A multi-objective optimization-based core resource intelligent scheduling method and system

Abstract

This invention discloses a multi-objective optimization-based intelligent scheduling method and system for fiber core resources, relating to the field of data processing technology. The method includes: real-time acquisition of OSNR, PMD, and XTALK damage parameters of each fiber core and storage as a historical sequence; prediction of the health of each fiber core using an LSTM-Attention neural network model based on the historical sequence, the health being obtained by fusing the logarithmic gain term of OSNR, the square root attenuation term of PMD, and the exponential attenuation term of XTALK; establishment of a multi-objective optimization function including resource utilization, energy consumption, blocking rate, and service quality risk; constrained iterative solution using the NSGA-III algorithm, in which wavelength continuity genetic operations are sequentially performed to generate offspring schemes and digital twin simulations are used to verify the feasibility of the predicted schemes, outputting a Pareto optimal solution set; training a graph neural network with this solution set, and outputting a scheduling scheme through forward inference in the online stage.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for intelligent scheduling of fiber core resources based on multi-objective optimization. Background Technology

[0002] With the explosive growth of data traffic, multi-core optical fibers have become an important development direction for next-generation optical networks due to their high space utilization and large-capacity transmission capabilities. Fiber core resource scheduling, which aims to allocate appropriate fiber cores and wavelength resources according to service needs, is a key technology for multi-core optical fiber networks.

[0003] Existing fiber core resource scheduling technologies primarily employ multi-objective optimization and heuristic algorithms. In terms of optimization modeling, existing technologies use resource utilization, energy consumption, and congestion rate as optimization objectives to construct multi-objective optimization models. Regarding constraint handling, existing technologies add physical layer constraints as penalty terms to the objective function or use the degree of constraint violation to determine feasibility. For solution algorithms, existing technologies use heuristic algorithms such as genetic algorithms and particle swarm optimization for iterative solutions. In terms of real-time scheduling, existing technologies use pre-computed routing tables or simple greedy strategies to respond to service requests.

[0004] However, existing technologies have at least the following shortcomings: First, multi-objective optimization models mainly focus on resource-level indicators and lack awareness of physical layer damage parameters such as optical signal-to-noise ratio, polarization mode dispersion, and crosstalk between fiber cores, which may cause scheduling schemes to fail due to physical layer performance degradation.

[0005] Secondly, when physical layer constraints are treated as penalty terms, the penalty coefficient is difficult to set. If it is too small, the constraints are easily violated; if it is too large, the search space is excessively compressed, making it difficult to achieve a balance between feasibility and optimization performance.

[0006] Third, standard genetic operators are prone to generating invalid solutions with discontinuous wavelengths after crossover and mutation operations, requiring additional repair operations and reducing search efficiency.

[0007] Fourth, pre-computed routing tables are difficult to adapt to dynamic changes in network conditions, and the iterative solution of heuristic algorithms is time-consuming, making it difficult to meet the real-time response requirements of business requests.

[0008] Therefore, there is an urgent need for a smart scheduling technology for fiber core resources that can achieve physical layer damage perception, hierarchical processing of constraints and targets, efficient search, and fast reasoning. Summary of the Invention

[0009] In view of the shortcomings of the prior art, the present invention provides a method and system for intelligent scheduling of fiber core resources based on multi-objective optimization, in order to solve the technical problems of lack of physical layer damage perception, improper constraint processing, low search efficiency and insufficient real-time performance in the prior art multi-objective optimization method.

