operation is performed according to the optimal solution of the nth iteration.

[0038] Wherein, T1=NT2, n=1, 2, 3, …, N, and N is a positive integer.

[0039] ​Traditional parameter tuning methods often fail to effectively incorporate optimal solution information during the iteration process. This step, by iterating the genetic algorithm at specific time points and promptly tuning parameters based on the optimal solution, achieves dynamic and real-time tuning, accelerating the convergence speed of parameter tuning and improving its accuracy and efficiency.

[0040] This application provides an innovative parameter tuning method for ROADM systems. This method achieves intelligent and dynamic tuning of system parameters through the organic combination of deep learning models and genetic algorithms.

[0041] Specifically, in step 30, a deep learning model is used to predict the OSNR value, providing forward-looking guidance for subsequent operations; in step 40, key elements of the genetic algorithm are flexibly adjusted based on the prediction results, enhancing the algorithm's adaptability; and in step 50, through iterative operations and timely optimization based on the optimal solution, the efficiency and accuracy of parameter tuning are improved. Through the synergistic effect of these three steps, this method can effectively cope with the complex parameter interactions and dynamic changes in ROADM systems, reducing the complexity and workload of manual debugging, while simultaneously improving system stability and transmission performance. It provides an intelligent and automated solution for parameter tuning in ROADM systems, possessing significant technical advantages and practical application value.

[0042] In this embodiment, parameter optimization is based at least on a first correspondence, a second correspondence, and a third correspondence. Wherein:

[0043] The first correspondence records the change in OSNR for a single wavelength for every 1dB change in transmit power, reflecting the nonlinear relationship between power and signal-to-noise ratio. Specifically, when the power is less than 15dBm, increasing power significantly improves OSNR; however, when the power exceeds 15dBm, further increases in power have a significantly diminishing effect on OSNR improvement. For example, increasing power from 14dBm to 15dBm improves OSNR by 2dB; but increasing from 15dBm to 16dBm improves OSNR by less than 0.5dB. Its main function is to help quantify the marginal effect of power increases on OSNR, preventing the algorithm from entering the high-power, low-efficiency region. For example, the calculation expression for the first correspondence is: ,in This represents the transmit power at a single wavelength. This indicates the optical signal-to-noise ratio at that wavelength. This represents the partial derivative of OSNR with respect to power. This calculation expression describes the law of diminishing marginal returns to power increases.

[0044] The second correspondence records the change in power difference between adjacent wavelength channels at any given site for every 1dB change in VOA attenuation. For example, a 2dB increase in VOA results in a 1dB decrease in the power difference between wavelengths λ1 and λ2. This relationship guides the directional adjustment of the VOA to balance power differences between different wavelengths. For example, the calculation expression for the second correspondence is: ,in This indicates the adjustment amount for the VOA attenuation value. This represents the linear control coefficient of attenuation on flatness. This represents the change in power flatness between wavelengths. This calculation expression achieves a direct mapping between VOA adjustment and power flatness. For example, The value is 0.5.

[0045] The third correspondence relates to the change in multi-wavelength gain uniformity for every 0.1 dB / nm change in the gain slope of the OA (Optical Area Correlation) property. For example, when the gain slope increases from 0.2 dB / nm to 0.3 dB / nm, the flatness improves by 0.3 dB. Its main function is to constrain the solution space boundary, ensuring that the initial solution meets the flatness requirements. For example, the calculation expression for the third correspondence is: ,in This indicates the multi-wavelength gain uniformity index. Indicates the gain slope of OA. This represents the partial derivative of the OA gain slope with respect to multi-wavelength gain uniformity. This calculated expression describes the impact of slope fine-tuning on system-level performance.

[0046] Based on the above three correspondences, in this embodiment of the application, after predicting the OSNR change trend after a duration of T1 (e.g., 1 hour) using a deep learning model, each wavelength is pre-adjusted. Based on the pre-adjustment results, the optimal solution is obtained every T2 (e.g., 5 minutes) using a genetic algorithm, and the parameters are optimized according to the obtained optimal solution. Here, T1=N*T2, n=1,2,3,……,N, where N is a positive integer.

[0047] In an exemplary embodiment, the input layer of the deep learning model sets the weight value of the transmission power based on the first correspondence, and the hidden layer is set with an activation function obtained based on the third correspondence;

[0048] The core purpose of this step is to identify wavelength channels that may degrade in advance, providing a basis for subsequent pre-adjustments. The deep learning model analyzes historical data to predict the OSNR values ​​of each wavelength at the future time T1+t.

[0049] In the input layer, weight values ​​are set based on the first correspondence, which can deweight the features of high-power wavelengths (>15dBm). For example, when the power of a certain wavelength reaches 16dBm, its input weight is reduced to 0.27. This design can effectively suppress the interference of nonlinear effects (such as four-wave mixing) on ​​the prediction and improve the prediction accuracy.

[0050] In the hidden layer, a third correspondence is used to construct the activation function. Specifically, the parameter γ is used as a parameter in the activation function to ensure that the prediction results conform to the physical laws of optical communication. For example, when the slope of OA changes by 0.1 dB / nm, the model can accurately reflect the improvement trend of flatness by 0.3 dB.

[0051] By using predictions constrained by physical laws, the risk of OSNR degradation at edge wavelengths (such as λ1 dropping from 22dB to 19dB) can be identified in advance at a time T1 in scenarios with rising temperatures, providing a time window for proactive management.

[0052] Figure 2 for Figure 1 Another flowchart illustrating the method is shown. Figure 2 As shown, the method further explains steps 40 and 50:

[0053] In one exemplary embodiment, step 40 includes steps 41 to 43, wherein:

[0054] Step 41: When the predicted OSNR value of the k-th wavelength is less than the lower limit of OSNR (e.g., 20dB), if the OTU power of the k-th wavelength is less than the upper limit of OTU power (e.g., 15dBm), perform at least one of operation 1 and operation 2.

