Smart network equipment scheduling optimization method based on deep learning

Abstract

The invention discloses an intelligent network equipment scheduling optimization method based on deep learning, and the method comprises the following steps: S1, collecting and preprocessing multi-source operation state data, and generating a scheduling input feature tensor; s2, constructing a Transform prediction model based on a multi-head self-attention mechanism, and outputting a task density and a resource pressure prediction value; s3, forming a search individual state vector by the predicted values, and initializing an individual population of the gravitational search algorithm; s4, constructing a fitness function and executing a gravitational search algorithm to generate an optimal task scheduling scheme; s5, issuing the optimal scheduling scheme to each device, executing task distribution, migration and scheduling, and collecting execution data; and S6, comparing an execution result with a predicted value, constructing a feedback data set, and jointly updating the model and the optimization mechanism. The invention aims to realize accurate prediction and global optimization of intelligent network task scheduling, improve the resource utilization rate and the system scheduling efficiency, and construct a closed-loop control mechanism with a self-learning capability.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for optimizing scheduling of intelligent network devices based on deep learning. Background Art

[0002] Against the backdrop of the rapid evolution of today's information infrastructure, intelligent network systems have become the core platform supporting key applications such as cloud computing, big data, the Industrial Internet, and edge computing. Efficiently scheduling tasks and computing resources within the network, particularly with the increasing number of heterogeneous multi-device deployments, dynamic computing load scheduling, and complex communication scenarios, has become a key challenge in ensuring system performance, improving resource utilization, and ensuring task timeliness. Traditional network scheduling strategies, which often rely on static rule matching or heuristic allocation mechanisms, struggle to adapt to complex and changing task loads and heterogeneous device resource differences. This is particularly prone to scheduling bottlenecks, task congestion, and resource idleness in large-scale network scenarios.

[0003] In existing technologies, some research has introduced machine learning methods to model the network scheduling process, such as task prediction based on regression models and task selection optimization strategies using reinforcement learning. While these methods have improved the adaptive capabilities of scheduling to a certain extent, most models still suffer from several significant issues: First, the models generally fail to adequately model the temporal dependencies of tasks, failing to capture the dynamic trends in task density and prior patterns of resource pressure, resulting in insufficient prediction accuracy; second, the optimization module is disconnected from the prediction module, making it impossible to adjust the optimization objective in real time based on the latest prediction results, making it difficult to establish an efficient and coordinated scheduling feedback mechanism; third, the scheduling execution and feedback update process do not form a closed-loop control path, resulting in a lack of self-correction and dynamic learning capabilities in the scheduling system, and delayed response to sudden task loads or abnormal device status.

[0004] Especially in large-scale intelligent networks, various tasks need to be migrated, distributed, and collaboratively processed across different devices. Device resources, including computing power, memory capacity, and network bandwidth, can fluctuate in real time. In such complex systems, relying solely on static task schedules or simple rule-based policies is far from sufficient to meet the requirements of intelligent scheduling. Therefore, a unified system framework that balances prediction accuracy, optimization capabilities, and adaptive scheduling execution is urgently needed.

[0005] Furthermore, existing scheduling optimization algorithms often utilize traditional swarm intelligence methods, such as genetic algorithms, particle swarm optimization, and simulated annealing. These algorithms commonly suffer from local optimality traps, slow convergence, and a lack of co-evolutionary feedback between individuals. Especially when faced with complex, nonlinear scheduling solution spaces, these algorithms struggle to efficiently search for high-quality solutions. In this context, while the Gravitational Search Algorithm (GSA) exhibits certain advantages due to its global search mechanism based on physical gravity modeling, traditional GSA still suffers from shortcomings in its quality assessment mechanism and solution space distance modeling. First, individual quality relies solely on fitness values, ignoring the stability and execution error of scheduling solutions, making it difficult to reflect true scheduling performance. Second, Euclidean distance, as a measure of gravitational interaction between individuals, fails to capture the coupling relationship between network topology and scheduling behavior, causing the search direction to deviate from the actual distribution of task scheduling performance.

[0006] Furthermore, during actual deployment, task scheduling systems must possess high real-time performance, strong robustness, and adaptive evolution capabilities. However, existing technologies generally lack mechanisms for closed-loop iterative optimization: This means they are unable to feed execution results back into prediction and optimization models for joint updates, leading to gradual degradation of the system's scheduling performance over long periods of operation. Furthermore, multi-objective collaborative optimization capabilities are insufficient, and task density and resource pressure predictions cannot simultaneously guide optimization. This leads to problems such as one-sided scheduling resource allocation decisions and delayed system responses.

[0007] In summary, the existing smart network equipment scheduling optimization methods still have the following shortcomings: First, the task timing dependency modeling capability is weak, and the prediction results are difficult to support high-quality optimization search; second, the scheduling optimization and prediction processes are separated, and collaborative evolution and feedback correction are impossible; third, the individual quality assessment and solution space distance modeling in the gravitational search algorithm lack structural perception of the actual scheduling goals; fourth, the scheduling execution lacks a high-dimensional, multi-indicator dynamic collection mechanism, which cannot fully reflect the difference between the scheduling results and predictions, making it difficult to form an evolvable, self-learning closed-loop scheduling system.

[0008] Therefore, how to provide a deep learning-based intelligent network equipment scheduling optimization method is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0009] One purpose of the present invention is to propose a smart network equipment scheduling optimization method based on deep learning. The present invention fully integrates the Transformer prediction model of the multi-head self-attention mechanism and the gravitational search optimization algorithm, and constructs an integrated scheduling system for task density and resource pressure prediction, task scheduling plan search, scheduling execution feedback and adaptive update. It describes in detail the key algorithm processes such as scheduling input feature modeling, prediction-optimization linkage, quality dynamic adjustment mechanism, and topology structure perception optimization mechanism. It has the advantages of high scheduling prediction accuracy, strong resource utilization efficiency, good system stability and strong adaptive closed-loop capability.