[0010] In a first aspect, the present invention provides a method for intelligent scheduling of fiber core resources based on multi-objective optimization, comprising: S1: Damage parameter acquisition: Real-time acquisition of damage parameters of each fiber core and storage as a historical sequence. The damage parameters include optical signal-to-noise ratio (OSNR), polarization mode dispersion (PMD), and inter-fiber crosstalk coefficient (XTALK). S2: Health Prediction: Based on the historical sequence, the health of each fiber core is predicted using an LSTM-Attention neural network model; the health is a comprehensive evaluation value characterizing the transmission quality of the fiber core, obtained by fusing OSNR, PMD, and XTALK. S3: Multi-objective optimization modeling: Establish a multi-objective optimization function, with optimization objectives including maximizing fiber core resource utilization, minimizing energy consumption, minimizing service blockage rate, and minimizing service quality risk; the service quality risk is calculated based on the health status. S4: Constraint-Dominated Iterative Solution: The NSGA-III algorithm is used to iteratively solve the multi-objective optimization function and generate candidate scheduling schemes. During the iteration process, steps S5 and S6 are executed sequentially. The OSNR of the fiber cores involved in the candidate scheduling schemes is not lower than a first preset threshold as a constraint condition. The selection is made using constraint dominance relations. The constraint dominance relations mean that the schemes that do not meet the constraints are dominated by the schemes that meet the constraints. The schemes that all meet the constraints are dominated by Pareto dominance relations. After the iteration is completed, the Pareto optimal solution set is output. S5: Wavelength continuity genetic operation: Crossover and mutation operations are performed using genetic operators that maintain wavelength continuity; wavelength continuity means that the service uses the same wavelength on its end-to-end transmission path; S6: Digital twin simulation verification: Construct a digital twin of the physical state of the fiber core to perform physical simulation of the candidate scheduling scheme, and predict whether the OSNR of the fiber core involved in the scheme meets the constraints. S7: Graph Neural Network Inference Output: The graph neural network is trained using the Pareto optimal solution set as training data. The trained graph neural network is used to score business requests and output a scheduling scheme.

[0011] Secondly, the present invention provides a fiber core resource intelligent scheduling system based on multi-objective optimization, comprising: a processor and a memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the intelligent scheduling method for fiber core resources based on multi-objective optimization as described in the first aspect.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) By integrating the three damage parameters OSNR, PMD and XTALK into a health index and using the LSTM-Attention neural network model for prediction, the transmission quality of the fiber core can be comprehensively evaluated and its changing trend can be predicted, providing a basis for scheduling decisions. (2) By incorporating health into the business quality risk objective in the multi-objective optimization function, the optimization process has a tendency to actively avoid low-health fiber cores, rather than only processing them when constraints are violated. (3) By using the constraint dominance relationship for selection, the scheme that meets the constraint conditions has absolute priority in the selection operation, thus avoiding the problem of difficulty in setting the penalty coefficient; (4) By performing wavelength continuity genetic operations to generate offspring schemes first and then performing digital twin simulation verification during the iteration process, the feasibility of the scheme can be predicted before the actual implementation of the scheme, thus avoiding infeasible schemes from entering the subsequent process; (5) By training the graph neural network with the Pareto optimal solution set, only forward inference needs to be performed in the online stage to output the scheduling scheme, which significantly reduces the time complexity of online scheduling. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating an intelligent scheduling method for fiber core resources based on multi-objective optimization, provided in an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of a fiber core resource intelligent scheduling system based on multi-objective optimization, provided as an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0016] Reference manual attached Figure 1 The diagram illustrates a flowchart of an intelligent scheduling method for fiber core resources based on multi-objective optimization, provided by an embodiment of the present invention.

[0017] This invention provides a method for intelligent scheduling of fiber core resources based on multi-objective optimization, which may include the following steps: S1: Damage parameter acquisition: Real-time acquisition of damage parameters of each fiber core and storage as a historical sequence. The damage parameters are optical signal-to-noise ratio (OSNR), polarization mode dispersion (PMD), and inter-core crosstalk coefficient (XTALK).

[0018] Specifically, the multi-core fiber optic network comprises J fiber cores, and the damage parameters of each fiber core are acquired in real time using optical performance monitoring equipment. For fiber core j, the acquired damage parameters include: optical signal-to-noise ratio. Polarization mode dispersion and inter-core crosstalk coefficient Where j represents the total number of fiber cores in the network and j=1,2,...,J, t represents the time step index and t=1,2,...,T, T represents the length of the historical sequence. This represents the OSNR of fiber core j acquired at time step t. The unit of OSNR is dB. This represents the PMD of fiber core j acquired at time step t. The unit of PMD is ps. XTALK represents the XTALK of fiber core j acquired at time step t. XTALK is a dimensionless coefficient with a value range of 0 to 1.

[0019] Furthermore, the damage parameters collected from each fiber core over T consecutive time steps are stored as a historical sequence for subsequent health prediction.

[0020] In this embodiment of the invention, the damage parameters are organized as a time series containing T time steps rather than a single moment's sample value, so that the input of the subsequent prediction model contains information about the evolution of the damage parameters over time, thereby enabling prediction based on historical trends.

[0021] S2: Health Prediction: Based on historical data, the health of each fiber core is predicted using an LSTM-Attention neural network model. The health score is a comprehensive evaluation value characterizing the transmission quality of the fiber core, obtained by fusing OSNR, PMD, and XTALK.