[0055] Step 42: Determine if k is greater than or equal to K;

[0056] If yes, proceed to step 50; otherwise, proceed to step 43.

[0057] Step 43: Update the value of k to k+1, and then execute step 41.

[0058] This step is a wavelength-level pre-adjustment operation. It dynamically configures the search rules of the genetic algorithm by performing virtual adjustment operations on the prediction results instead of directly issuing physical commands.

[0059] Operation 1 (OTU / VOA adjustment):

[0060] If the current power is less than the upper limit of the OTU power, increase the OTU transmit power of the current wavelength as the initial solution generation condition for the current wavelength in the genetic algorithm; otherwise, adjust the VOA attenuation value of the neighboring stations of the current wavelength as the generation range of the initial solution of the current wavelength in the neighboring stations in the genetic algorithm.

[0061] For example, when the predicted OSNR value for a certain wavelength is below the lower limit (e.g., 20dB) and the current power is <15dBm, the initial solution tendency of the genetic algorithm is set. For example, if the current power of λ1 is 14dBm, the initial solution range is limited to [14, 14.5]dBm; if the power has reached the upper limit of 15dBm, the adjustment range of VOA for adjacent stations is calculated according to the second correspondence (e.g., when β=0.5, if the flatness needs to be improved by 0.5dB, then the VOA is reduced by 1dB).

[0062] Operation 2 (adjusting OSNR constraints):

[0063] Increase the lower limit value in the constraints of ONSR to construct a preventative safety buffer zone.

[0064] The adjusted lower limit can be Γ+|ΔOSNR pred |, where Γ is the preset OSNR security threshold, and ΔOSNR pred This represents the change in OSNR predicted by the deep learning model.

[0065] For example, the OSNR constraint for this wavelength can be actively tightened from the basic threshold "OSNR≥20dB" to "OSNR≥22dB", forcing the genetic algorithm to search for a higher-performance solution space.

[0066] By modifying the search rules of the genetic algorithm for each wavelength, the algorithm avoids blindly searching for invalid regions, improves the efficiency of finding the optimal solution, and prevents short-term optimization from leading to long-term degradation.

[0067] In one exemplary embodiment, step 50 includes:

[0068] Step 51: At time n*T2+t, based on the constraint solution space boundary of the nth iteration determined by the third correspondence, obtain the initial solution of the nth iteration; for example, when the current flatness is 1.5dB, the target is 0.8dB, and γ=3dB / (dB / nm), the initial solution range of the OA slope is set to [0.23, 0.47]dB / nm to ensure that the initial population meets the physical feasibility.

[0069] Step 52: Obtain the value of the objective function for the nth iteration, and obtain the fitness of the nth iteration based on the value of the objective function. The objective function includes the weight value of the single-wavelength power penalty term determined based on the first correspondence. The weight of the single-wavelength penalty term in the objective function is dynamically set according to the first correspondence. For example, when the power is >15dBm, due to the diminishing marginal return of the OSNR gain, the weight is increased, severely penalizing high-power, inefficient solutions.

[0070] Step 53: Under the constraints of the nth iteration, based on the fitness of the nth iteration, perform genetic computation on the initial solution of the nth iteration until a preset termination condition is met. The genetic computation includes mutation operation based on the second correspondence. Non-random mutation is performed using the second correspondence. For example, when it is necessary to improve the flatness by 0.5dB (β=0.5), the VOA attenuation value is directly calculated to reduce by 1dB to avoid invalid perturbation.

[0071] Step 54: When the optimal solution is obtained in the nth iteration, perform parameter tuning based on the optimal solution obtained in the nth iteration; wherein, the optimal solution includes the transmit power of each wavelength, the VOA attenuation value (unit: dB) of each site, the OA increase and the gain slope.

[0072] Step 55: Determine whether n is greater than or equal to N. If n is less than N, execute step 51. If n is greater than or equal to N, the iteration process of step 50 ends. After that, you can continue to execute step 30 or end the tuning operation.

[0073] Where T1 = N * T2, where n = 1, 2, 3, ..., N, and N is a positive integer.

[0074] Based on the above analysis, it can be seen that the method provided in this application has the following technical advantages:

[0075] 1. Solve the problem of delayed response in dynamic environments.

[0076] Rapid Response: Risks are identified in advance through prediction by deep learning models (T1 cycle), and minute-level closed-loop optimization is achieved in genetic algorithms (T2 cycle). Compared with the hour-level response of traditional manual debugging, the response speed is improved, thus enabling rapid prediction and optimization.

[0077] Multi-level response mechanism: By decoupling in the time dimension, a long-term risk scan (hour-level) is performed in the T1 cycle, and a short-term rapid response (minute-level) is achieved in the T2 cycle, forming a "prediction-optimization" dual-rate control (T1=N*T2), which effectively balances the foresight of prediction and the timeliness of optimization.

[0078] 2. Overcoming the bottleneck of multi-objective optimization conflict

[0079] Physical law embedding: Operation 1 (OTU / VOA bias setting) avoids invalid search regions, such as inefficient regions with power >15dBm; Operation 2 (OSNR constraint dynamic tightening) reduces the solution space and improves Pareto front convergence efficiency.

[0080] Global-local coordination: Combining pre-adjustment (wavelength level) with genetic optimization (system level) to simultaneously ensure single-wavelength performance and network-wide security, breaking through the bottleneck of multi-objective optimization conflict.

[0081] 3. Suppressing systemic risks caused by nonlinear effects

[0082] Risk mitigation mechanism: Combining wavelength-level pre-adjustment and system-level genetic optimization, pre-adjustment avoids blindly increasing single-wave power, and the genetic algorithm embeds a penalty term in its objective function to automatically reduce the weight of high-power wavelengths, thereby reducing the probability of four-wave mixing risk.

[0083] Physical laws drive AI: the first correspondence suppresses high-power and inefficient operation, the second correspondence makes VOA adjustment more precise, and the third correspondence ensures the physical feasibility of the initial solution. Physical laws are deeply integrated into the AI ​​model, promoting the transformation of ROADM system from passive maintenance to proactive maintenance.