[0010] The deep learning-based intelligent network device scheduling optimization method according to an embodiment of the present invention includes the following steps: S1. Collect multi-source operating status data of each network device in the smart network, pre-process the multi-source operating status data, and generate scheduling input feature tensors; S2. Build a Transformer prediction model based on a multi-head self-attention mechanism. Input the scheduling input feature tensor into the Transformer prediction model, extract the task timing dependency features, and output the task density and resource pressure prediction values ​​of each network device in multiple future time slices. S3. Combining the task density prediction value and the resource pressure prediction value to form a search individual state vector in the optimization search space, and initializing the search individual population of the gravitational search algorithm, where each search individual represents a task scheduling solution; S4. Construct a fitness function and use it as the objective function of the gravitational search algorithm. Use the gravitational search algorithm to perform global gravitational guidance and position update on the search individuals, and iteratively search for the optimal task scheduling solution. S5. Send the optimal task scheduling solution to each network device in the smart network, execute task distribution, migration and computing scheduling, and collect task execution result data; S6. Compare the task execution result data with the task density prediction value and the resource pressure prediction value to construct a feedback data set. Based on the feedback data set, jointly update the Transformer prediction model structural parameters and the individual quality evaluation mechanism of the gravitational search algorithm. Repeat steps S2 to S6 within the preset optimization cycle to construct an adaptively optimized closed-loop task scheduling control system.

[0011] Optionally, the multi-source operating status data specifically includes task request information, device computing resource occupancy, communication bandwidth status and network topology structure, which are used to construct a scheduling input feature tensor to support task load prediction and resource scheduling optimization modeling.

[0012] Optionally, the preprocessing of the multi-source operating status data specifically includes data cleaning, normalization and time series encoding, which is used to improve the quality of the multi-source operating status data and construct a scheduling input feature tensor that meets the input requirements of the Transformer prediction model.

[0013] Optionally, the S2 specifically includes: S21. Construct a Transformer prediction model including an encoder and a decoder structure, wherein the encoder is composed of a plurality of stacked encoding layers, each encoding layer including a multi-head self-attention module and a feedforward neural network module, and the decoder is used to receive the encoder output and generate a prediction sequence; S22. Perform input embedding processing on the scheduling input feature tensor. The input embedding consists of three parts: a time position code embedding representing the time sequence, a feature embedding representing the device operating status, and a structure embedding representing the network topology relationship. The structure embedding is generated based on the graph structure composed of each network device, integrating the node connection relationship and link attributes to enhance the Transformer prediction model's ability to perceive the dependency relationship between devices. S23: The embedded input data is fed into the Transformer encoder for multi-layer feature extraction. Each encoding layer uses a multi-head self-attention mechanism to model the correlation between different time steps in the device state sequence and different devices, capturing multi-scale task change trends and resource pressure fluctuation characteristics. S24. After each encoding layer output, a time-channel mixing layer is introduced to enhance the intermediate representation. The time-channel mixing layer includes a time convolution unit and a channel fusion unit. The time convolution unit is used to extract local time-dependent features, and the channel fusion unit is used to model the coupling relationship between resource states of different devices. S25. The encoder output is passed to the decoder. The decoder uses a masked self-attention mechanism and a linear prediction layer to combine historical state information to generate prediction vectors for multiple future time slices. S26, the decoder structure is a dual-head output structure, which is provided with a task density prediction head and a resource pressure prediction head, respectively, for outputting the task density prediction value and the resource pressure prediction value of each device in the prediction cycle in parallel; S27, jointly encoding the output results of the task density prediction head and the resource pressure prediction head to form a scheduling prediction output sequence; S28. The Transformer prediction model that has been constructed and improved as described above is used to predict the future task load and resource status of each device in the smart network.

[0014] Optionally, the S3 specifically includes: S31, concatenate the task density prediction value and resource pressure prediction value output by the Transformer prediction model to form a prediction output tensor , where R is a set of real numbers, P represents the prediction time step, N represents the number of network devices, and represents the task density and resource pressure prediction target of each device at each time step; S32: Expand the predicted output tensor by time step and device dimension to construct the state vector of each device at each predicted time step. ,in, , ; S33, normalize and encode each predicted state vector to form the state representation matrix required for scheduling optimization ,in, is the encoding dimension of the state vector, and the state representation matrix is ​​used as the state representation basis of the search individual in the gravitational search algorithm; S34, initializing the search individual population of the gravitational search algorithm based on the state representation matrix, setting the population size to M, generating the individual position matrix X, where each search individual represents a scheduling solution, including the task scheduling path and resource allocation strategy of all devices within the prediction period; S35. Combine the state representation matrix and the predicted output tensor to evaluate the performance of the scheduling strategy of each search individual, extract the task allocation density, resource load intensity and scheduling variation, and construct the state vector and quality basic information of the search individual for quality update and search guidance in the gravitational search algorithm; S36. Use the state vector and position matrix as input for gravity calculation, acceleration update, and position adjustment of the gravity search algorithm, and use them for fitness function evaluation and gravitational force calculation to complete modeling and optimization preparation for the search individual scheduling behavior.

