Service architecture optimization method based on multi-expert lightweight model

Abstract

The invention discloses a service architecture optimization method based on a multi-expert lightweight model, and the method comprises the following steps: S1, collecting real-time state data, and carrying out the preprocessing of the real-time state data; s2, the multi-expert lightweight model performs flow prediction on the preprocessed data, dynamically optimizes resource allocation, performs joint training and output fusion on the output of each model, and generates a comprehensive decision; s3, generating a dynamic optimization strategy of the service architecture in combination with the comprehensive decision; s4, automatically adjusting the service architecture according to the dynamic optimization strategy; s5, monitoring the performance of the optimized service architecture in real time to adjust the weighting coefficient and parameter of the multi-expert lightweight model; and S6, updating parameters of the model by using incremental learning and an online training method, and adjusting a service architecture. Through combination of the multi-expert lightweight model and the adaptive multi-modal fusion algorithm, the method has the characteristics of high efficiency, intelligence and sustainable optimization.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and edge computing, and in particular to a service architecture optimization method based on a multi-expert lightweight model. Background Art

[0002] With the development of cloud computing, edge computing, and distributed service systems, modern service architectures are characterized by numerous nodes, complex topologies, and frequent dynamic resource changes. To improve system resource utilization and service response efficiency, service architecture optimization has become a key issue in system operation and management. Currently, common service optimization methods often employ static rule configuration, single prediction models, or centralized optimization strategies to address tasks such as resource allocation, load balancing, and traffic forecasting. However, in practice, these methods often suffer from issues such as delayed response, low prediction accuracy, and difficulty adapting to dynamic environmental changes.

[0003] Traditional optimization methods generally rely on a single model structure, failing to address the diverse needs of multiple tasks. Furthermore, existing models are complex and computationally intensive, making them difficult to deploy in lightweight environments like edge devices. Model training and prediction processes are mostly performed offline and statically, lacking continuous learning mechanisms. This makes it difficult for the system to adjust strategies in a timely manner in response to changes in network topology or fluctuations in service requests, thus impacting overall performance.

[0004] When it comes to information fusion, existing technologies often use fixed weights or simple averaging to fuse the outputs of different models, failing to consider the varying importance of each model in different contexts and lacking adaptive adjustment mechanisms. Furthermore, the connections between service nodes are often overlooked, and optimization decisions fail to fully leverage network topology and inter-node data dependencies, leading to problems such as unbalanced resource scheduling and severe local bottlenecks.

[0005] In summary, the existing technologies in service architecture optimization have the following main shortcomings: lack of an expert model system for multi-task collaboration, which cannot support differentiated processing between tasks; single fusion mechanism and lack of dynamic weighting capability; lack of feedback adjustment and online update mechanism in the model, making it difficult to cope with environmental changes; lack of the ability to jointly model network topology and service data, which limits the global optimization effect of the scheduling strategy.

[0006] There is an urgent need for a service architecture optimization method that supports multi-expert model collaboration, adaptive fusion, and feedback-driven updates to improve the system's resource scheduling capabilities and overall operating efficiency in complex dynamic environments. Summary of the Invention

[0007] One purpose of the present invention is to propose a service architecture optimization method based on a multi-expert lightweight model. The present invention fully combines multi-expert modeling, self-attention mechanism, graph neural network, adaptive reinforcement learning and multimodal fusion algorithm, and describes in detail the mechanism for realizing coordinated optimization of traffic prediction, resource scheduling and load balancing in a complex dynamic service environment. It has the advantages of lightweight model structure, high decision-making efficiency, strong adaptability and suitability for edge deployment.

[0008] A service architecture optimization method based on a multi-expert lightweight model according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect the real-time status data of each service node in the service architecture and pre-process the real-time status data;

[0010] S2, a multi-expert lightweight model uses a self-attention mechanism to predict traffic flow on pre-processed data, dynamically optimizes resource allocation using an adaptive reinforcement learning algorithm, and introduces an adaptive multimodal fusion algorithm to jointly train and fuse the outputs of each model to generate a comprehensive decision;

[0011] S3. Combine comprehensive decision-making and generate dynamic optimization strategies for service architecture through multi-dimensional feedback learning mechanisms;

[0012] S4. Automatically adjust resource allocation, service scheduling, network routing, and load balancing of the service architecture based on dynamic optimization strategies.

[0013] S5. Monitor the performance of the optimized service architecture in real time. Based on the real-time monitoring results, adjust the weighting coefficients and parameters of the multi-expert lightweight model through a multi-level adaptive feedback mechanism.

[0014] S6. Use incremental learning and online training methods to update the parameters of the multi-expert lightweight model and adjust the service architecture according to network topology changes and service requirements.

[0015] Optionally, the real-time status data includes performance indicators of each service node: CPU usage, memory occupancy, network bandwidth, network delay, traffic, and load.

[0016] Optionally, the preprocessing step includes denoising, data normalization, and extracting key features including load fluctuations, service node resource utilization, and network traffic patterns from the real-time status data.