[0022] In one possible implementation, the health score is obtained by weighted fusion of a logarithmic gain term based on OSNR, a square root decay term based on PMD, and an exponential decay term based on XTALK. The formula for calculating the health score is shown in formula (1): in, This indicates the health status of fiber core j, with a value ranging from 0 to 1, where j represents the index of all fiber cores in the network, and j = 1, 2, ..., J. This represents the OSNR of fiber core j, in dB. , This represents the PMD of fiber core j, with the unit being ps. , Let XTALK represent the core j, which is dimensionless and ranges from 0 to 1. Indicates the OSNR reference value and the unit is dB. , This represents the upper limit of OSNR normalization, in dB. , The coefficient of the square root attenuation term is expressed in units of 1 / 2. and , The coefficient of the exponentially decaying term is dimensionless and , Represents the natural logarithm function. This represents an exponential function.

[0023] It should be noted that in formula (1) , , The value is the current or future time value predicted based on historical sequences, or the latest time value from the historical sequence. , , The value, This represents the OSNR of fiber core j acquired at time step T. This indicates the PMD of fiber core j acquired at time step T. This indicates the XTALK of fiber core j acquired at time step T.

[0024] In this embodiment of the invention, formula (1) uses a fusion method of multiplying the logarithmic gain term, the square root decay term, and the exponential decay term, which is different from the method of linearly weighted summing of each damage parameter. When any damage parameter deteriorates, the decrease in the value of the corresponding term will lead to an overall decrease in health, thus avoiding the situation where the abnormality of a single parameter is masked by the normal value of other parameters.

[0025] Optionally, The value is 20dB. The value is 40dB. The value is 0.1 , The value is 2.0. It is understood that those skilled in the art can adjust the value of the above parameter according to the actual situation, and the embodiments of the present invention do not specifically limit this.

[0026] In one possible implementation, the LSTM-Attention neural network model includes an input layer, an LSTM layer, a multi-head self-attention layer, and an output layer. The input layer receives the historical sequence, the LSTM layer extracts time-dependent features, the multi-head self-attention layer learns the contribution weights of each time step to the prediction result, and the output layer outputs the health score.

[0027] Furthermore, the LSTM layer processes the input historical sequence, selectively retaining or forgetting historical information through the combined action of the forget gate, input gate, and output gate, outputting the hidden state vector at each time step. Where t represents the time step index and t = 1, 2, ..., T, Represents the time step t. 3D real vector This represents the dimension of the hidden state vector.

[0028] Optionally, The value is 128. It is understood that those skilled in the art can adjust the value according to the actual situation. The value of is adjusted, but the embodiments of the present invention do not impose specific limitations on this.

[0029] Furthermore, the multi-head self-attention layer uses the hidden state sequence output by the LSTM layer as input for attention weighting. This involves processing the hidden states from T time steps. Select the K most recent time steps. Participating in attention calculation, among which, Within the multi-head self-attention layer, the selected K time steps are renumbered, with indices k and k=1,2,...,K, where k=K corresponds to the T-th time step in the original sequence. The formula for calculating the attention weights is shown in formula (2): in, Denotes the attention weight at the k-th time step and and k represents the time step index involved in the attention calculation, and k = 1, 2, ..., K, where K represents the number of time steps involved in the attention calculation. This represents the hidden state vector of the last time step involved in the computation, i.e., the Kth time step, and is... 3D real vector Let the hidden state vector at time step k be... 3D real vector Denotes the attention weight matrix and is OK A real matrix of columns, The scaling factor is used to stabilize the gradient. For the summation index independent of k and traversing all time steps from 1 to K, This represents the transpose of a vector. This represents an exponential function.

[0030] Furthermore, the output layer maps the attention-weighted feature vectors to health prediction values ​​and outputs the health of each fiber core.

[0031] In this embodiment of the invention, the multi-head self-attention layer enables the contribution weight of each historical moment to the prediction result to be dynamically learned by the model based on the input data. This differs from the method of using fixed equal weights or preset exponential decay weights for all historical moments, and can adaptively determine the importance of each moment for different input sequences.

[0032] S3: Multi-objective optimization modeling: Establish a multi-objective optimization function, with optimization objectives including maximizing fiber core resource utilization, minimizing energy consumption, minimizing service blockage rate, and minimizing service quality risk. Service quality risk is calculated based on health status.