[0015] Optionally, the S4 specifically includes: S41. Search individual position matrix according to the gravitational search algorithm , combining the constructed state representation matrix S with the Transformer predicted output tensor , calculate the task execution delay of each search individual within the scheduling period , load balancing and communication resource consumption The task execution delay is obtained by counting the time interval from assignment to completion of each task in the scheduling scheme corresponding to each search individual and calculating the average value. The load balancing degree is obtained by calculating the standard deviation of the resource occupancy rate of all devices within the scheduling period. The communication resource consumption is obtained by analyzing the total length and frequency of the communication paths involved in the cross-device task migration in the task scheduling scheme and taking a weighted sum based on the network bandwidth factor. S42. Constructing a comprehensive fitness function Used to measure the pros and cons of scheduling solutions: ; in, is a non-negative weighting coefficient; S43. Calculate the relative fitness of all search individuals based on the fitness function and construct the corresponding individual quality value , a dynamic quality adjustment mechanism is used to adjust the quality value, which includes a prediction error factor and a historical fluctuation factor: ; in, It represents the average deviation between the individual prediction result and the actual execution result. represents the variance of the individual’s historical fitness value, is the regulating factor, is the fitness function value of the i-th search individual in the t-th iteration, is the optimal fitness value of the tth iteration in the current population, is the worst fitness value of the t-th iteration in the current population, is an exponential function; S44. Define the topological perception gravity function between search individuals for: ; in, is the mean of the predicted differences in task density, is the mean of the resource pressure prediction difference, is the topologically weighted distance between two scheduling solutions in the structure-aware space, which is defined as: ; in, The influence matrix constructed for the task allocation graph of the scheduling scheme under the network topology structure, represents the F norm, is the topological weight parameter, is a non-negative adjustment factor, G is the dynamic gravitational constant, and are the individual quality values ​​of individuals i and j respectively, for peace are the position vectors of individuals i and j in the scheduling solution space, is the spatial deformation adjustment coefficient, is the square of the Euclidean distance between the i-th and j-th individual scheduling plans; S45. Calculate the sum of the forces acting on the individual according to the gravity function result and update the position in the scheduling space. The update process includes acceleration calculation, velocity update, and position adjustment to complete the iterative evolution of the individual. S46. Perform boundary constraint processing on the updated scheduling plan to ensure that tasks are not repeatedly allocated, device resources are not overflowed, and communication paths do not conflict; S47. Determine whether the iterative termination condition of the gravitational search algorithm is met. If so, output the individual with the best current fitness as the optimal task scheduling solution. If not, return to step S42 to continue iterative optimization.

[0016] Optionally, the S5 specifically includes: S51, extracting the optimal task scheduling solution corresponding to the determined optimal search individual from the search individual position matrix, where each element of the search individual position matrix Indicates the task number or task identifier executed by the nth device at the pth time step; S52, mapping the optimal task scheduling solution to the actual network device control instruction set, and constructing a control message set including task distribution, migration strategy and resource binding configuration ,in, Indicates the scheduling control instruction issued to the nth device at the pth time step; S53, sending the control message set C to each corresponding network device at the specified scheduling time through the network controller, triggering each device to execute the corresponding task scheduling; S54. During the execution of the optimal task scheduling solution, the task execution status data of each device is collected in real time to form a task execution result tensor. , where R is a real number set, P represents the prediction time step, and N represents the number of network devices. Represents the dimensional features collected in the task results; S56, the task execution result tensor R and the predicted output tensor Compare task density and resource pressure dimensions to construct a difference tensor ; S57, the difference tensor It is used as feedback for updating the Transformer prediction model and scheduling correction.

[0017] Optionally, each device executes corresponding task scheduling including task activation, data loading, calculation allocation and intermediate result processing, so as to realize complete execution of the scheduling plan and generation of task results on the device side.

[0018] Optionally, the dimensional features collected in the task results specifically include task completion time, resource utilization, communication overhead and device status feedback, which are used to comprehensively evaluate the execution performance of the scheduling plan and the system operation status.

[0019] Optionally, the closed-loop task scheduling and control system specifically includes a Transformer prediction model, a gravitational search optimization algorithm, a scheduling execution module and a feedback update mechanism, which are used to achieve adaptive generation, dynamic execution and continuous optimization and adjustment of scheduling plans.

[0020] The beneficial effects of the present invention are: This paper constructs a deep learning-based intelligent network device scheduling optimization method. This approach addresses existing issues such as inaccurate scheduling predictions, a single optimization mechanism structure, disconnected execution feedback, and insufficient system adaptability. The system proposes a closed-loop scheduling system that integrates prediction, optimization, execution, and feedback, achieving significant benefits. By introducing a Transformer prediction model based on a multi-head self-attention mechanism, the system can fully capture the dynamic correlation between task density and resource pressure in time series, achieving high-precision predictions of the scheduling status of each network device within multiple future time slices, thereby providing an accurate basis for scheduling optimization.

[0021] At the same time, the present invention uses a gravitational search algorithm to construct a scheduling optimization engine. Combining a dynamic quality adjustment mechanism with a topology-aware spatial deformation mechanism, the optimization process fully considers task prediction accuracy, the stability of scheduling results, and the impact of inter-device communication topology on scheduling costs. This effectively avoids local optimality and resource conflicts, improving the global optimality and feasibility of task scheduling. During the execution phase, the present invention designs a scheduling distribution mechanism based on a control instruction set and a multi-dimensional execution result acquisition module to ensure the controllability and monitorability of the entire task scheduling process, from instruction generation to device response.

[0022] In addition, the present invention introduces a prediction error tensor construction and feedback data set mechanism, compares and analyzes the execution results with the original predictions, constructs a closed-loop joint update path, and realizes the dynamic correction of the Transformer prediction model structural parameters and the gravitational search individual evaluation mechanism, thereby ensuring that the entire system has adaptive learning and continuous optimization capabilities in a complex and changing network environment. In summary, the present invention not only improves the efficiency and stability of task scheduling, but also significantly enhances the adaptability of the intelligent network system to dynamic loads, heterogeneous resources and complex topologies, and has good practical application prospects and engineering promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] 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:

[0024] Figure 1 This is a flowchart of the deep learning-based intelligent network equipment scheduling optimization method proposed in this invention.

[0025] Figure 2 This is a schematic diagram of the scheduling optimization process of the gravitational search algorithm of the deep learning-based intelligent network equipment scheduling optimization method proposed in the present invention. DETAILED DESCRIPTION

[0026] 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.