[0017] Optionally, the dynamic optimization strategy includes resource allocation strategy of service nodes, network routing adjustment strategy, load balancing strategy and service scheduling strategy. All dynamic optimization strategies make comprehensive decisions based on real-time network status data, historical network traffic patterns, real-time load information and future traffic forecast results.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Input the preprocessed data and analyze and predict different optimization tasks using a multi-expert lightweight model, wherein the multiple multi-expert lightweight models include a traffic prediction model, a load balancing model, and a resource scheduling model;

[0020] S22, the traffic prediction model first calculates the traffic volume based on the historical traffic data X t and network topology data T t After preprocessing, the self-attention mechanism module calculates the weighted matrix A at each moment t , historical traffic data X t and network topology data T t With the weight matrix A t Perform weighted interaction to obtain weighted feature representation Represent the features Input to the LSTM network, after processing by the LSTM network, the output will generate a network traffic prediction value at a future moment

[0021] S23, real-time network load data L t , historical load pattern data L hist and network topology data T t , will be input into the load balancing model, based on deep Q learning and multi-channel feedback mechanism, to learn the real-time network load data L t , historical load pattern data L hist and network topology data T t The complex nonlinear relationship between them; through a multi-channel feedback structure, dynamically adjust the resource allocation decision of service nodes under real-time network load changes; through an adaptive reward mechanism, adjust the load balancing strategy;

[0022] S24. The reward mechanism is designed based on the system's resource utilization, response time, and load fluctuations. The load balancing model dynamically adjusts the resource allocation of service nodes through the complex nonlinear relationship learned, the resource allocation decision of the service node, and the load balancing strategy. The resource allocation strategy R adjusted by the adaptive reinforcement learning algorithm t As an output of load balancing decisions;

[0023] S25. Resource scheduling model uses graph neural network to analyze network topology data T t and traffic forecast results Perform joint analysis: Through graph convolution operations, the resource scheduling model can learn information from the adjacent nodes of each node. In the graph convolution operation, the weight of the node is based on the traffic prediction data. and real-time load information for dynamic adjustment; after multiple graph convolution operations and adaptive adjustments, the final feature representation of each node h v The final feature representation h of all nodes is combined to form the optimized state of the node in the global network. v , the resource scheduling model will generate a globally optimized resource scheduling strategy S opt :

[0024] S26. The input of the adaptive multimodal fusion algorithm is the network traffic prediction value Resource allocation strategy R t and resource scheduling strategy S opt , introduce a feedback learning mechanism to dynamically weight the output results of each multi-expert lightweight model; multiply the output results of each model by its weight coefficient and sum them to obtain the final comprehensive decision S final .

[0025] The present invention provides a service architecture optimization method based on a multi-expert lightweight model. By constructing multiple lightweight models including traffic prediction, load balancing and resource scheduling, multi-source data such as historical and real-time traffic, load, network topology, etc. are modeled and predicted respectively. Among them, the traffic prediction model adopts the self-attention mechanism and LSTM network to jointly estimate future traffic. The load balancing model realizes dynamic resource adjustment based on deep Q learning and multi-channel feedback mechanism. The resource scheduling model uses graph neural network to fuse traffic and load information to generate a global optimal scheduling strategy. Finally, an adaptive multimodal fusion algorithm and feedback learning mechanism are used to dynamically weight and fuse the output results of each model to form a comprehensive optimization decision, which significantly improves the resource utilization and response performance of the service architecture in a complex network environment.

[0026] Optionally, the S3 specifically includes:

[0027] S31. Use a multi-dimensional feedback learning mechanism to optimize the comprehensive decision result S final The multi-dimensional feedback learning mechanism adjusts the weighted coefficients of each multi-expert lightweight model output based on historical feedback and real-time feedback information.

[0028] S32. For each result output by the multi-expert lightweight model, the weighting coefficient will be dynamically adjusted according to the actual performance of the current network environment. Each weighting coefficient w i They will be adjusted in real time according to the reward function in the feedback learning mechanism;

[0029] S33. Dynamic optimization strategy for generating service architecture S dynamic , dynamic optimization strategy S dynamic It includes resource allocation of service nodes, network routing selection, load balancing and service scheduling.

[0030] The present invention further introduces a multi-dimensional feedback learning mechanism to optimize the comprehensive decision-making results. This mechanism combines historical feedback and real-time feedback information to dynamically adjust the weighting coefficients of the output results of each multi-expert lightweight model; by real-time evaluation of the current network environment performance and adaptive weighting adjustment based on the reward function, a dynamic optimization strategy for the service architecture covering resource allocation, network routing selection, load balancing and service scheduling is generated, thereby improving the intelligent response capability and operational efficiency of the overall system.

[0031] Optionally, the S4 specifically includes:

[0032] S41, according to the dynamic optimization strategy S dynamic Adjust resource allocation for each service node in the service architecture, including key resource indicators such as CPU usage, memory usage, network bandwidth, network latency, traffic flow, and load;

[0033] S42, according to the dynamic optimization strategy S dynamic The network routing adjustment strategy in the network is used to optimize the routing path in the network. The network routing adjustment is based on the traffic prediction data Y t+1 and real-time network topology data, using the shortest path algorithm;

[0034] S43. Dynamically adjust the load distribution of the service nodes through an adaptive algorithm according to the load balancing strategy in the dynamic optimization strategy;

[0035] S44, according to the service scheduling strategy S in the dynamic optimization strategy opt ,automatically adjust the task scheduling of service nodes, including task allocation, service start and stop operations;

[0036] S45. Automatically optimize the service architecture based on the resource allocation, network routing, load balancing, and service scheduling strategies adjusted above.

[0037] The present invention comprehensively adjusts the service architecture based on a dynamic optimization strategy, covering the key resource configuration of service nodes (such as CPU, memory, bandwidth, latency, traffic and load), and uses the shortest path algorithm combined with traffic prediction and real-time topology optimization to optimize the network routing path; at the same time, it dynamically adjusts the load distribution of service nodes according to the adaptive load balancing strategy, and automatically manages task allocation and service start and stop operations according to the service scheduling strategy, ultimately achieving the coordinated implementation of resource allocation, routing optimization, load balancing and task scheduling, and improving the overall operating efficiency and intelligence level of the service architecture.