[0084] In summary, by using deep learning model-guided genetic algorithm dynamic constraints, rapid response, multi-objective optimization, and risk suppression in ROADM system parameter tuning were achieved, providing strong support for the efficient and stable operation of optical communication networks.

[0085] In one exemplary embodiment, a monitoring and assessment mechanism for the risk value of four-wave mixing is introduced, and corresponding operations are performed accordingly to ensure the performance and reliability of the system.

[0086] Four-wave mixing is a nonlinear optical effect. In ROADM systems, when multiple wavelengths of optical signals are transmitted through optical fibers, the nonlinear characteristics of the fiber may generate new, unwanted optical frequencies (i.e., idler frequencies). These idler frequencies interfere with the signal light, reducing the OSNR and consequently increasing the BER, thus affecting the system's transmission performance and reliability. Therefore, monitoring the risk value of four-wave mixing is of great significance. Its purpose is to provide early warning of potential risks so that timely measures can be taken to avoid system performance degradation caused by the four-wave mixing effect.

[0087] During the time interval from time t to time T1+t, the risk value of four-wave mixing is obtained, and if the risk value is greater than a preset risk threshold, at least one of operations 3, 4, and 5 is executed, wherein:

[0088] Operation 3 involves lowering the upper limit value of the total power constraint in the genetic algorithm;

[0089] Operation 4 involves adjusting the weight of the deviation between the total power and the optimal total power in the objective function of the genetic algorithm when the effective operating bandwidth of the optical amplifier meets the preset narrow bandwidth condition or when the optical fiber used is an optical fiber including G.657 fiber.

[0090] Operation 5 involves increasing the weight of the single-wave power penalty term in the objective function of the genetic algorithm.

[0091] For example, the preset risk threshold is 20 dBm 2 / THz, when the risk value exceeds the risk threshold, the idler optical power generated by four-wave mixing will significantly interfere with the signal, resulting in OSNR degradation and BER increase.

[0092] By performing operation 3, the upper limit of single-wave power constraints can be reset based on the total power increase predicted by the deep learning model and the preset safety margin. This operation forces the algorithm to reduce the total power in subsequent iterations, thereby actively avoiding the risk of nonlinear effects caused by excessive total power, reducing the idler optical power generated by four-wave mixing, and avoiding serious interference to the signal.

[0093] By performing operation 4, the weight of the total power deviation is increased, causing the algorithm to prioritize the optimization of total power during the optimization process, thus suppressing the risk of four-wave mixing that may result from high power. This allows the system to prioritize reducing total power while meeting other performance indicators, thereby reducing the risk of four-wave mixing.

[0094] By performing operation 5, the situation of excessive single-wave power can be effectively avoided, thereby effectively limiting the total power and reducing the risk value of four-wave mixing.

[0095] By performing steps 3 to 5, the fundamental conflict between improving OSNR and avoiding FWM caused by total power exceeding limits during ROADM tuning can be effectively resolved. This ensures that signal quality is not significantly affected by the four-wave mixing effect while meeting system performance requirements.

[0096] Optionally, the risk value can be obtained in the following ways:

[0097] The risk value is obtained by multiplying the square of the actual total power value by the total number of current wavelength channels, and then calculating the ratio of the product to the effective operating bandwidth of the optical amplifier.

[0098] Based on the above method, potential four-wave mixing risks can be detected in a timely manner through simple calculations, thus avoiding system performance degradation.

[0099] In one exemplary implementation, a total power prediction response mechanism is added to address the risk of system-level nonlinear effects through a dual guarantee of dynamic constraints and weight adjustment. The specific implementation is as follows:

[0100] In step 30, the predicted value of the total power is obtained using the deep learning model, thereby enabling the solution space to be actively shrunk before the genetic algorithm iteration;

[0101] In step 40, if the predicted value of the total power is greater than the total power threshold, the upper limit value of the constraint condition of the total power in the genetic algorithm is lowered; the content of operation 4 is to adjust the weight value of the deviation between the total power and the optimal total power in the objective function of the genetic algorithm.

[0102] Among them, the total power threshold is ,in This is the upper limit for hardware security (e.g., 23dBm). To predict the power increase, Set a safety margin (e.g., 2dB).

[0103] The technical features in the above exemplary embodiments are designed to avoid amplifier saturation. Specifically, when the deep learning model predicts that the total power is about to approach the hardware limit (e.g., 23dBm), the system will reserve a certain safety margin (e.g., 2dB) in advance. This measure can effectively prevent optical module overload problems caused by sudden traffic surges. By tightening the power constraint, the genetic algorithm is forced to reduce the total power distribution of the entire network during the optimization process.

[0104] In summary, setting an absolute upper limit on power prevents the system from entering a dangerous state and effectively avoids the vicious cycle problem that may occur in traditional solutions: "power reduction → OSNR degradation → power increase → nonlinear effects".

[0105] In one exemplary embodiment, a physical model of ambient temperature and equipment aging (the fourth and fifth correspondences) is further introduced, and dynamic environmental adaptive optimization is achieved through the collaborative processing of deep learning and genetic algorithms.

[0106] The fourth correspondence records the change in power attenuation per unit length of optical fiber after a preset time period for every 1°C increase in temperature. For example, the calculation expression for the fourth correspondence is ΔA = 0.05*ΔT*L, where ΔT is the temperature change (°C), L is the optical fiber length (km), and ΔA is the attenuation increment per unit length of optical fiber after 2 hours (dB / km).

[0107] This fourth correspondence serves as a physical parameter for the hidden layer in the deep learning model, used to predict future fiber attenuation, thereby enabling accurate prediction of power attenuation caused by temperature changes.

[0108] The fifth correspondence records the change in OA gain attenuation for each additional month of equipment operation time; for example, the calculation expression for the fifth correspondence is ΔG = k*t, where k is the monthly OA gain attenuation rate (e.g., 0.1dB / month) and t is the number of months the equipment has been in operation.