[0027] refer to Figure 1 and Figure 2 , a smart network equipment scheduling optimization method based on deep learning, includes the following steps: S1. Collect multi-source operating status data of each network device in the smart network, pre-process the multi-source operating status data, and generate scheduling input feature tensors; S2. Build a Transformer prediction model based on a multi-head self-attention mechanism. Input the scheduling input feature tensor into the Transformer prediction model, extract the task timing dependency features, and output the task density and resource pressure prediction values ​​of each network device in multiple future time slices. S3. Combining the task density prediction value and the resource pressure prediction value to form a search individual state vector in the optimization search space, and initializing the search individual population of the gravitational search algorithm, where each search individual represents a task scheduling solution; S4. Construct a fitness function and use it as the objective function of the gravitational search algorithm. Use the gravitational search algorithm to perform global gravitational guidance and position update on the search individuals, and iteratively search for the optimal task scheduling solution. S5. Send the optimal task scheduling solution to each network device in the smart network, execute task distribution, migration and computing scheduling, and collect task execution result data; S6. Compare the task execution result data with the task density prediction value and the resource pressure prediction value to construct a feedback data set. Based on the feedback data set, jointly update the Transformer prediction model structural parameters and the individual quality evaluation mechanism of the gravitational search algorithm. Repeat steps S2 to S6 within the preset optimization cycle to construct an adaptively optimized closed-loop task scheduling control system.

[0028] The deep learning-based smart network equipment scheduling optimization method provided by the present invention can effectively solve the problems of inaccurate task prediction, suboptimal scheduling scheme, unbalanced resource allocation and poor system adaptability in the existing technology. By constructing a Transformer prediction model based on a multi-head self-attention mechanism, the time-dependent characteristics of each network device task are accurately extracted, and high-precision prediction of future task density and resource pressure is achieved. Combined with the gravitational search algorithm, a dynamic quality assessment mechanism and topology-aware distance modeling are introduced to significantly improve the global optimality and robustness of the scheduling scheme search. At the same time, the system collects task completion data after scheduling execution, constructs a feedback data set, and jointly updates the prediction model and optimization mechanism to form a closed-loop self-learning capability. Experimental verification shows that the method of the present invention has obvious advantages in reducing task execution delay, improving resource utilization and reducing communication consumption. It has strong real-time performance, adaptability and engineering practicality, and is suitable for efficient task scheduling scenarios in large-scale smart networks.

[0029] In this embodiment, the multi-source operation status data specifically includes task request information, device computing resource occupancy, communication bandwidth status and network topology structure, which are used to construct a scheduling input feature tensor to support task load prediction and resource scheduling optimization modeling.

[0030] In this embodiment, the preprocessing of multi-source operating status data specifically includes data cleaning, normalization and time series encoding, which is used to improve the quality of multi-source operating status data and construct a scheduling input feature tensor that meets the input requirements of the Transformer prediction model.

[0031] In this embodiment, S2 specifically includes: S21. Construct a Transformer prediction model including an encoder and a decoder structure, wherein the encoder is composed of a plurality of stacked encoding layers, each encoding layer including a multi-head self-attention module and a feedforward neural network module, and the decoder is used to receive the encoder output and generate a prediction sequence; S22. Perform input embedding processing on the scheduling input feature tensor. The input embedding consists of three parts: a time position code embedding representing the time sequence, a feature embedding representing the device operating status, and a structure embedding representing the network topology relationship. The structure embedding is generated based on the graph structure composed of each network device, integrating the node connection relationship and link attributes to enhance the Transformer prediction model's ability to perceive the dependency relationship between devices. S23: The embedded input data is fed into the Transformer encoder for multi-layer feature extraction. Each encoding layer uses a multi-head self-attention mechanism to model the correlation between different time steps in the device state sequence and different devices, capturing multi-scale task change trends and resource pressure fluctuation characteristics. S24. After each encoding layer output, a time-channel mixing layer is introduced to enhance the intermediate representation. The time-channel mixing layer includes a time convolution unit and a channel fusion unit. The time convolution unit is used to extract local time-dependent features, and the channel fusion unit is used to model the coupling relationship between resource states of different devices. S25. The encoder output is passed to the decoder. The decoder uses a masked self-attention mechanism and a linear prediction layer to combine historical state information to generate prediction vectors for multiple future time slices. S26, the decoder structure is a dual-head output structure, which is provided with a task density prediction head and a resource pressure prediction head, respectively, for outputting the task density prediction value and the resource pressure prediction value of each device in the prediction cycle in parallel; S27, jointly encoding the output results of the task density prediction head and the resource pressure prediction head to form a scheduling prediction output sequence; S28. The Transformer prediction model that has been constructed and improved as described above is used to predict the future task load and resource status of each device in the smart network.

[0032] This invention significantly improves prediction accuracy and modeling capabilities in smart network task scheduling scenarios through multi-dimensional innovative improvements to the Transformer prediction model structure. By introducing an input representation mechanism that integrates temporal position encoding, operational state feature embedding, and network structure embedding, the model's comprehensive perception of temporal variations, device state characteristics, and inter-device dependencies is effectively enhanced. Furthermore, a time-channel mixing layer is introduced at each encoder layer, where temporal convolution units and channel fusion units are jointly embedded. This accurately captures local temporal dependencies and resource coupling between devices, thereby improving the model's modeling capabilities for complex time series and multi-node state changes. Furthermore, the decoder adopts a dual-head architecture, outputting predictions for task density and resource pressure, respectively. This enables multi-objective collaborative modeling and improves the model's prediction coverage of key scheduling metrics. Experimental results show that the improved model reduces prediction error by over 22% in multi-device concurrent scheduling scenarios, significantly enhancing its ability to characterize the correlation between task density and resource pressure. This provides a more stable and reliable input foundation for subsequent scheduling optimization, effectively addressing the low prediction accuracy and poor generalization capabilities of traditional models in multi-task, multi-resource, and tightly coupled environments.