[0033] In one possible implementation, the multi-objective optimization function contains four optimization objectives. The calculation formula for the multi-objective optimization function is shown in formula (3): Where min represents the minimization operation. This represents a multi-objective optimization function and is a 4-dimensional real vector. Indicates the candidate scheduling scheme, Indicates the utilization rate of fiber core resources. Indicates energy consumption in watts (W). Indicates the business blocking rate. This indicates business quality risk; the negative sign indicates that the goal of maximizing is being transformed into the goal of minimizing.

[0034] Furthermore, fiber core resource utilization rate Energy consumption and business blocking rate All are candidate scheduling schemes Post-execution network status metrics. Fiber core resource utilization rate. Defined as the ratio of the resource usage of all fiber cores to the total capacity. Where j represents the entire network fiber core index and j=1,2,...,J, This represents the resource utilization rate of fiber core j after the candidate scheduling scheme is executed, and its value ranges from 0 to 1. This represents the total capacity of fiber core j, expressed in Gbps. Greater than 0. Energy consumption Defined as the sum of transmission energy consumption of all fiber cores. ,in, This represents the transmission energy consumption of fiber core j after the candidate scheduling scheme is executed, and the unit is W. Business blocking rate Defined as the ratio of blocked service requests to the total number of service requests. Where B represents the number of blocked services and is a non-negative integer, and W represents the total number of service requests and is a positive integer.

[0035] Furthermore, candidate scheduling schemes It involves M fiber cores, where M represents the number of fiber cores involved in the candidate scheduling scheme and To distinguish between the fiber cores involved in the entire network and those involved in the candidate scheduling schemes, a fiber core index s is introduced, where s = 1, 2, ..., M. The s-th fiber core involved in the scheme corresponds to a specific fiber core in the entire network. The health of this fiber core is denoted as... The OSNR value is denoted as Resource utilization rate is recorded as Transmission energy consumption is denoted as All values ​​are taken from the corresponding values ​​of the fiber core in the entire network fiber core parameters.

[0036] In one possible implementation, business quality risk is calculated based on a weighted average of health level and business priority; the lower the health level or the higher the business priority, the greater the business quality risk. The calculation formula is shown in formula (4): in, M represents the business quality risk, and M represents the candidate scheduling scheme. The number of fiber cores involved, s represents the fiber core index involved in the scheme and s=1,2,...,M. This indicates that the s-th root scheme involves the health of the fiber core and its value ranges from 0 to 1. This indicates the normalized priority of the business involving the fiber core using the s-th root scheme, and its value ranges from 0 to 1. .

[0037] In this embodiment of the invention, formula (4) uses the complement of health. With business priority The summation after multiplication is incorporated into the optimization objective, allowing health status to participate in both constraint determination and objective optimization simultaneously, unlike methods that only consider physical layer state as a constraint. By integrating health status into the optimization objective, the optimization process proactively avoids low-health fiber cores, rather than only addressing them when constraints are violated.

[0038] S4: Constraint-Dominated Iterative Solution: The NSGA-III algorithm is used to iteratively solve the multi-objective optimization function and generate candidate scheduling schemes. Steps S5 and S6 are executed sequentially during the iteration. The constraint condition is that the OSNR of all fiber cores involved in the candidate scheduling schemes is not lower than a first preset threshold. A constraint dominance relationship is used for selection, meaning that schemes that do not meet the constraints are dominated by schemes that do. Schemes that all meet the constraints are subject to Pareto dominance. After iteration, the Pareto optimal solution set is output.

[0039] It should be noted that those skilled in the art can set the size of the first preset threshold according to actual needs, and the present invention does not limit this.

[0040] In one possible implementation, the constraint condition is that the OSNR of all fiber cores involved in the candidate scheduling scheme is not lower than a first preset threshold, and the constraint condition is expressed as follows: This holds true for all schemes involving fiber core s, where s represents the fiber core index involved in the scheme and s = 1, 2, ..., M. This indicates that the OSNR of the s-th scheme involves the fiber core and is expressed in dB. This represents the OSNR threshold, and the unit is dB.