[0109] When the historical database detects a continuous drift in OA gain (e.g., a monthly decrease of 0.1 dB for three consecutive months), the gain constraint range will be automatically adjusted to allow the genetic algorithm to consider gain changes caused by equipment aging during the iteration process. The specific calculation expression is as follows:

[0110] ;

[0111] Where Gmin0 is the initial lower limit of gain (5dB), Gmax0 is the initial upper limit of gain (25dB), k is the monthly attenuation rate statistically analyzed in the historical database (e.g., 0.1dB / month), and t is the number of months the device has been in operation.

[0112] For example, a certain span of optical fiber is 80km long, and the equipment has been running for 18 months (k=0.1dB / month), with a temperature rise warning ΔT=8℃.

[0113] In the deep learning model, based on the fourth correspondence, ΔA = 0.05 × 8 × 80 = 3.2 dB is calculated, predicting that the decay will increase by 3.2 dB after 2 hours.

[0114] Genetic algorithm dynamic constraints: Total power limit: Pmax = 23 - 3.2 = 19.8 dBm (to prevent nonlinear effects); OA gain boundary: G ∈ [5 + 0.1 × 18, 25 + 0.1 × 18] = [6.8, 26.8] dB.

[0115] In summary, by transforming temperature-hysteresis (the fourth correspondence) and aging-gain drift (the fifth correspondence) into a computable model, we can overcome the limitations of traditional empirical tuning.

[0116] In one exemplary embodiment, a closed-loop knowledge iteration mechanism is established to continuously optimize the deep learning model using quadruple data, thus addressing the mismatch between static models and the environment in traditional solutions. The specific implementation is as follows:

[0117] exist Figure 1 The method shown further includes the following steps:

[0118] Step 10: Obtain the operating data of the ROADM system from time t-T1 to the current time t, where each set of operating data includes environmental conditions, aging coefficient, optimal solution and actual performance;

[0119] Step 20: Perform incremental training on the deep learning model based on the running data, and then execute step 30.

[0120] In step 10, a quadruple data structure is used to comprehensively record key information such as environmental state, aging coefficient, optimal solution, and actual performance. Specifically, this includes:

[0121] Environmental conditions (St): such as external factors like temperature and humidity, for example, St=35℃.

[0122] Aging factor (Kt): such as the monthly degradation rate of the equipment, for example Kt=0.1 dB / month.

[0123] Optimal solution (X*): The configuration output by the genetic algorithm, for example [P{λ1}=14.5 dBm, V{R2}=9 dB].

[0124] Actual performance (Yt): such as measured indicators like OSNR and BER, for example [OSNR=21 dB, BER=1e-13].

[0125] This quadruple structure can systematically integrate data from different dimensions, providing rich training materials for the optimization of deep learning models.

[0126] In step 20, the incremental training process includes the following main steps:

[0127] Data acquisition: Collect quadruple data from time t-T1 to time t.

[0128] Model input: The input layer receives the environmental state St and the aging coefficient Kt.

[0129] Hidden layer supervision: The hidden layer uses the optimal solution X* as the supervision signal to learn optimization strategies.

[0130] Output layer supervision: The output layer performs supervised learning based on the actual performance Yt.

[0131] Backpropagation update: The weights of the deep learning model are updated through the backpropagation algorithm, thereby optimizing the prediction model.

[0132] The aforementioned incremental training process starts from the real-world environment, employing genetic algorithm iteration, quadruple generation, and LSTM incremental training to ultimately achieve high-precision predictions, forming a closed-loop feedback loop that continuously improves model accuracy. Furthermore, the optimal solution X* from the quadruple plays a supervisory role in the hidden layer, aiding in learning and optimizing strategies and improving decision consistency; the actual performance Yt is used for supervision in the output layer, ensuring that the prediction results closely reflect the actual physical environment and guaranteeing the accuracy and reliability of the predictions. This enables the deep learning model to dynamically adapt to environmental changes and device aging, continuously improving prediction accuracy.

[0133] In summary, the technical effects achieved by the above exemplary embodiments include:

[0134] Eliminating cold start error: The initial model relies on theoretical physical relationships. By injecting actual network data through incremental training, the prediction error in the early stages of operation can be effectively reduced.

[0135] Adaptive regional differences: To address the differences in fiber optic attenuation caused by different environments such as dryness in the north and humidity in the south, the quadruplet data dynamically corrects the prediction model, ensuring that the OSNR prediction error is within a preset range when deployed across regions.

[0136] Decision knowledge transfer: The optimal solution X* output by the genetic algorithm serves as a supervision signal, guiding the optimization strategy of the deep learning model and making the prediction structure approximate the decision logic of the genetic algorithm.

[0137] Avoid long-term degradation: When the actual performance Yt deteriorates (such as a sudden increase in BER), the prediction bias can be corrected in a timely manner.

[0138] In one exemplary embodiment, a real-time local fine-tuning mechanism is added to address the problem of instantaneous power imbalance in single waves that may occur after global optimization.

[0139] The mechanism is triggered by real-time detection of single-wavelength power deviation, specifically whether the difference between the actual power and the target power exceeds a preset threshold (e.g., 1 dB). This detection is based on second-level real-time data acquisition from the OTU / OCM device, rather than predictions, ensuring timely and accurate response.

[0140] The adjustment operation targets only the station where the power deviation occurs at the target wavelength, rather than interfering with the entire network. For example, if the power deviation of wavelength λ1 at station R2 reaches 1.5dB, the adjustment command is only sent to VOA_λ1 at station R2, achieving precise and localized power adjustment.

[0141] Specifically, the adjustment command generation directly references the VOA value from the optimal solution output by the genetic algorithm. When a power deviation exceeding the limit is detected in real time, the system automatically issues a command to set the VOA of the corresponding site and wavelength to the optimal value pre-calculated by the genetic algorithm. This process avoids real-time optimization calculations, thus greatly improving response speed.