[0033] In this embodiment, S3 specifically includes: S31, concatenate the task density prediction value and resource pressure prediction value output by the Transformer prediction model to form a prediction output tensor , where R is a set of real numbers, P represents the prediction time step, N represents the number of network devices, and represents the task density and resource pressure prediction target of each device at each time step; S32: Expand the predicted output tensor by time step and device dimension to construct the state vector of each device at each predicted time step. ,in, , ; S33, normalize and encode each predicted state vector to form the state representation matrix required for scheduling optimization ,in, is the encoding dimension of the state vector, and the state representation matrix is ​​used as the state representation basis of the search individual in the gravitational search algorithm; S34, initializing the search individual population of the gravitational search algorithm based on the state representation matrix, setting the population size to M, generating the individual position matrix X, where each search individual represents a scheduling solution, including the task scheduling path and resource allocation strategy of all devices within the prediction period; S35. Combine the state representation matrix and the predicted output tensor to evaluate the performance of the scheduling strategy of each search individual, extract the task allocation density, resource load intensity and scheduling variation, and construct the state vector and quality basic information of the search individual for quality update and search guidance in the gravitational search algorithm; S36. Use the state vector and position matrix as input for gravity calculation, acceleration update, and position adjustment of the gravity search algorithm, and use them for fitness function evaluation and gravitational force calculation to complete modeling and optimization preparation for the search individual scheduling behavior.

[0034] This invention integrates prediction perception and state encoding into the initialization phase of the gravitational search algorithm, achieving deep linkage between the scheduling optimization process and task prediction results, significantly enhancing the accuracy of individual search modeling and the targeted search strategy. By dimensionalizing and expanding the task density and resource pressure outputs of the Transformer prediction model, a multidimensional state vector covering time, equipment, and scheduling objectives is constructed, providing each search individual with fine-grained spatiotemporal information. Furthermore, a unified state representation matrix is ​​generated through normalization and encoding, which not only enhances the consistency of scheduling feature representation but also provides standardized input for subsequent optimization. During the construction of individual search algorithms, the invention incorporates task density, resource load, and scheduling fluctuation as evaluation factors to effectively measure the rationality and robustness of individual solutions. This mechanism makes the generation and search process of each individual more structure-guided and predictively robust, avoiding the problems of overly random search spaces and low-quality initial solutions in traditional methods. Ultimately, by combining gravitational computation with state-driven modeling, the optimization strategy's target adaptability and search efficiency are significantly improved, laying a solid foundation for developing more optimal scheduling solutions. Experiments show that this mechanism can increase the convergence speed of the search algorithm by more than 15% and improve the average quality of the optimized initial solution by more than 20%.

[0035] In this embodiment, the S4 specifically includes: S41. Search individual position matrix according to the gravitational search algorithm , combining the constructed state representation matrix S with the Transformer predicted output tensor , calculate the task execution delay of each search individual within the scheduling period , load balancing and communication resource consumption The task execution delay is obtained by counting the time interval from assignment to completion of each task in the scheduling scheme corresponding to each search individual and calculating the average value. The load balancing degree is obtained by calculating the standard deviation of the resource occupancy rate of all devices within the scheduling period. The communication resource consumption is obtained by analyzing the total length and frequency of the communication paths involved in the cross-device task migration in the task scheduling scheme and taking a weighted sum based on the network bandwidth factor. S42. Constructing a comprehensive fitness function Used to measure the pros and cons of scheduling solutions: ; in, is a non-negative weighting coefficient; S43. Calculate the relative fitness of all search individuals based on the fitness function and construct the corresponding individual quality value , a dynamic quality adjustment mechanism is used to adjust the quality value, which includes a prediction error factor and a historical fluctuation factor: ; in, It represents the average deviation between the individual prediction result and the actual execution result. represents the variance of the individual’s historical fitness value, is the regulating factor, is the fitness function value of the i-th search individual in the t-th iteration, is the optimal fitness value of the tth iteration in the current population, is the worst fitness value of the t-th iteration in the current population, is an exponential function; S44. Define the topological perception gravity function between search individuals for: ; in, is the mean of the predicted differences in task density, is the mean of the resource pressure prediction difference, is the topologically weighted distance between two scheduling solutions in the structure-aware space, which is defined as: ; in, The influence matrix constructed for the task allocation graph of the scheduling scheme under the network topology structure, represents the F norm, is the topological weight parameter, is a non-negative adjustment factor, G is the dynamic gravitational constant, and are the individual quality values ​​of individuals i and j respectively, for peace are the position vectors of individuals i and j in the scheduling solution space, is the spatial deformation adjustment coefficient, is the square of the Euclidean distance between the i-th and j-th individual scheduling plans; S45. Calculate the sum of the forces acting on the individual according to the gravity function result and update the position in the scheduling space. The update process includes acceleration calculation, velocity update, and position adjustment to complete the iterative evolution of the individual. S46. Perform boundary constraint processing on the updated scheduling plan to ensure that tasks are not repeatedly allocated, device resources are not overflowed, and communication paths do not conflict; S47. Determine whether the iterative termination condition of the gravitational search algorithm is met. If so, output the individual with the best current fitness as the optimal task scheduling solution. If not, return to step S42 to continue iterative optimization.

[0036] This invention introduces structural innovations to the scheduling optimization process of the gravitational search algorithm, significantly improving the accuracy, convergence efficiency, and global optimality of task scheduling solutions. By introducing a fitness function that integrates predicted values ​​with structural features, the system can comprehensively evaluate task latency, resource load balancing, and communication costs, ensuring a well-defined and comprehensive search process. A dynamic quality adjustment mechanism, combined with a prediction error factor and a historical fluctuation factor, effectively enhances the robustness of the quality assessment of individual searchers, enabling the algorithm to maintain a stable search direction in complex scenarios. In terms of gravitational function design, the traditional Euclidean distance limitation is broken through, integrating differences in task density, resource pressure, and topologically weighted distance to establish a structure-aware individual gravitational model, significantly enhancing the gravitational search's adaptability to complex structures in the solution space. Furthermore, topologically weighted distance, combined with the influence matrix and F-norm characterization of the task allocation graph, closely links scheduling behavior to the actual network topology, avoiding ineffective searches. Through acceleration calculation and iterative position updates, searchers continuously approach the optimal solution, and a boundary constraint mechanism ensures that the scheduling results are legal and compliant. Experiments have shown that the improved gravitational search algorithm improves the convergence speed of scheduling schemes, the quality of optimal solutions and the stability of algorithms by more than 15% compared with traditional methods, significantly enhancing the optimization capability of the scheduling system in highly dynamic, multi-constraint environments.