[0041] Furthermore, the constraint violation degree is used to measure the extent to which a candidate scheduling scheme violates the constraints. The formula for calculating the constraint violation degree is shown in formula (5): in, Indicate candidate scheduling schemes The constraint violation degree is expressed in dB, M represents the number of fiber cores involved in the candidate scheduling scheme, and s represents the fiber core index involved in the scheme, where s = 1, 2, ..., M. This indicates that the OSNR of the s-th scheme involves the fiber core and is expressed in dB. This represents the OSNR threshold, expressed in dB. This indicates taking the larger of the two values. When... When the candidate scheduling scheme satisfies the constraints. At that time, the candidate scheduling scheme does not meet the constraints.

[0042] Optionally, The value is set at 15 dB. It is understood that those skilled in the art can adjust the value based on the specific circumstances. The value of is adjusted, but the embodiments of the present invention do not impose specific limitations on this.

[0043] Furthermore, constraint dominance refers to the situation where a scheme that does not meet the constraints is dominated by a scheme that does meet the constraints. For two candidate scheduling schemes... and ,plan Constraint Domination Scheme The scheme is proposed if and only if one of the following conditions is met: The constraints are satisfied and the solution is The constraints are not met. Alternatively, the solution... With the plan All constraints are met and the scheme Domination schemes under Pareto domination .

[0044] In one possible implementation, S4 specifically includes sub-steps S401 and S402: S401: Project the reference point of the NSGA-III algorithm onto the boundary of the feasible region that satisfies the constraints.

[0045] S402: When selecting the environment, each candidate scheduling scheme is associated with the nearest reference point on the feasible domain boundary. For schemes that are associated with the same reference point, all meet the constraints, and do not dominate each other, the variance of the fiber cores involved in each scheme in terms of health is calculated, and the scheme with the smallest variance is selected first.

[0046] Specifically, for candidate scheduling schemes For the M fiber cores involved, first calculate the average health value. Then calculate the variance. Where s represents the scheme involving fiber core indices and s = 1, 2, ..., M, This indicates that the s-th root scheme involves the health of the fiber core. This represents the average health level. This represents the variance of fiber health. A smaller variance indicates a more uniform distribution of health across all fiber cores, preventing some fiber cores from becoming bottlenecks due to excessively low health.

[0047] Furthermore, in each generation of population update in the NSGA-III algorithm, step S5 is first executed to perform genetic operations on the current population to generate candidate scheduling schemes for offspring. Then, step S6 is executed to perform physical simulation verification on each candidate scheduling scheme for offspring and update the constraint violation degree according to the verification results, until the preset iteration termination condition is reached and the Pareto optimal solution set is output.

[0048] In this embodiment of the invention, the constraint dominance relationship places candidate scheduling schemes that satisfy the constraints and those that do not at different dominance levels. Schemes that satisfy the constraints have absolute priority in the selection operation, which is different from the method of adding constraint violation as a penalty term to the objective function. The latter requires setting a penalty coefficient; if the penalty coefficient is too small, the constraints are easily violated, and if the penalty coefficient is too large, the search space is excessively compressed.

[0049] S5: Wavelength Continuity Genetic Operation: Crossover and mutation operations are performed using genetic operators that maintain wavelength continuity. Wavelength continuity means that services use the same wavelength along their end-to-end transmission path.

[0050] In one possible implementation, the genetic operators include a continuous crossover operator, a phase mutation operator, and a connectivity repair operator.

[0051] Furthermore, the continuous crossover operator is used to exchange fiber core allocation segments within path segments using the same wavelength, ensuring the wavelength continuity of the offspring scheme after the crossover operation.

[0052] Furthermore, the phase variation operator is used to apply an offset to the wavelength allocation and synchronously adjust the entire path to ensure the wavelength continuity of the scheme after the variation operation.

[0053] Furthermore, the connectivity repair operator is used to repair schemes that do not meet wavelength continuity by searching for and using resources with the same wavelength in adjacent fiber cores.

[0054] In this embodiment of the invention, the wavelength continuity constraint is embedded in the operational logic of the genetic operator. The offspring schemes satisfy the wavelength continuity constraint during generation, unlike the standard genetic operator method which generates schemes and then checks and corrects for constraint compliance. The latter generates a large number of invalid schemes that do not meet the constraints in high-constraint-density scenarios, requiring additional computational resources for repair or discarding.

[0055] S6: Digital Twin Simulation Verification: Construct a digital twin of the physical state of the fiber core to perform physical simulation on the candidate scheduling scheme and predict whether the OSNR of the fiber core involved in the scheme meets the constraints.

[0056] In one possible implementation, the digital twin is updated according to the damage parameters at a preset period and noise reduction is performed using Kalman filtering.