[0142] For example, after global optimization, if the real-time power of wavelength λ1 at site R2 drops from 14.5 dBm to 13.0 dBm (a deviation of 1.5 dB exceeds the 1 dB threshold), a local fine-tuning mechanism is triggered. The specific process is as follows:

[0143] Real-time detection: OCM periodically reports power data and identifies power deviations exceeding the standard for λ1.

[0144] Command generation: Query the latest optimal solution of the genetic algorithm to find that VOA_R2_λ1=9dB, and generate the command: "SET VOA_R2_λ1 = 9dB".

[0145] Execution: The command is directed to the R2 site via the NETCONF protocol.

[0146] As explained above, the global genetic algorithm handles cross-site power balancing with a 5-minute cycle, but may experience response delays during optimization intervals, which can be ignored due to second-level fluctuations. Local VOA adjustment, on the other hand, achieves real-time (second-level) single-wavelength power fine-tuning, reducing service interruption time and computational overhead. Furthermore, computation-free adjustment based on the genetic solution directly calls the VOA value from the optimal solution of the genetic algorithm, eliminating the need for real-time optimization calculations and improving response speed from seconds to milliseconds. It also avoids cross-site power oscillations that may result from local optimization.

[0147] In an exemplary embodiment, the multidimensional fusion mechanism of the genetic algorithm objective function is further improved. In addition to the two sets of key parameter instructions mentioned above, at least one of the following three sets of key parameters can be dynamically weighted, which effectively solves the problem of Pareto convergence in multi-objective optimization.

[0148] In one implementation, the objective function is calculated as follows:

[0149] ;

[0150] in, To optimize variables, , , , All represent weighting coefficients (values ​​can be 0.5, 0.3, 0.1, or 0.1).

[0151] in, This represents the actual bit error rate. Indicates the bit error rate threshold. This represents the standard deviation of OSNR for each wavelength. This is the actual value of the total power. This is the optimal total power point;

[0152] in, ;

[0153] in, For the k-th single-wave power, Let K be the OSNR power efficiency coefficient, k = 1, 2, 3, ..., K; when Exponential decay occurs at >15dBm, forcing the algorithm to avoid the high-power inefficient region.

[0154] In another implementation, a prediction bias penalty term is added to the objective function F(x) mentioned above. This improvement forces the algorithm to optimize current performance while also considering prediction accuracy, thereby avoiding long-term degradation caused by short-term optimization. The calculation expression is as follows:

[0155]

[0156] in, These are the predicted values ​​from the deep learning model. These are the actual collected values. For prediction time windows.

[0157] Based on the above explanation, it can be seen that the objective function at time n*T2+t is also based on at least one of the following determined factors:

[0158] The deviation between the actual bit error rate at time n*T2+t and the preset bit error rate threshold is used to penalize solutions with high bit error rates.

[0159] The deviation between the variance of OSNR at each wavelength and the average value of the sum of OSNR at time n*T2+t is used to suppress power imbalance between wavelengths.

[0160] The prediction error of OSNR at time n*T2+t is used to force the genetic algorithm to take into account the long-term trend; wherein, the prediction error is the ratio of the absolute value of the difference between OSNR1 and OSNR2 to the time difference, where OSNR1 represents the actual value of OSNR at time n*T2+t, OSNR2 represents the predicted value of OSNR at time T1+t, and the time difference is T1-n*T2.

[0161] Among them, the bit error rate and OSNR variance are derived from real-time data collected by OTU / OCM and are mainly used to optimize current performance and solve the problem of instantaneous service degradation. The prediction error is calculated by comparing the predicted value of the deep learning model with the actual value to ensure long-term stability and avoid optimization methods that result in short-term gains but long-term deterioration.

[0162] By monitoring key system indicators in real time (such as bit error rate, OSNR uniformity, and nonlinearity risk), and combining network scenario characteristics (such as backbone nodes, DWDM architecture, and fiber type) and environmental changes (such as temperature fluctuations), the values ​​of each weight in the objective function are automatically adjusted. Specifically, this includes:

[0163] For weight w 1 (Bit Error Rate Weight):

[0164] When the actual BER value approaches or exceeds the BER threshold, especially at core nodes of the backbone network, it is necessary to prioritize adjusting the OTU power / OA gain to quickly suppress the bit error rate. At this point, the weights... w 1. The adjustment range is from 0.5 to between 0.6 and 0.7.

[0165] If the actual BER value is less than 10 -4 Furthermore, when balancing other metrics, to avoid over-optimization leading to excessive total power or deterioration of OSNR (optical signal-to-noise ratio), the weights should be adjusted accordingly.w 1. Decrease, the adjustment range is from 0.5 to between 0.3 and 0.4.

[0166] For weight w 2 (OSNR uniformity weight):

[0167] In DWDM (Dense Wavelength Division Multiplexing) / long-distance links, if the OSNR standard deviation is greater than 3.0 dB, the VOA attenuation value should be adjusted first to suppress edge wavelength degradation. w 2. The adjustment range is from 0.3 to between 0.4 and 0.5.

[0168] When the number of wavelengths is less than 16 or the uniformity requirement is low, in order to free up resources for optimizing other indicators, the weights are adjusted. w 2 decreased, and the adjustment range was changed from 0.3 to between 0.1 and 0.2.

[0169] For weight w 3 (Total Power Deviation Weight):

[0170] In the case of narrow bandwidth amplifiers (less than 40 THz) or G.657 fiber, if the risk of four-wave mixing exceeds 20 dBm. 2 / THz, in order to dynamically lower the power limit and prevent nonlinear effects caused by excessive power, weighting w The value increased by 3, and its adjustment range was from 0.2 to between 0.3 and 0.4.

[0171] For scenarios involving high-power tolerant amplifiers (greater than 25dBm) and transmission distances less than 80km, weighting is used to improve power allocation flexibility. w The value decreased by 3, and its adjustment range was adjusted from 0.2 to between 0.05 and 0.1.