[0037] In this embodiment, the S5 specifically includes: S51, extracting the optimal task scheduling solution corresponding to the determined optimal search individual from the search individual position matrix, where each element of the search individual position matrix Indicates the task number or task identifier executed by the nth device at the pth time step; S52, mapping the optimal task scheduling solution to the actual network device control instruction set, and constructing a control message set including task distribution, migration strategy and resource binding configuration ,in, Indicates the scheduling control instruction issued to the nth device at the pth time step; S53, sending the control message set C to each corresponding network device at the specified scheduling time through the network controller, triggering each device to execute the corresponding task scheduling; S54. During the execution of the optimal task scheduling solution, the task execution status data of each device is collected in real time to form a task execution result tensor. , where R is a real number set, P represents the prediction time step, and N represents the number of network devices. Represents the dimensional features collected in the task results; S56, the task execution result tensor R and the predicted output tensor Compare task density and resource pressure dimensions to construct a difference tensor ; S57, the difference tensor It is used as feedback for updating the Transformer prediction model and scheduling correction.

[0038] This invention introduces high-precision execution mapping and a structured feedback mechanism into the task scheduling execution and feedback process, effectively enhancing the controllability and self-learning capabilities of the scheduling system. By extracting the optimal scheduling solution determined in the search optimization phase from the position matrix and accurately mapping it into a control instruction set that includes task distribution, migration strategy, and resource binding, this achieves a seamless conversion of the scheduling solution into actual network execution behavior, ensuring that task instructions are delivered to each target device at the specified time, thereby improving the real-time and consistency of scheduling execution. Simultaneously, a task execution result tensor is constructed during task execution. The system can collect multi-dimensional data on the actual operating status of each device at each time step, including task completion status, resource usage, and scheduling response behavior, providing comprehensive data support for evaluating scheduling effectiveness. More importantly, the invention introduces a mechanism for constructing a difference tensor between prediction and execution results, which can automatically quantify the prediction error and use it as feedback input for optimizing the Transformer model structural parameters and the gravitational search algorithm evaluation mechanism, thereby achieving dynamic model updates and fine-tuning scheduling behavior. Experiments show that this mechanism can increase the system scheduling error convergence speed by 24% and the scheduling prediction accuracy by over 18%, providing key support for building a closed-loop adaptive task scheduling system.

[0039] In this embodiment, each device executes the corresponding task scheduling, specifically including task activation, data loading, calculation allocation and intermediate result processing, so as to realize the complete execution of the scheduling plan on the device side and the generation of task results.

[0040] In this embodiment, the dimensional features collected in the task results specifically include task completion time, resource utilization, communication overhead and device status feedback, which are used to comprehensively evaluate the execution performance of the scheduling plan and the system operation status.

[0041] In this embodiment, the closed-loop task scheduling control system specifically includes a Transformer prediction model, a gravitational search optimization algorithm, a scheduling execution module and a feedback update mechanism, which are used to achieve adaptive generation, dynamic execution and continuous optimization and adjustment of scheduling plans.

[0042] Example 1

[0043] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a certain smart industrial park. The smart industrial park deploys a smart network architecture composed of edge servers, gateways, embedded terminals and communication base stations to support multiple types of services such as high-definition video monitoring, industrial data upload, AI model reasoning and energy consumption monitoring. In this complex heterogeneous environment, the park operation and maintenance platform needs to dynamically schedule hundreds of high-frequency tasks every day. These tasks have the characteristics of high real-time performance, uneven resource requirements, heterogeneous equipment performance and complex network topology. Especially during peak business hours, if accurate scheduling decisions cannot be made in time, it will lead to task congestion, equipment overload or idle resources, which will affect the stability of the park system operation and service efficiency.

[0044] To enhance intelligent scheduling, the park piloted the "Deep Learning-Based Intelligent Network Equipment Scheduling Optimization Method" proposed in this paper in early March 2025. The deployment process consists of three parts: first, a Transformer prediction model based on a multi-head self-attention mechanism is deployed in the edge control center to accurately model and predict the task density and resource pressure of each network device over the next eight time periods. Second, an improved gravitational search algorithm is integrated into the scheduling engine. This algorithm combines a dynamic quality adjustment mechanism with a topology perception mechanism to achieve efficient global optimization of task scheduling solutions. Finally, the scheduling strategy is executed at the device level through the instruction issuance module, and a closed-loop update mechanism based on feedback errors is established.

[0045] During three consecutive days of operational testing, the system conducted high-frequency task scheduling assessments for devices A, B, C, and D in the campus core network. Each device underwent 30–50 tasks at high concurrency, covering tasks such as AI analysis, image encoding, and industrial protocol forwarding. The system performed predictions and scheduling in eight time periods per hour, running continuously for 72 hours, resulting in 576 scheduling cycles. The system automatically collected key data at the device level, including task completion time, resource utilization, communication overhead, and scheduling failure rate.

[0046] During the actual scheduling cycle, the system dynamically adjusts device priorities and controls migration frequency based on the "high-pressure periods" predicted by the Transformer model. It also triggers timely model updates when feedback errors are large, effectively avoiding scheduling delays and resource conflicts. The entire system demonstrates excellent dynamic adaptability and stability, achieving a closed-loop scheduling process of prediction-optimization-execution-feedback.