[0038] Optionally, the S5 specifically includes:

[0039] S51, real-time monitoring of the performance indicators of each service node in the optimized service architecture, said performance indicators including CPU usage C i , memory usage M i 、Network bandwidth B i , network delay D i , network traffic T i and system load L i , where i represents the i-th service node;

[0040] S52: Construct a multi-dimensional performance state vector V based on the real-time performance indicators of each service node i =

[0041] [C i ,M i ,B i ,D i ,T i ,L i ], and V i Input into the multi-level adaptive feedback mechanism to generate the feedback vector F i ;

[0042] S53, the multi-level adaptive feedback mechanism includes a local feedback layer, a multi-expert lightweight model feedback layer and a fusion feedback layer;

[0043] S54, local feedback layer according to the real-time state vector V of each node i With the historical state vector The difference between ΔV i Calculate primary feedback value

[0044] S55, multi-expert lightweight model feedback layer based on network traffic prediction value and the actual observed value y ij The error between Generate feedback value Where j represents the j-th multi-expert lightweight model;

[0045] S56, the fusion feedback layer generates a fusion feedback value f according to the overall performance index of the system g , and acts on the weighted coefficients of the output results of the multi-expert lightweight model in the adaptive multimodal fusion algorithm;

[0046] S57, based on primary feedback value Feedback value and fusion feedback value f g Construct comprehensive feedback weight α ij , used to adjust the output weighting coefficient of the j-th multi-expert lightweight model in the adaptive multimodal fusion algorithm;

[0047] S58, while adjusting the weighting coefficient, according to the feedback vector F i and performance change trends, the internal parameters θ of the multi-expert lightweight model j Update and output the updated weighted coefficient and multi-expert lightweight model parameters

[0048] This invention introduces a multi-level adaptive feedback mechanism to achieve real-time monitoring and dynamic adjustment of the performance of each service node in the optimized service architecture. The system first collects multidimensional performance indicators for each node and constructs a performance state vector. Feedback values ​​are then generated at different levels, sequentially in the local feedback layer (analyzing node state changes), the multi-expert model feedback layer (evaluating prediction errors), and the fusion feedback layer (measuring overall system performance). Based on these feedback values, a comprehensive feedback weight is constructed, dynamically adjusting the weighting coefficients of each expert model output. Simultaneously, the model's internal parameters are updated in real time based on the feedback vectors and performance change trends, thereby enhancing the system's responsiveness to environmental changes and its ability to continuously optimize.

[0049] Optionally, the S6 specifically includes:

[0050] S61, based on real-time monitoring data V i and feedback vector F i , by the data difference ΔV within the time window [t,t+Δt] i (t) Construct incremental sample set D t , and input into the expert lightweight model;

[0051] S62, using the incremental learning mechanism to adjust the parameters θ of the traffic prediction expert model, the load balancing expert model and the resource scheduling expert model j Make updates;

[0052] S63, introduce online training mechanism to adjust the weight coefficient α of the adaptive multimodal fusion algorithm ij Conduct continuous optimization and adopt a sliding window feedback fusion strategy;

[0053] S64, according to the updated expert model parameters and weighting coefficient Combined with the current network topology diagram G t =(N t ,E t )Node N t and edge E t The structural changes of the graph neural network are carried out, and the adjusted node features are output.

[0054] S65. Dynamically adjust the service node deployment density ρ based on the real-time growth rate r(t) of service requests and the distribution density function P(s,t) i (t);

[0055] S66, based on the updated and ρ i (t+1), automatically perform real-time reconstruction of service node resource configuration, scheduling path and service distribution strategy, and complete adaptive optimization and adjustment of service architecture.

[0056] The present invention realizes adaptive optimization and adjustment of service architecture through incremental learning and online training mechanism: it uses real-time monitoring data and feedback vectors to construct incremental sample sets within the time window, and dynamically updates the parameters of multiple expert models such as traffic prediction, load balancing and resource scheduling; and continuously optimizes the weighted coefficients of the fusion algorithm through the sliding window feedback fusion strategy; executes the graph neural network re-propagation process in combination with changes in the network topology structure, and updates the node feature representation; and dynamically adjusts the service node deployment density according to the real-time growth rate and distribution density of service requests, ultimately realizing real-time reconstruction of resource allocation, scheduling path and service distribution strategy, and improving the system's intelligent adaptation and optimization capabilities.

[0057] The beneficial effects of the present invention are:

[0058] This paper introduces a multi-expert lightweight model and combines it with an adaptive multimodal fusion algorithm to achieve division of labor modeling and collaborative decision-making for key optimization tasks in the service architecture. First, by constructing three types of multi-expert lightweight models for traffic prediction, load balancing, and resource scheduling, it can be optimized specifically for different task characteristics. This overcomes the limitation of existing technologies that a single model cannot balance the accuracy and efficiency of multiple tasks, significantly improving the processing effect and responsiveness of each subtask.

[0059] This method utilizes a self-attention mechanism to extract traffic and topological features, combining deep reinforcement learning with graph neural networks to model complex resource relationships. This approach improves prediction accuracy and scheduling robustness without increasing system burden. It is particularly suitable for large-scale service systems with numerous nodes and frequently changing structures. An adaptive multimodal fusion algorithm dynamically adjusts weighting coefficients based on the performance of multiple model outputs, making the combined optimization results more stable and adaptable. This addresses the problem that existing fixed fusion strategies are unable to cope with changing network states.