[0172] For weight w 4 (Single-wave power penalty):

[0173] When the single-wave power prediction exceeds 15 dBm or the four-wave mixing risk exceeds 20 dBm 2 At / THz, to avoid inefficiency in the high-power region, the weight w4 is increased, and the adjustment range is from 0.1 to between 0.15 and 0.2.

[0174] If the temperature increases by a factor Δ T If the temperature is greater than or equal to 2℃, then the weight is... w 4 will follow Δ w 4 = 0.02 × Δ T The rules are adjusted to compensate for power fluctuations caused by temperature.

[0175] When the single-wave power prediction for all wavelengths is less than 12 dBm and the four-wave mixing risk is less than 10 dBm 2 At / THz, in order to free up power adjustment space, weighting w The value decreased by 4, from 0.1 to 0.05.

[0176] For weight w 5 (Prediction Error Penalty):

[0177] When the prediction error rate of a deep learning model exceeds 3% or exceeds the error rate three times consecutively, in order to force the algorithm to improve prediction reliability and avoid long-term degradation, the weights are adjusted. w 5. The adjustment range is from 0.05 to between 0.08 and 0.1.

[0178] If the prediction error rate of the deep learning model is less than 1% and stable, then the focus should be on optimizing the current performance, and the weights... w The value decreased from 0.05 to 0.02.

[0179] The exemplary embodiment described above introduces a dynamic weighting mechanism to achieve multi-timescale fusion of second-level real-time parameters and hour-level prediction parameters. Utilizing three types of parameters—bit error rate deviation, OSNR uniformity deviation, and prediction error penalty term—the genetic algorithm balances current optimal performance with long-term stability during optimization. The dynamic weight allocation mechanism adjusts weights based on the remaining duration of the prediction time window, resolving the conflict between short-term optimization and long-term degradation, and improving the efficiency of Pareto front search.

[0180] Optionally, the constraints of the genetic algorithm include not only total power constraints and OSNR constraints for each wavelength, but also VOA constraints and OA gain constraints.

[0181] For example, the constraints of genetic algorithms include: ; ; ; .

[0182] In one exemplary embodiment, to support data optimization operations, the data access and preprocessing methods are further defined, and the specific operations are as follows:

[0183] The system synchronously receives single-wavelength emitting power and receiving power data from the OTU service unit, as well as spectral information acquired by the OCM. This data is synchronized through a timestamp alignment mechanism. For example, the OTU service unit obtains single-wavelength emitting power and receiving power data by directly reading the electrical signal conversion values ​​from the optical module, while the spectral information acquired by the OCM includes power, flatness, and OSNR for each wavelength. All of this data needs to be integrated within a unified timestamp framework to ensure the accuracy of subsequent processing.

[0184] After data is received, outlier filtering is required to ensure data quality. The system uses the 3σ principle to eliminate power jump anomalies. For example, if the single-wavelength optical power suddenly exceeds the normal range by ±20%, it is identified as an outlier and removed according to this principle. For invalid data segments with OSNR below 15dB, linear interpolation is used to complete the data to ensure data continuity and accurately reflect the actual operating status of the system.

[0185] After outlier filtering, the data is standardized. Specifically, power values ​​collected by different devices are converted to dBm units to ensure comparability of data from various devices. Simultaneously, spectral flatness is normalized and expressed as a deviation value relative to a reference wavelength, thereby eliminating dimensional differences between measurement results from different devices and providing standardized data support for subsequent data analysis and model training.

[0186] Furthermore, to efficiently store and manage data, the system employs a tiered storage architecture. On one hand, a real-time database is constructed, using an SQLite database to build time-series tables and storing raw collected data at 10-second intervals. This real-time database supports millisecond-level query responses, ensuring rapid access to the latest data. On the other hand, a historical database is established, extracting aggregated data from the real-time database every 30 seconds. For example, it calculates the power mean and OSNR variance over a 5-minute period and synchronizes this data to the historical database via the NETCONF protocol. This aggregated data forms the dataset for algorithm training, providing a foundation for training deep learning models.

[0187] Through the comprehensive design of multi-source data fusion processing, outlier filtering, standardization, and hierarchical storage architecture, the data access and preprocessing module provides high-quality, standardized, and efficiently accessible data support for subsequent operations of the ROADM system parameter tuning platform. This meticulously processed data not only ensures the accuracy of model training but also lays a solid foundation for real-time system monitoring and rapid response.

[0188] The method provided in the embodiments of this application is illustrated below with application examples:

[0189] Genetic algorithms are optimization algorithms that simulate natural selection and genetic mechanisms. Through operations such as population initialization, fitness calculation, selection, crossover, and mutation, they progressively search for the optimal solution. Applying this algorithm to ROADM system parameter tuning primarily benefits from its powerful global search capability and excellent handling of complex, multi-constraint problems.

[0190] In this application example, during population initialization, an initial population containing 100 feasible solutions is generated, centered on the current configuration. The OTU power is randomly generated within ±3dB of the current value. In the fitness calculation phase, for each solution, a simulation configuration is issued through the control / processing layer. The feedback OSNR and BER data are collected and substituted into the objective function to calculate the fitness, which is defined as 1 / F(x), where F(x) is the objective function.

[0191] The selection operation employs a roulette wheel approach, retaining the top 30% of high-quality solutions. The crossover operation crosses parameters between selected solutions, such as swapping 10% of the VOA attenuation and OA gain parameters. The mutation operation applies a 5% probability to the single-wave power with a random perturbation of ±0.5dB to avoid getting trapped in local optima. The iteration terminates when the fitness improvement is less than 1% for five consecutive generations, or when the upper limit of 100 iterations is reached, at which point the optimal configuration combination is output.

[0192] The deep learning model used in this example is the LSTM (Long Short-Term Memory) model. LSTM is a special type of recurrent neural network (RNN) designed to solve the long-term dependency problem of traditional RNNs. It can effectively process and predict long-term dependencies in time series data, thereby providing predictive assistance for the optimization process.