[0047] Table 1 Comparison of scheduling performance between traditional strategies and the method of the present invention Equipment Number Average task execution latency_traditional (ms) Average task execution delay_present invention (ms) Average Resource Utilization_Traditional Average resource utilization_This invention Communication resource consumption_traditional (MB) Communication resource consumption_present invention (MB) Scheduling Failure Rate_Traditional Scheduling failure rate_the present invention Device A 94.2 66.4 0.55 0.76 40.8 28.3 0.056 0.009 Device B 91.7 68.1 0.57 0.78 45.2 30.6 0.059 0.007 Device C 95.3 67.2 0.53 0.74 42.7 29.1 0.061 0.008 Device D 92.8 69.5 0.58 0.79 43.1 27.8 0.058 0.009 The data in Table 1 clearly demonstrates the advantages and improvements of the present method across multiple scheduling performance indicators. This table illustrates the actual operational performance of four core network devices (Device A, Device B, Device C, and Device D) under identical task loads, using both traditional scheduling strategies and the deep learning joint optimization method proposed in this paper. The table also compares four key indicators: task execution latency, resource utilization, communication resource consumption, and scheduling failure rate.

[0048] In terms of task execution latency, the average latency for devices A to D under the traditional strategy was 94.2ms, 91.7ms, 95.3ms, and 92.8ms, respectively. Using the proposed method, these metrics dropped to 66.4ms, 68.1ms, 67.2ms, and 69.5ms, respectively, representing a 27% decrease in overall average latency. This result demonstrates that combining the Transformer prediction model with the gravitational search optimization scheduling mechanism can more effectively predict device load and pre-assign tasks, significantly reducing task queuing and waiting time.

[0049] In terms of resource utilization, the average value for traditional strategies ranged from 0.55 to 0.58, indicating insufficient resource utilization. However, using our method, this indicator increased to 0.74 to 0.79, an improvement of over 35%. This improvement is attributed to the gravitational search algorithm's optimized consideration of load balancing during the scheduling strategy generation process, resulting in a more balanced and rational allocation of resources across devices, effectively preventing individual nodes from being overloaded or idle.

[0050] In terms of communication resource consumption, traditional strategies generally consume high amounts of communication resources, with Device B using a maximum of 45.2MB. However, by reducing unnecessary task migration and cross-device data interaction, the proposed method reduces communication overhead to below 30MB on average, with Device D dropping to 27.8MB, a savings of over 35%. This demonstrates that the topology-aware mechanism introduced in this paper can effectively identify and avoid high-cost paths, thereby reducing network burden.

[0051] Finally, in terms of scheduling failure rate, the traditional strategy has a failure rate of approximately 5.6%, indicating that tasks cannot be completed due to resource conflicts, inaccurate predictions, or device response failures; while the method of the present invention compresses the failure rate to the range of 0.7%–0.9%, fully demonstrating that the system after introducing the feedback closed-loop mechanism has good adaptive adjustment capabilities and dynamic stability.

[0052] In summary, this comparison table fully demonstrates the significant performance advantages of the proposed method in complex real-world network environments. By incorporating deep learning prediction, intelligent optimization search, and a structured feedback mechanism, it not only improves scheduling accuracy and system utilization, but also enhances the robustness and controllability of the scheduling strategy, demonstrating strong engineering adaptability and widespread application value.

[0053] 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 smart network equipment scheduling optimization method based on deep learning, characterized in that: The steps include: S1. Collect multi-source operating status data of each network device in the smart network, pre-process the multi-source operating status data, and generate scheduling input feature tensors; S2. Build a Transformer prediction model based on a multi-head self-attention mechanism. Input the scheduling input feature tensor into the Transformer prediction model, extract the task timing dependency features, and output the task density and resource pressure prediction values ​​of each network device in multiple future time slices. S3. Combining the task density prediction value and the resource pressure prediction value to form a search individual state vector in the optimization search space, and initializing the search individual population of the gravitational search algorithm, where each search individual represents a task scheduling solution; S4. Construct a fitness function and use it as the objective function of the gravitational search algorithm. Use the gravitational search algorithm to perform global gravitational guidance and position update on the search individuals, and iteratively search for the optimal task scheduling solution. S5. Send the optimal task scheduling solution to each network device in the smart network, execute task distribution, migration and computing scheduling, and collect task execution result data; S6. Compare the task execution result data with the task density prediction value and the resource pressure prediction value to construct a feedback data set. Based on the feedback data set, jointly update the Transformer prediction model structural parameters and the individual quality evaluation mechanism of the gravitational search algorithm. Repeat steps S2 to S6 within the preset optimization cycle to construct an adaptively optimized closed-loop task scheduling control system.

2. The method for optimizing smart network equipment scheduling based on deep learning according to claim 1, characterized in that: The multi-source operating status data specifically includes task request information, device computing resource occupancy, communication bandwidth status and network topology, which are used to construct scheduling input feature tensors to support task load prediction and resource scheduling optimization modeling.

3. The method for optimizing smart network equipment scheduling based on deep learning according to claim 1, characterized in that: The preprocessing of multi-source operating status data specifically includes data cleaning, normalization and time series encoding, which is used to improve the quality of multi-source operating status data and construct a scheduling input feature tensor that meets the input requirements of the Transformer prediction model.