[0060] This invention incorporates a multi-level feedback mechanism and an online incremental update strategy, establishing a closed-loop control system of "monitoring-feedback-learning-optimization." This system dynamically adjusts the parameters and fusion strategies of each expert model based on real-time performance changes, enabling continuous model evolution and adaptive reconstruction of the service architecture. This mechanism improves the system's responsiveness and resource utilization efficiency in environments with sudden loads and network fluctuations.

[0061] In summary, the present invention not only improves the intelligent optimization capability of the service architecture, but also has comprehensive advantages such as lightweight structure, easy deployment, strong adaptability, and stable operation. It is suitable for various complex application scenarios such as cloud computing platforms, edge computing systems, and intelligent service networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0063] Figure 1 This is an overall flow chart of a service architecture optimization method based on a multi-expert lightweight model proposed in the present invention; DETAILED DESCRIPTION

[0064] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0065] refer to Figure 1 , a service architecture optimization method based on a multi-expert lightweight model, comprising the following steps:

[0066] S1. Collect the real-time status data of each service node in the service architecture and pre-process the real-time status data;

[0067] S2, a multi-expert lightweight model uses a self-attention mechanism to predict traffic flow on pre-processed data, dynamically optimizes resource allocation using an adaptive reinforcement learning algorithm, and introduces an adaptive multimodal fusion algorithm to jointly train and fuse the outputs of each model to generate a comprehensive decision;

[0068] S3. Combine comprehensive decision-making and generate dynamic optimization strategies for service architecture through multi-dimensional feedback learning mechanisms;

[0069] S4. Automatically adjust resource allocation, service scheduling, network routing, and load balancing of the service architecture based on dynamic optimization strategies.

[0070] S5. Monitor the performance of the optimized service architecture in real time. Based on the real-time monitoring results, adjust the weighting coefficients and parameters of the multi-expert lightweight model through a multi-level adaptive feedback mechanism.

[0071] S6. Use incremental learning and online training methods to update the parameters of the multi-expert lightweight model and adjust the service architecture according to network topology changes and service requirements.

[0072] The present invention collects real-time status data of each service node in the service architecture, constructs a multi-expert lightweight model, divides the work and models tasks such as traffic prediction and resource allocation, uses the self-attention mechanism to improve prediction accuracy, combines the adaptive reinforcement learning algorithm to achieve dynamic resource optimization, and uses the adaptive multimodal fusion algorithm to jointly train and fuse the outputs of each model to generate a comprehensive optimization decision; further generates a dynamic optimization strategy based on a multi-dimensional feedback learning mechanism, automatically adjusts the resource allocation, scheduling and routing strategies in the service architecture, and introduces a multi-level adaptive feedback mechanism and an online incremental learning method to update the model parameters in real time and adaptively adjust the service architecture according to changes in network topology and service demand, thereby realizing an efficient, intelligent and self-evolving service optimization process.

[0073] In this embodiment, the real-time status data includes performance indicators of each service node: CPU usage, memory occupancy, network bandwidth, network delay, traffic, and load.

[0074] In this embodiment, the preprocessing step includes denoising, data normalization, and extracting key features including load fluctuation, service node resource utilization, and network traffic patterns from real-time status data.

[0075] In this embodiment, a service architecture optimization method based on a multi-expert lightweight model is characterized in that the dynamic optimization strategy includes a resource allocation strategy of a service node, a network routing adjustment strategy, a load balancing strategy, and a service scheduling strategy. All dynamic optimization strategies make comprehensive decisions based on real-time network status data, historical network traffic patterns, real-time load information, and future traffic forecast results.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. The preprocessed data is input, and different optimization tasks are analyzed and predicted through multiple expert models, wherein the multiple expert models include a traffic prediction expert model, a load balancing expert model, and a resource scheduling expert model. The multiple expert models are jointly trained and output fused through an adaptive multimodal fusion algorithm. The adaptive multimodal fusion algorithm dynamically adjusts the output weight of each expert model according to the task characteristics and prediction accuracy of each expert model using a weighting mechanism.

[0078] S22. The traffic prediction expert model adopts a multi-layer time series modeling method based on the self-attention mechanism, combined with local network topology information, to dynamically adjust the weight of the traffic prediction task. t and network topology data T tIt will first be input into the self-attention mechanism module for processing. The self-attention mechanism will calculate the weight of each moment, which reflects the importance of historical data and network topology in traffic prediction. The specific calculation formula is as follows:

[0079]

[0080] Where: X t is the historical traffic data, T t is the network topology data at the current moment, W q and W k is the weight matrix in the attention mechanism, d k is the dimension of the query vector, A t is the calculated weighting matrix;

[0081] The weighted matrix A calculated by the self-attention mechanism t The original input features are weighted. Specifically, the original data X t It will interact with the calculated weight matrix to obtain a weighted feature representation:

[0082]

[0083] in It is a weighted input feature that contains a weighted combination of historical traffic and network topology information. The weighted features are input into the LSTM network to capture temporal and long-term dependencies.

[0084] The input of LSTM can be expressed as:

[0085]

[0086] Where: h t and c t are the hidden state and memory unit of the LSTM unit, is the weighted input.

[0087] After LSTM network processing, the output will generate a network traffic prediction value at a future moment Right now:

[0088]

[0089] Where: f(·) is the activation function of the neural network, W f are the weights in the network, For the predicted future traffic.