[0057] It should be noted that those skilled in the art can set the size of the preset cycle according to actual needs, and this invention does not limit this.

[0058] Furthermore, physical simulations are used to predict the OSNR change, PMD drift, and XTALK increment of the fiber cores involved in the candidate scheduling schemes.

[0059] The formula for calculating the change in OSNR is shown in formula (6): in, This indicates the decrease in OSNR of the fiber core involved in the s-th scheme, expressed in dB. s indicates that the scheme involves fiber core indices and s = 1, 2, ..., M. Let r represent the set of adjacent fiber cores involved in the s-th root scheme, and let r represent the index of the adjacent fiber core. Furthermore, adjacent fiber cores are part of the entire network of fiber cores. This indicates that the s-th root scheme involves the crosstalk coupling coefficient between the fiber core and the r-th adjacent fiber core, and , This represents the resource utilization rate of the r-th adjacent fiber core, taken from the fiber core parameters of the entire network. The value corresponding to j=r, and its range is from 0 to 1. This indicates that the s-th root scheme involves the gain coefficient of the fiber core and , This represents the resource utilization rate of the fiber core involved in the s-th root scheme, with a value ranging from 0 to 1. Represents the logarithmic function with base 10. This indicates a series of multiplication operations.

[0060] In this embodiment of the invention, formula (6) uses a multiplicative form to model the cumulative effect of crosstalk among multiple adjacent fiber cores. The crosstalk effects of each adjacent fiber core act on the signal-to-noise ratio of the target fiber core in a multiplicative manner, which is different from the method of linearly superimposing the crosstalk effects of each adjacent fiber core in an additive manner. The multiplicative form can reflect the cascade amplification characteristics of crosstalk.

[0061] The formula for calculating the PMD drift is shown in formula (7): in, This indicates the PMD drift of the fiber core involved in the s-th scheme, in ps. s indicates that the scheme involves fiber core indices and s = 1, 2, ..., M. This represents the PMD drift coefficient of the fiber core involved in the s-th root scheme, in ps. , This represents the transmission energy consumption of the fiber core involved in the s-th root scheme, expressed in W. , This represents a reference value for energy consumption, and the unit is W. .

[0062] It should be noted that the transmission energy consumption in formula (7) With PMD drift The physical connection is that increased transmission energy consumption leads to an increase in the internal temperature of the optical fiber, and the temperature change causes a change in the stress distribution of the optical fiber material, which in turn leads to a drift in polarization mode dispersion.

[0063] Optionally, The value is 0.5ps. The value is 1W. It is understood that those skilled in the art can adjust the value of the above parameters according to the actual situation, and the embodiments of the present invention do not specifically limit this.

[0064] The formula for calculating the XTALK increment is shown in formula (8): in, This indicates that the s-th root scheme involves the XTALK increment of the fiber core and s indicates that the scheme involves fiber core indices and s = 1, 2, ..., M. Let r represent the set of adjacent fiber cores involved in the s-th root scheme, and let r represent the index of the adjacent fiber core. , This represents the crosstalk growth factor of the s-th root scheme involving the influence of the r-th adjacent root on the fiber core, and... , It represents the resource utilization rate of the r-th adjacent fiber core, and its value ranges from 0 to 1, and the value comes from the same source as formula (6).

[0065] Optionally, The value is 0.01. It is understood that those skilled in the art can adjust the value according to the actual situation. The value of is adjusted, but the embodiments of the present invention do not impose specific limitations on this.

[0066] Furthermore, the OSNR value of the fiber core is updated based on the OSNR decrease obtained from the simulation. ,in, This represents the updated OSNR value after simulation, used to determine whether the candidate scheduling scheme satisfies the constraints. If there exists any fiber core such that... If the candidate scheduling scheme is not feasible, then mark it as an infeasible scheme.

[0067] In this embodiment of the invention, physical simulation pre-verifies candidate scheduling schemes during the iterative solution phase. This allows for the determination of whether a scheme will violate constraints after execution, unlike methods that execute the scheme first and then check for constraint violations. Pre-verification eliminates infeasible schemes during the iterative phase, preventing service interruptions due to physical layer performance degradation after actual execution.

[0068] S7: Graph Neural Network Inference Output: The graph neural network is trained using training data constructed from the Pareto optimal solution set. The trained graph neural network is then used to score business requests and output a scheduling scheme.