[0193] In this application example, the LSTM model is introduced into the ROADM system tuning process to predict the OSNR change trend over a short period of time in the future, so as to achieve early warning or pre-adjustment.

[0194] The input layer of the LSTM model receives 72 hours of historical data on temperature, power at each wavelength, and OSNR, with a time window sliding step of 10 minutes. The hidden layer contains two layers of 128-node LSTM units, capable of capturing the hysteresis effect of temperature-power attenuation; for example, for every 1°C increase in temperature, fiber attenuation increases by 0.05 dB / km after 2 hours. The output layer predicts the OSNR trend at each wavelength for the next hour, with an error rate controlled within 3%.

[0195] By combining LSTM prediction with genetic algorithm optimization, and through mechanisms such as dynamic constraint injection, prediction bias penalty, and feedback from the optimal solution to the LSTM prediction model, the two work together dynamically to improve optimization performance. A genetic algorithm iteration is triggered every 5 minutes, synchronously updating global and local configurations to form a closed loop of "collection-optimization-distribution-feedback," enabling dynamic and intelligent tuning of ROADM system parameters.

[0196] The following example, using a scenario of rising temperature, details the execution process of this application instance:

[0197] The temperature rise scenario demonstrates how the intelligent tuning system can automatically optimize ROADM system parameters and ensure stable system performance when changes in ambient temperature cause changes in fiber optic attenuation. This is achieved through data acquisition, optimization engine processing, hierarchical command issuance, closed-loop feedback verification, and knowledge iterative storage.

[0198] Step 1: Start Data Acquisition

[0199] The ambient temperature abnormally rose from 23℃ to 35℃, triggering data acquisition. The system collected data from the OTU service unit, including the emitting power at a specific wavelength (λ1) of +14dBm and the received power. At the same time, it collected data from the OCM monitoring unit, including the average OSNR decreasing by 1.5dB (from 22dB to 20.5dB), the BER deteriorating from 1e-13 to 5e-13, and the spectral flatness deviation increasing to 2.1dB.

[0200] Step 2: Optimize engine co-processing

[0201] LSTM prediction engine activated

[0202] The LSTM prediction engine receives 72 hours of historical temperature, power at each wavelength, and OSNR data (with a 10-minute sliding window). A 2×128-node LSTM unit in the hidden layer captures the temperature-attenuation hysteresis effect, quantizing the relationship as follows: for every 1°C increase in temperature, fiber attenuation increases by 0.05 dB / km after 2 hours. The output layer predicts that the OSNR of λ1 will decrease by another 0.8 dB within 1 hour, with an error rate ≤3%. Based on the prediction results, a pre-adjustment strategy is triggered, updating the genetic algorithm constraints. Because the current power (+14 dBm) is below the +15 dBm upper limit, the plan is to increase the OTU power by 0.5 dB.

[0203] Genetic Algorithm Optimization Engine

[0204] During population initialization, 100 feasible solutions are generated centered on the current configuration, with Power(λk) = 14 + 3×Random(-1, 1) dBm. In the fitness calculation phase, the simulated configuration is distributed, OSNR / BER data is collected, and the F(x) value is calculated. In the genetic computation, the top 30% of solutions are selected using a roulette wheel selection method, and the parameters of VOA and OA gains are cross-parameterized (10% parameter swap), with a 5% probability of perturbing the λ1 power by ±0.5dB. Upon reaching the termination condition, the optimal solution is output: OTU power increases by 0.5dB to 14.5dBm, and the VOA of station R2 decreases by 1dB to 9dB. The optimal solution X* is then used as training data and fed back to the LSTM algorithm.

[0205] Step 3: Issuance of hierarchical instructions

[0206] During the global coarse adjustment phase, OA gain adjustment commands were issued in batches via the NETCONF protocol, increasing the OA gain of station A from 18dB to 19.2dB (temperature compensation formula: 0.2dB / km×100km×(35-23) / 10). During the local fine adjustment phase, commands were issued directionally to station R2, adjusting VOA_λ1 from 10dB to 9dB, while simultaneously increasing the power of OTU_λ1 from +14dBm to +14.5dBm.

[0207] Step 4: Closed-loop feedback verification

[0208] The monitoring indicators showed changes (within 5 minutes) that the OSNR of λ1 increased from 20.5dB to 21dB, the BER of λ1 recovered from 5e-13 to 1e-13, and the flatness improved from 2.1dB to 0.8dB. If the BER did not meet the target, the genetic algorithm was triggered to iterate again; if the temperature continued to rise, the LSTM initiated a secondary warning.

[0209] Step 5: Iterative Knowledge Storage

[0210] The historical database was updated by inserting the following data: VALUES (CURRENT_TIMESTAMP, 'Temperature Compensation', 35,14.5, 9, 21, 1e-13). The data structure includes timestamp (optimization operation time point, ISO 8601 format), optimization_type (trigger optimization scenario type, such as temperature compensation), temperature (ambient temperature, °C), adjusted_power (OTU adjusted power, dBm), adjusted_voa (VOA adjusted decay value, dB), achieved_osnr (optimized OSNR, dB), and achieved_ber (optimized BER, scientific notation). Simultaneously, the LSTM model underwent incremental training, incorporating the temperature-decay correlation data; a genetic algorithm optimized the solution space, recording high-quality solutions to accelerate convergence.

[0211] Through the above steps, the system achieves automatic parameter optimization in scenarios with rising temperatures, effectively addressing the impact of environmental changes on the performance of the ROADM system and ensuring stable system operation and transmission quality.

[0212] In addition, this application also provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described above when running.

[0213] A parameter tuning apparatus for a ROADM system includes a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the methods described above. This apparatus can be implemented based on an FPGA.