[0090] S23, real-time network load data L t , historical load pattern data L histand network topology data T t , this data is fed into the load balancing expert model. The load balancing expert model uses an adaptive reinforcement learning algorithm, deep Q-learning, and a multi-channel feedback mechanism to learn the complex nonlinear relationships between network states. The core goal of deep Q-learning is to optimize resource allocation strategies by maximizing expected long-term rewards. The specific Q-value update formula is as follows:

[0091] Q(s t ,a t )=(1-α)Q(s t ,a t )+α(r t +γmax a′ Q(s t+1 ,a ′ ))

[0092] Where: Q(s t ,a t ) is the current state s t and action a t Q value; α is the learning rate; r t is the instant reward at the current moment; γ is the discount factor; max a′ Q(s t+1 ,a ′ ) is the next state s t+1 The maximum Q value of all possible actions.

[0093] The multi-channel feedback mechanism uses a multi-channel feedback structure to dynamically adjust service node resource allocation decisions under real-time load fluctuations. Each channel reflects the feedback signal of a subsystem. By integrating multiple feedback signals, the model can make accurate optimization decisions under complex load fluctuations.

[0094] The load balancing strategy is adjusted through an adaptive reward mechanism. The reward mechanism is usually designed based on the system's resource utilization, response time, and load fluctuations to ensure that the system always maintains efficient operation. The formula for dynamic adjustment is as follows:

[0095] r t =f(L t ,L hist ,T t );

[0096] Where: r t is the reward at the current moment; L t is the real-time network load data; L hist is the historical load pattern data T t is the network topology data; f(·) is the reward function, which is designed according to the actual needs of the system.

[0097] The model dynamically adjusts the resource allocation of service nodes using the learned strategy. The goal of this adjustment process is to avoid the single-point optimization problem in traditional reinforcement learning models and achieve intelligent balancing of network load. The adjustment formula is as follows:

[0098] R t =Optimize(L t ,L hist ,T t );

[0099] Where: R t is the adjusted resource allocation strategy; L t is the real-time load data; L hist is the historical load data; T t is the network topology data; Optimize(·) is the optimization process.

[0100] The resource allocation strategy R adjusted by the adaptive reinforcement learning algorithm t As the output of load balancing decisions, it ensures dynamic balancing of resources under real-time load changes, avoids overload or idle resources, and optimizes service quality and resource utilization.

[0101] S24. The resource scheduling expert model uses a graph neural network (GNN)-based optimization method, combined with a deep graph convolutional network and a graph adaptive update strategy, to jointly analyze network topology and traffic prediction results. The resource scheduling optimization process can be represented by the following graph neural network formula:

[0102]

[0103] in, is the representation of node v at the k+1 layer, mathcalN(v) is the neighbor set of node v, C vu is the connection weight between neighbor node u and node v, mathbfW (k) is the weight matrix of the kth layer, σ is the activation function, mathbfb (k) This formula implements the adaptive update of network topology and traffic prediction data to optimize resource scheduling.

[0104] S25. The output results of the multiple expert models are fused using an adaptive multimodal fusion algorithm. The fusion algorithm dynamically weights the output results of each expert model by introducing a multi-dimensional feedback learning mechanism based on the prediction results of each expert model. The output includes traffic prediction results, load balancing decisions, and resource scheduling strategies. The adaptive multimodal fusion algorithm uses weighted averaging and ensemble learning techniques to achieve optimal fusion of the output results based on the real-time performance, prediction accuracy, and historical optimization results of each expert model. It is specifically expressed by the following formula:

[0105]

[0106] Among them, O final is the final optimization decision output, α i is the weighted coefficient of expert model i, O i is the output decision of expert model i, and N is the number of expert models. The output of each expert model is weighted and fused according to its real-time performance and task adaptability to ensure the best optimization results in different network environments.

[0107] In this embodiment, S3 specifically includes:

[0108] S32. Use a multi-dimensional feedback learning mechanism to optimize the comprehensive decision result S final This mechanism adjusts the weighting coefficients of each expert model's output based on historical feedback and real-time feedback information. Historical feedback helps update model weights by evaluating past optimization strategies; real-time feedback fine-tunes the output of each model based on the current network status (such as resource utilization, response time, traffic load, etc.). This feedback mechanism can automatically adjust the weighting coefficients in a changing network environment to ensure that the best comprehensive decision is made at every moment.

[0109] S32. Through the multi-dimensional feedback learning mechanism, the weighting coefficients of each expert model output will be dynamically adjusted according to the actual performance of the current network environment. These coefficients are not only based on the real-time network status, but also take into account the comprehensive evaluation of historical performance and optimization effects. Specifically, each weighting coefficient w i The reward function in the feedback learning mechanism will be adjusted in real time to adapt to changes in the network and ensure that the output of each model contributes optimally to the final decision at every moment.

[0110] S34. Through the continuous adjustment and optimization of the multi-dimensional feedback learning mechanism, combined with the output results of each expert model and its adjusted weighted coefficient, a dynamic optimization strategy S for the service architecture is generated. dynamic The strategy includes resource allocation of service nodes, network routing selection, load balancing and service scheduling. Dynamic optimization strategy S dynamic The weighted sum S of each model output will be final , and the weighted coefficients adjusted by the feedback learning mechanism, calculate and generate optimization strategies to ensure that the various resources of the service architecture are optimally configured and can cope with future changes in network traffic and load.