[0069] Specifically, based on the health of each fiber core and damage parameters , , Using the comprehensive scores of each candidate scheduling scheme in the Pareto optimal solution set as the input features of the graph neural network, a training dataset is constructed to supervise the training of the graph neural network.

[0070] In one possible implementation, the graph neural network models each fiber core and its connections as a graph structure, where nodes represent optical cross-connect devices and edges represent fiber links. Edge features include health and XTALK. Health is used as the aggregation weight when aggregating neighbor features.

[0071] Furthermore, the graph neural network employs a message-passing mechanism for feature aggregation. The calculation formula for node feature update is shown in formula (9): in, Let v be the feature vector of node v in the (n+1)th layer and be 3D real vector Let v represent the dimension of the graph neural network feature vector, v represent the graph node index, n represent the graph neural network layer index and n = 0, 1, ..., R-1, and R represent the number of layers in the graph neural network. Let u be the feature vector of the nth layer node and be 3D real vector Let v represent the set of neighboring nodes of node v, and u represent the index of a neighboring node. , This represents the health of the fiber core corresponding to the edge between node v and node u, and is taken from the overall network fiber core health calculated by formula (1). The corresponding value in the table, and the value range is from 0 to 1. Denotes the smoothing constant and This is used to avoid the denominator being zero and to ensure that subsequent formulas use this value. Let n be the weight matrix of the nth layer and be ... OK A real matrix of columns, For the summation index independent of u, and traversing all neighboring nodes of node v, This represents the ReLU activation function.

[0072] In this embodiment of the invention, formula (9) uses health status As a weighting coefficient for neighbor feature aggregation, neighbor nodes with higher health contribute more information to the current node's feature updates during message passing, unlike the method of equal-weighted aggregation for all neighbor nodes. After multiple layers of message passing, the node features implicitly contain preference information propagating along the high-health path.

[0073] Optionally, R can be set to 3. The value is 64. It is understood that those skilled in the art can adjust the value of the above parameter according to the actual situation, and the embodiments of the present invention do not specifically limit this.

[0074] Furthermore, after message passing through the R layer, the feature vectors of all nodes involved in the candidate scheduling scheme are summed and pooled, and then mapped to the predicted score of the scheme through a fully connected layer. During the training phase, the multi-objective comprehensive score of each candidate scheduling scheme in the Pareto optimal solution set is calculated using formula (10) as the training label. During the inference phase, the graph neural network directly outputs the predicted score. The formula for calculating the multi-objective comprehensive score is shown in formula (10): in, Let represent the overall score of candidate scheduling scheme 'a', where 'a' represents the index of the candidate scheduling scheme, Q represents the number of optimization objectives (Q=4), and q represents the index of the optimization objective (q=1,2,...,Q). Let the weight coefficients of the q-th optimization objective be... and , The value of candidate scheduling scheme a on the qth optimization objective is represented by formula (3). exist The value is for scheme a. Let represent the maximum value of the q-th optimization objective in the Pareto optimal solution set. Let represent the minimum value of the q-th optimization objective in the Pareto optimal solution set. The smoothing constant is used to avoid the denominator being zero.

[0075] Optionally, All values ​​are set to 0.25. It is understood that those skilled in the art can adjust the values ​​of the weighting coefficients according to the actual situation, and the embodiments of the present invention do not impose specific limitations on this.

[0076] In one possible implementation, S7 specifically includes sub-steps S701 and S702: S701: Determine the complexity of a business request.

[0077] S702: If the complexity is lower than the second preset threshold, a graph neural network is used for scoring and a scheduling scheme is output. If the complexity is not lower than the second preset threshold, a scheduling scheme is selected from the Pareto optimal solution set according to the constraints of the business request.

[0078] It should be noted that those skilled in the art can set the size of the second preset threshold according to actual needs, and this invention does not limit this.

[0079] In this embodiment of the invention, the iterative solution process of multi-objective optimization is completed in the offline stage, and the Pareto optimal solution set output by it is used to construct training data to train the graph neural network. In the online stage, only the forward inference of the graph neural network needs to be performed to output the scheduling scheme for the business request. The time complexity of the online stage is reduced from the complexity of iterative solution to the complexity of the forward inference of the neural network.

[0080] Reference manual attached Figure 2 The diagram shows a structural schematic of a fiber core resource intelligent scheduling system based on multi-objective optimization provided by an embodiment of the present invention.