[0214] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term "computer storage medium" includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

Claims

1. A parameter tuning method for a ROADM system, characterized in that, include: Step 30: At the current time t, use a deep learning model to obtain the predicted OSNR values ​​for each wavelength at time T1+t; Step 40: Based on the predicted OSNR value for each wavelength, perform the following operations for each wavelength: For the current wavelength, if the predicted OSNR value for the current wavelength is less than the lower limit of OSNR, and if the OTU power for the current wavelength is less than the upper limit of OTU power, then at least one of the constraints of the genetic algorithm and the initial solution generation conditions should be adjusted. Step 50: Based on the operation results of step 40, perform the iterative operation of the genetic algorithm at time n*T2+t. When the optimal solution is obtained in the nth iteration, perform parameter tuning operation based on the optimal solution obtained in the nth iteration. Where T1 = N * T2, and n = 1, 2, 3, ..., N, where N is a positive integer; In step 30, the deep learning model satisfies at least one of condition 1 and condition 2, wherein: Condition 1 is that the input layer of the deep learning model sets the weight value of the transmit power based on the first correspondence, wherein the first correspondence records the change in OSNR of a single wavelength for every 1dB change in the transmit power of that wavelength. Condition 2 is that the hidden layer of the deep learning model is set with an activation function based on the third correspondence, wherein the third correspondence records the change in multi-wavelength gain uniformity for every 0.1dB / nm change in the OA gain slope; In step 40, the genetic algorithm is adjusted by performing at least one of operations 1 and 2, wherein: Operation 1 involves increasing the OTU transmit power at the current wavelength if the current power is less than the upper limit of the OTU power, as a condition for generating the initial solution for the current wavelength in the genetic algorithm; otherwise, adjusting the VOA attenuation value of the adjacent stations of the current wavelength station, as a range for generating the initial solution for the current wavelength among the adjacent stations in the genetic algorithm. Operation 2 involves raising the lower limit value in the ONSR constraint conditions.

2. The method according to claim 1, characterized in that, Step 50 includes: Step 51: At time n*T2+t, obtain the initial solution of the nth iteration based on the boundary of the constraint solution space of the nth iteration determined by the third correspondence. Step 52: Obtain the value of the objective function for the nth iteration, and obtain the fitness of the nth iteration based on the value of the objective function, wherein the objective function includes the weight value of the single-wave power penalty term determined based on the first correspondence. Step 53: Under the constraints of the nth iteration, based on the fitness of the nth iteration, perform genetic computation on the initial solution of the nth iteration until a preset termination condition is met, wherein the genetic computation includes mutation operation based on the second correspondence. Step 54: When the optimal solution is obtained in the nth iteration, perform parameter tuning based on the optimal solution obtained in the nth iteration. Step 55: Determine whether n is greater than or equal to N, and if n is less than N, execute step 51; The first correspondence records the change in OSNR for a single wavelength when the transmit power changes by 1 dB; the second correspondence records the change in power difference between adjacent wavelength channels at any station when the VOA attenuation value changes by 1 dB; and the third correspondence records the change in multi-wavelength gain uniformity when the OA gain slope changes by 0.1 dB / nm.

3. The method according to claim 1 or 2, characterized in that, The method further includes: At the start of each iteration of the genetic algorithm, the risk value of the four-wave mixing is calculated, and if the risk value is greater than a preset risk threshold, at least one of operations 3, 4, and 5 is executed, wherein: Operation 3 involves lowering the upper limit value of the total power constraint in the genetic algorithm; Operation 4 involves adjusting the weight of the deviation between the total power and the optimal total power in the objective function of the genetic algorithm when the effective operating bandwidth of the optical amplifier meets the preset narrow bandwidth condition or when the optical fiber used is an optical fiber including G.657 fiber. Operation 5 involves increasing the weight of the single-wave power penalty term in the objective function of the genetic algorithm.

4. The method according to claim 3, characterized in that, The methods for obtaining the risk value include: The risk value is obtained by multiplying the square of the actual total power value by the total number of current wavelength channels, and then calculating the ratio of the product to the effective operating bandwidth of the optical amplifier.

5. The method according to claim 1, characterized in that: In step 30, the predicted value of the total power is obtained using the deep learning model; In step 40, if the predicted value of the total power is greater than the total power threshold, the upper limit value of the constraint condition of the total power in the genetic algorithm is lowered.

6. The method according to claim 1, characterized in that: At least one of the fourth and fifth correspondences is pre-recorded, wherein: The fourth correspondence records the change in power attenuation per unit length of optical fiber after a preset time period for every 1°C increase in temperature; the hidden layer of the deep learning model is deployed with the fourth correspondence, and in step 30, the deep learning model is used to obtain the total power attenuation value; in step 40, the total power attenuation value is used to determine the constraints on the total power in the genetic algorithm. The fifth correspondence records the change in OA gain attenuation for each additional month of equipment operation time; wherein, in step 50, the constraints on OA gain in the genetic algorithm are determined based on the fifth correspondence.

7. The method according to claim 6, characterized in that, The method further includes: Step 10: Obtain the operating data of the ROADM system from time t-T1 to the current time t, where each set of operating data includes environmental conditions, aging coefficient, optimal solution and actual performance; Step 20: Perform incremental training on the deep learning model based on the running data, and then execute step 30.

8. The method according to claim 1, characterized in that, The method includes: Based on real-time data, the target wavelength with a single-wavelength power deviation greater than the power deviation threshold is determined, and a VOA adjustment command is issued to the station corresponding to the target wavelength.

9. The method according to claim 1, characterized in that, The objective function of the genetic algorithm at time n*T2+t is determined based on at least one of the following: The deviation between the actual bit error rate at time n*T2+t and the preset bit error rate threshold; The deviation between the variance of OSNR at each wavelength and the average value of the sum of OSNR at time n*T2+t; The prediction error of OSNR at time n*T2+t; wherein the prediction error is the ratio of the absolute value of the difference between OSNR1 and OSNR2 to the time difference, where OSNR1 represents the actual value of OSNR at time n*T2+t, OSNR2 represents the predicted value of OSNR at time T1+t, and the time difference is T1-n*T2.

10. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method described in any one of claims 1 to 9 when it is run.

11. A parameter tuning device for a ROADM system, 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 as described in any one of claims 1 to 9.

Images