[0111] In this embodiment, the S4 specifically includes:

[0112] S41, according to the dynamic optimization strategy Sdynamic Adjust resource allocation for each service node in the service architecture, including key resource indicators such as CPU usage, memory usage, network bandwidth, network latency, traffic flow, and load;

[0113] S42, according to the dynamic optimization strategy S dynamic The network routing adjustment strategy in the network is used to optimize the routing path in the network. The network routing adjustment is based on the traffic prediction data Y t+1 and real-time network topology data, using the shortest path algorithm;

[0114] S43. Dynamically adjust the load distribution of the service nodes through an adaptive algorithm according to the load balancing strategy in the dynamic optimization strategy;

[0115] S44, according to the service scheduling strategy S in the dynamic optimization strategy opt ,automatically adjust the task scheduling of service nodes, including task allocation, service start and stop operations;

[0116] S45. Automatically optimize the service architecture based on the resource allocation, network routing, load balancing, and service scheduling strategies adjusted above.

[0117] In this embodiment, the S5 specifically includes:

[0118] S51, real-time monitoring of the performance indicators of each service node in the optimized service architecture, including but not limited to CPU usage C i , memory usage M i 、Network bandwidth B i , network delay D i , network traffic T i and system load L i , where i represents the i-th service node;

[0119] S52: Construct a multi-dimensional performance state vector V based on the real-time performance indicators of each service node i =

[0120] [C i ,M i ,B i ,D i ,T i ,L i ], and V i Input into the multi-level adaptive feedback mechanism to generate the feedback vector F i ;

[0121] S53, the multi-level adaptive feedback mechanism includes a local feedback layer, an expert model feedback layer and a fusion feedback layer:

[0122] The local feedback layer is based on the real-time state vector V of each node i With the historical state vector The difference between Calculate primary feedback value

[0123] The expert model feedback layer outputs the predictions of each expert model. The error between the actual observation value yij Generate feedback value fij e , where j represents the jth expert model (including traffic prediction expert model, load balancing expert model and resource scheduling expert model);

[0124] The fusion feedback layer generates the fusion feedback value f according to the overall performance index of the system g , and acts on the weighted coefficients of the output results of each expert model in the adaptive multimodal fusion algorithm;

[0125] S54, based on and f g Construct comprehensive feedback weight α ij , which is used to adjust the output weight coefficient of the j-th expert model in the adaptive multimodal fusion algorithm in S2. The update formula is:

[0126]

[0127] Where η is the learning rate, λ1, λ2, and λ3 are feedback weighting coefficients, satisfying λ1+λ2+λ3=1;

[0128] S55, while adjusting the weighting coefficient, according to the feedback vector F i and performance change trends, and the internal parameters θ of each expert model in S2 j To update, use the following gradient update formula:

[0129]

[0130] in represents the loss function of the j-th expert model, and γ is the learning rate for updating model parameters;

[0131] S56, the updated weighted coefficient and expert model parameters Applied to the adaptive multimodal fusion module and expert models in S2 to drive the execution of the next round of service architecture optimization tasks.

[0132] In this embodiment, S6 specifically includes:

[0133] S61, based on real-time monitoring data V i =[C i,M i ,B i ,D i ,T i ,L i ] and feedback vector F i , by the data difference ΔV within the time window [t,t+Δt] i (t) = V i (t+Δt)-V i (t) Construct incremental sample set D t , and input into each expert model;

[0134] S62, using the incremental learning mechanism to adjust the parameters θ of the traffic prediction expert model, the load balancing expert model and the resource scheduling expert model j To update, the incremental update formula is as follows:

[0135]

[0136] in Indicates that the j-th expert model is in the sample set D t The loss function on , γ t is the adaptive learning rate at the current moment;

[0137] S63, introduce online training mechanism to adjust the weight coefficient α of the adaptive multimodal fusion algorithm ij Continuous optimization is performed, and a sliding window feedback fusion strategy is adopted. The weight update formula is:

[0138]

[0139] in is the optimal weighted value calculated for the current feedback, λ t is the fusion factor based on dynamic adjustment of the window, satisfying 0<λ t <1;

[0140] S64, according to the updated expert model parameters and weighting coefficient Combined with the current network topology diagram G t =(N t ,E t )Node N t and edge E t The structural changes of the graph neural network are carried out, and the adjusted node features are output.

[0141] S65. Dynamically adjust the service node deployment density ρ based on the real-time growth rate r(t) of service requests and the distribution density function P(s,t) i (t), the adjustment strategy formula is as follows:

[0142] ρ i (t+1)=ρ i (t)+κ·(r(t)·P(i,t)-ρ i (t));

[0143] Where κ is the adjustment rate control factor, P(i,t) represents the service demand density of the i-th node at time t;

[0144] S66, based on the updated and ρ i (t+1), automatically perform real-time reconstruction of service node resource configuration, scheduling path and service distribution strategy, and complete adaptive optimization and adjustment of service architecture.

[0145] Example 1:

[0146] To verify the feasibility of the present invention in practice, the present invention was applied to a provincial government cloud platform resource scheduling system. This system primarily handles the daily processing and response services for government business data. Deployed in a provincial data center, the system includes over 500 service nodes, supporting concurrent requests from over 80 government business systems, processing over 3TB of data daily, and averaging over 15 million requests per day. The core issues facing this system include: delayed resource allocation due to inaccurate traffic forecasts during peak business periods, uneven load distribution between service nodes, increased scheduling response delays, and the inability to update optimization strategies in a timely manner when network topology is adjusted or nodes are added. These issues directly impact business continuity and government response speed.