[0081] This invention also provides a multi-objective optimization-based intelligent scheduling system 20 for fiber core resources, comprising: a processor 201 and a memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-mentioned intelligent scheduling method for fiber core resources based on multi-objective optimization and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0082] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0083] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0084] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0085] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0086] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A method for intelligent scheduling of fiber core resources based on multi-objective optimization, characterized in that, include: S1: Real-time acquisition of damage parameters of each fiber core and storage as a historical sequence. The damage parameters include optical signal-to-noise ratio (OSNR), polarization mode dispersion (PMD), and inter-fiber crosstalk coefficient (XTALK). S2: Based on the historical sequence, the health of each fiber core is predicted using an LSTM-Attention neural network model; the health is a comprehensive evaluation value characterizing the transmission quality of the fiber core, obtained by fusing OSNR, PMD, and XTALK. S3: Establish a multi-objective optimization function, the optimization objectives of which include maximizing fiber core resource utilization, minimizing energy consumption, minimizing service blockage rate, and minimizing service quality risk; the service quality risk is calculated based on the health status. S4: The NSGA-III algorithm is used to iteratively solve the multi-objective optimization function and generate candidate scheduling schemes. During the iteration process, steps S5 and S6 are executed sequentially. The OSNR of the fiber cores involved in the candidate scheduling schemes is not lower than a first preset threshold is used as a constraint condition. The selection is based on constraint dominance relationship. The constraint dominance relationship means that the scheme that does not meet the constraint condition is dominated by the scheme that meets the constraint condition. The schemes that all meet the constraint condition are subject to Pareto dominance relationship. After the iteration is completed, the Pareto optimal solution set is output. S5: Use a genetic operator that maintains wavelength continuity to perform crossover and mutation operations; wavelength continuity means that the service uses the same wavelength on its end-to-end transmission path; S6: Construct a digital twin of the physical state of the fiber core to perform physical simulation of the candidate scheduling scheme and predict whether the OSNR of the fiber core involved in the scheme meets the constraints. S7: Use the Pareto optimal solution set to construct training data to train a graph neural network, and use the trained graph neural network to score business requests and output a scheduling scheme.

2. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, The health score is obtained by weighted fusion of the logarithmic gain term based on the OSNR, the square root decay term based on the PMD, and the exponential decay term based on the XTALK.

3. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, The LSTM-Attention neural network model includes an input layer, an LSTM layer, a multi-head self-attention layer, and an output layer. The input layer is used to receive the historical sequence, the LSTM layer is used to extract time-dependent features, the multi-head self-attention layer is used to learn the contribution weights of each time step to the prediction result, and the output layer is used to output the health score.

4. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, The business quality risk is calculated based on a weighted average of the health score and the business priority. The lower the health score or the higher the business priority, the greater the business quality risk.

5. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, S4 specifically includes: S401: Project the reference point of the NSGA-III algorithm onto the boundary of the feasible region that satisfies the constraints. S402: When selecting the environment, each candidate scheme is associated with the nearest reference point on the boundary of the feasible domain. For schemes that are associated with the same reference point, all satisfy the constraints, and do not dominate each other, the mean and variance of the fiber core involved in each scheme in terms of the health are calculated, and the scheme with the smallest variance is selected first.

6. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, The genetic operators include the continuous crossover operator, the phase mutation operator, and the connectivity repair operator.

7. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, The digital twin is updated according to the damage parameters at a preset period and noise reduction is performed using Kalman filtering.

8. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, The graph neural network models each fiber core and its connection relationship as a graph structure, where nodes represent optical cross-connect devices and edges represent optical fiber links; the features of the edges include the health degree and the XTALK; the health degree is used as the aggregation weight when aggregating neighbor features.

9. The intelligent scheduling method for fiber core resources based on multi-objective optimization according to claim 1, characterized in that, Specifically, S7 includes: S701: Determine the complexity of the service request; S702: If the complexity is lower than the second preset threshold, the graph neural network is used to score and output the scheduling scheme; if the complexity is not lower than the second preset threshold, the scheduling scheme is selected from the Pareto optimal solution set according to the constraints of the service request.

10. A fiber core resource intelligent scheduling system based on multi-objective optimization, characterized in that, include: Processor and memory; The memory stores programs or instructions that can run on the processor, and when the program or instructions are executed by the processor, they implement the steps of the intelligent scheduling method for fiber core resources based on multi-objective optimization as described in any one of claims 1 to 9.

Images