[0147] In this application scenario, the present invention is deployed in the scheduling control center module to optimize the management of the entire government service architecture. The system first deploys the data acquisition module to collect the status data of each service node in real time, including CPU usage, memory usage, network bandwidth, node latency, real-time traffic, and request response load. Taking actual test data as an example, the acquisition cycle is 30 seconds, and about 120,000 feature samples can be generated per hour after data preprocessing. In actual business operations, such as the industrial and commercial tax system, the traffic fluctuates greatly during the high concurrency period (9:00-11:30), and the prediction error of the traditional rule scheduling model is as high as 38.5%, causing the average system delay to increase to more than 300ms. The following table shows the comparative experimental data of the present invention and the traditional method in the scheduling scenario of the provincial government platform (the sampling period is 7 consecutive days):

[0148] Table 1 Performance comparison of the method of the present invention and the traditional scheduling strategy in the real service architecture

[0149]

[0150] The present invention constructs three multi-expert lightweight models for traffic prediction, load balancing, and resource scheduling to handle key optimization tasks respectively. The traffic prediction model adopts a structure based on the fusion of the self-attention mechanism and LSTM, effectively capturing the fluctuation characteristics of each business in the historical time window, and combining the current network topology to model the relationship between nodes, achieving a prediction accuracy of 94.7% for the next 5 minutes. The load balancing expert model is based on deep Q learning and can learn the nonlinear relationship between the historical node load and the network status, realizing real-time resource reallocation. In the industrial and commercial system scheduling test, the load standard deviation was reduced from the original 0.34 to 0.09, and the service response stability was significantly improved. The resource scheduling expert model uses a graph neural network to model the global node topology and traffic distribution, realizing global resource scheduling optimization. In a simulation environment with dynamic topology adjustment (adding or closing 3-5 nodes per hour), the scheduling reconstruction of the present invention takes less than 1.2 seconds, while the original system takes an average of 7.8 seconds.

[0151] During the fusion phase, the adaptive multimodal fusion algorithm introduced in this paper dynamically adjusts weighting coefficients based on the output of the expert model. This effectively prevents the impact of single-model errors on overall optimization, especially in test scenarios with sudden increases in system service access (such as the social security service rush on the 15th of each month). The fusion results are fed back to the dynamic optimization strategy generation module, which adjusts resource scheduling paths, node task allocation, and routing schemes in real time.

[0152] Crucially, the present invention supports model self-evolution. Through incremental learning and online training mechanisms, when business data in government systems changes (e.g., when a new business system is launched or when node equipment is replaced), model parameters can be updated using incremental data samples within a sliding time window without retraining the entire model. This reduces system adaptation time from 10 minutes to less than 90 seconds, and improves resource reconfiguration accuracy to 97.2%.

[0153] The above data fully demonstrate that the present invention is superior to traditional methods in multiple dimensions such as traffic prediction accuracy, resource allocation balance, response speed, system stability and adaptability. It is particularly suitable for deployment in large-scale dynamic service architecture environments such as government cloud platforms, enterprise microservice systems, and intelligent edge node control, and has good engineering feasibility and promotion and application value.

[0154] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A service architecture optimization method based on a multi-expert lightweight model, characterized in that: The steps include: S1. Collect the real-time status data of each service node in the service architecture and pre-process the real-time status data; S2, a multi-expert lightweight model uses a self-attention mechanism to predict traffic flow on pre-processed data, dynamically optimizes resource allocation using an adaptive reinforcement learning algorithm, and introduces an adaptive multimodal fusion algorithm to jointly train and fuse the outputs of each model to generate a comprehensive decision; S3. Combine comprehensive decision-making and generate dynamic optimization strategies for service architecture through multi-dimensional feedback learning mechanisms; S4. Automatically adjust resource allocation, service scheduling, network routing, and load balancing of the service architecture based on dynamic optimization strategies. S5. Monitor the performance of the optimized service architecture in real time. Based on the real-time monitoring results, adjust the weighting coefficients and parameters of the multi-expert lightweight model through a multi-level adaptive feedback mechanism. S6. Use incremental learning and online training methods to update the parameters of the multi-expert lightweight model and adjust the service architecture according to network topology changes and service requirements.

2. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The real-time status data includes the performance indicators of each service node: CPU usage, memory occupancy, network bandwidth, network latency, traffic, and load.

3. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The preprocessing steps include denoising, data normalization, and extracting key features including load fluctuations, service node resource utilization, and network traffic patterns from real-time status data.

4. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The dynamic optimization strategy includes resource allocation strategy of service nodes, network routing adjustment strategy, load balancing strategy and service scheduling strategy. All dynamic optimization strategies are based on real-time network status data, historical network traffic patterns, real-time load information and future traffic forecast results for comprehensive decision-making.

5. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The S2 specifically includes: S21. Input the preprocessed data and analyze and predict different optimization tasks using a multi-expert lightweight model, wherein the multiple multi-expert lightweight models include a traffic prediction model, a load balancing model, and a resource scheduling model; S22, the traffic prediction model first calculates the traffic volume based on the historical traffic data X t and network topology data T t After preprocessing, the self-attention mechanism module calculates the weighted matrix A at each moment t , historical traffic data X t and network topology data T t With the weight matrix A t Perform weighted interaction to obtain weighted feature representation Represent the features Input to the LSTM network, after processing by the LSTM network, the output will generate a network traffic prediction value at a future moment S23, real-time network load data L t , historical load pattern data L hist and network topology data T t , will be input into the load balancing model, based on deep Q learning and multi-channel feedback mechanism, to learn the real-time network load data L t , historical load pattern data L hist and network topology data T t The complex nonlinear relationship between them; through a multi-channel feedback structure, dynamically adjust the resource allocation decision of service nodes under real-time network load changes; through an adaptive reward mechanism, adjust the load balancing strategy; S24. The reward mechanism is designed based on resource utilization, response time, and load fluctuation. The load balancing model dynamically adjusts the resource allocation of service nodes through the complex nonlinear relationship learned, the resource allocation decision of the service node, and the load balancing strategy. The resource allocation strategy R adjusted by the adaptive reinforcement learning algorithm t As an output of load balancing decisions; S25. Resource scheduling model uses graph neural network to analyze network topology data T t and traffic forecast results Perform joint analysis: Through graph convolution operations, the resource scheduling model can learn information from the adjacent nodes of each node. In the graph convolution operation, the weight of the node is based on the traffic prediction data. and real-time load information for dynamic adjustment; after multiple graph convolution operations and adaptive adjustments, the final feature representation of each node h v The optimized state of the node in the global network will be included; the final feature representation h of all nodes is combined v , the resource scheduling model will generate a globally optimized resource scheduling strategy S opt : S26. The input of the adaptive multimodal fusion algorithm is the network traffic prediction value Resource allocation strategy R t and resource scheduling strategy S opt , introduce the feedback learning mechanism to dynamically weight the output results of each multi-expert lightweight model; multiply the output results of each model by the weighted coefficient, and sum them to obtain the final comprehensive decision S final .

6. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The S3 specifically includes: S31. Use multi-dimensional feedback learning mechanism to optimize comprehensive decision-making results S final ,The multi-dimensional feedback learning mechanism adjusts the weighted coefficient of each multi-expert lightweight model output based on historical feedback and real-time feedback information; S32. For each result output by the multi-expert lightweight model, the weighting coefficient will be dynamically adjusted according to the actual performance of the current network environment. Each weighting coefficient w i They will be adjusted in real time according to the reward function in the feedback learning mechanism; S33. Dynamic optimization strategy for generating service architecture S dynamic , dynamic optimization strategy S dynamic It includes resource allocation of service nodes, network routing selection, load balancing and service scheduling.

7. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The S4 specifically includes: S41, according to the dynamic optimization strategy S dynamic Adjust resource allocation for each service node in the service architecture, including key resource indicators such as CPU usage, memory usage, network bandwidth, network latency, traffic flow, and load; S42, according to the dynamic optimization strategy S dynamic The network routing adjustment strategy in the network is used to optimize the routing path in the network; the network routing adjustment is based on the traffic prediction data Y t+1 and real-time network topology data, using the shortest path algorithm; S43. Dynamically adjust the load distribution of the service nodes through an adaptive algorithm according to the load balancing strategy in the dynamic optimization strategy; S44, according to the service scheduling strategy S in the dynamic optimization strategy opt ,automatically adjust the task scheduling of service nodes, including task allocation, service start and stop operations; S45. Automatically optimize the service architecture based on the resource allocation, network routing, load balancing, and service scheduling strategies adjusted above.

8. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The S5 specifically includes: S51, real-time monitoring of the performance indicators of each service node in the optimized service architecture, said performance indicators including CPU usage C i , memory usage M i 、Network bandwidth B i , network delay D i , network traffic T i and system load L i , where i represents the i-th service node; S52: Construct a multi-dimensional performance state vector V based on the real-time performance indicators of each service node i =[C i ,M i ,B i ,D i ,T i ,L i ], and V i Input into the multi-level adaptive feedback mechanism to generate the feedback vector F i ; S53, the multi-level adaptive feedback mechanism includes a local feedback layer, a multi-expert lightweight model feedback layer and a fusion feedback layer; S54, local feedback layer according to the real-time state vector V of each node i With the historical state vector The difference between ΔV i Calculate primary feedback value S55, multi-expert lightweight model feedback layer based on network traffic prediction value and the actual observed value y ij The error between Generate feedback value Where j represents the j-th multi-expert lightweight model; S56, the fusion feedback layer generates a fusion feedback value f according to the overall performance index g , and acts on the weighted coefficients of the output results of the multi-expert lightweight model in the adaptive multimodal fusion algorithm; S57, based on primary feedback value Feedback value and fusion feedback value f g Construct comprehensive feedback weight α ij , used to adjust the output weighting coefficient of the j-th multi-expert lightweight model in the adaptive multimodal fusion algorithm; S58, while adjusting the weighting coefficient, according to the feedback vector F i and performance change trends, the internal parameters θ of the multi-expert lightweight model j Update and output the updated weighted coefficient and multi-expert lightweight model parameters 9. The service architecture optimization method based on a multi-expert lightweight model according to claim 1 is characterized in that: The S6 specifically includes: S61, based on real-time monitoring data V i and feedback vector F i , by the data difference ΔV within the time window [t,t+Δt] i (t) Construct incremental sample set D t , and input into the expert lightweight model; S62, using the incremental learning mechanism to adjust the parameters θ of the traffic prediction expert model, the load balancing expert model and the resource scheduling expert model j Make updates; S63, introduce online training mechanism to adjust the weight coefficient α of the adaptive multimodal fusion algorithm ij Conduct continuous optimization and adopt a sliding window feedback fusion strategy; S64, according to the updated expert model parameters and weighting coefficient Combined with the current network topology diagram G t =(N t ,E t )Node N t and edge E t The structural changes of the graph neural network are carried out, and the adjusted node features are output. S65. Dynamically adjust the service node deployment density ρ based on the real-time growth rate r(t) of service requests and the distribution density function P(s,t) i (t); S66, based on the updated and ρ i (t+1), automatically perform real-time reconstruction of service node resource configuration, scheduling path and service distribution strategy, and complete adaptive optimization and adjustment of service architecture.

Images