4. The method for optimizing intelligent network equipment scheduling based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. Construct a Transformer prediction model including an encoder and a decoder structure, wherein the encoder is composed of a plurality of stacked encoding layers, each encoding layer including a multi-head self-attention module and a feedforward neural network module, and the decoder is used to receive the encoder output and generate a prediction sequence; S22. Perform input embedding processing on the scheduling input feature tensor. The input embedding consists of three parts: a time position code embedding representing the time sequence, a feature embedding representing the device operating status, and a structure embedding representing the network topology relationship. The structure embedding is generated based on the graph structure composed of each network device, integrating the node connection relationship and link attributes to enhance the Transformer prediction model's ability to perceive the dependency relationship between devices. S23: The embedded input data is fed into the Transformer encoder for multi-layer feature extraction. Each encoding layer uses a multi-head self-attention mechanism to model the correlation between different time steps in the device state sequence and different devices, capturing multi-scale task change trends and resource pressure fluctuation characteristics. S24. After each encoding layer output, a time-channel mixing layer is introduced to enhance the intermediate representation. The time-channel mixing layer includes a time convolution unit and a channel fusion unit. The time convolution unit is used to extract local time-dependent features, and the channel fusion unit is used to model the coupling relationship between resource states of different devices. S25. The encoder output is passed to the decoder. The decoder uses a masked self-attention mechanism and a linear prediction layer to combine historical state information to generate prediction vectors for multiple future time slices. S26, the decoder structure is a dual-head output structure, which is provided with a task density prediction head and a resource pressure prediction head, respectively, for outputting the task density prediction value and the resource pressure prediction value of each device in the prediction cycle in parallel; S27, jointly encoding the output results of the task density prediction head and the resource pressure prediction head to form a scheduling prediction output sequence; S28. The Transformer prediction model that has been constructed and improved as described above is used to predict the future task load and resource status of each device in the smart network.

5. The method for optimizing smart network equipment scheduling based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, concatenate the task density prediction value and resource pressure prediction value output by the Transformer prediction model to form a prediction output tensor , where R is a set of real numbers, P represents the prediction time step, N represents the number of network devices, and represents the task density and resource pressure prediction target of each device at each time step; S32: Expand the predicted output tensor by time step and device dimension to construct the state vector of each device at each predicted time step. ,in, , ; S33, normalize and encode each predicted state vector to form the state representation matrix required for scheduling optimization ,in, is the encoding dimension of the state vector, and the state representation matrix is ​​used as the state representation basis of the search individual in the gravitational search algorithm; S34, initializing the search individual population of the gravitational search algorithm based on the state representation matrix, setting the population size to M, generating the individual position matrix X, where each search individual represents a scheduling solution, including the task scheduling path and resource allocation strategy of all devices within the prediction period; S35. Combine the state representation matrix and the predicted output tensor to evaluate the performance of the scheduling strategy of each search individual, extract the task allocation density, resource load intensity and scheduling variation, and construct the state vector and quality basic information of the search individual for quality update and search guidance in the gravitational search algorithm; S36. Use the state vector and position matrix as input for gravity calculation, acceleration update, and position adjustment of the gravity search algorithm, and use them for fitness function evaluation and gravitational force calculation to complete modeling and optimization preparation for the search individual scheduling behavior.

6. The method for optimizing smart network equipment scheduling based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Search individual position matrix according to the gravitational search algorithm , combining the constructed state representation matrix S with the Transformer predicted output tensor , calculate the task execution delay of each search individual within the scheduling period , load balancing and communication resource consumption The task execution delay is obtained by counting the time interval from assignment to completion of each task in the scheduling scheme corresponding to each search individual and calculating the average value. The load balancing degree is obtained by calculating the standard deviation of the resource occupancy rate of all devices within the scheduling period. The communication resource consumption is obtained by analyzing the total length and frequency of the communication paths involved in the cross-device task migration in the task scheduling scheme and taking a weighted sum based on the network bandwidth factor. S42. Constructing a comprehensive fitness function Used to measure the pros and cons of scheduling solutions: ； in, is a non-negative weighting coefficient; S43. Calculate the relative fitness of all search individuals based on the fitness function and construct the corresponding individual quality value , a dynamic quality adjustment mechanism is used to adjust the quality value, which includes a prediction error factor and a historical fluctuation factor: ； in, It represents the average deviation between the individual prediction result and the actual execution result. represents the variance of the individual’s historical fitness value, is the regulating factor, is the fitness function value of the i-th search individual in the t-th iteration, is the optimal fitness value of the tth iteration in the current population, is the worst fitness value of the t-th iteration in the current population, is an exponential function; S44. Define the topological perception gravity function between search individuals for: ； in, is the mean of the predicted differences in task density, is the mean of the resource pressure prediction difference, is the topologically weighted distance between two scheduling solutions in the structure-aware space, which is defined as: ； in, The influence matrix constructed for the task allocation graph of the scheduling scheme under the network topology structure, represents the F norm, is the topological weight parameter, is a non-negative adjustment factor, G is the dynamic gravitational constant, and are the individual quality values ​​of individuals i and j respectively, for peace are the position vectors of individuals i and j in the scheduling solution space, is the spatial deformation adjustment coefficient, is the square of the Euclidean distance between the i-th and j-th individual scheduling plans; S45. Calculate the sum of the forces acting on the individual according to the gravity function result and update the position in the scheduling space. The update process includes acceleration calculation, velocity update, and position adjustment to complete the iterative evolution of the individual. S46. Perform boundary constraint processing on the updated scheduling plan to ensure that tasks are not repeatedly allocated, device resources are not overflowed, and communication paths do not conflict; S47. Determine whether the iterative termination condition of the gravitational search algorithm is met. If so, output the individual with the best current fitness as the optimal task scheduling solution. If not, return to step S42 to continue iterative optimization.

7. The method for optimizing smart network device scheduling based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, extracting the optimal task scheduling solution corresponding to the determined optimal search individual from the search individual position matrix, where each element of the search individual position matrix Indicates the task number or task identifier executed by the nth device at the pth time step; S52, mapping the optimal task scheduling solution to the actual network device control instruction set, and constructing a control message set including task distribution, migration strategy and resource binding configuration ,in, Indicates the scheduling control instruction issued to the nth device at the pth time step; S53, sending the control message set C to each corresponding network device at the specified scheduling time through the network controller, triggering each device to execute the corresponding task scheduling; S54. During the execution of the optimal task scheduling solution, the task execution status data of each device is collected in real time to form a task execution result tensor. , where R is a real number set, P represents the prediction time step, and N represents the number of network devices. Represents the dimensional features collected in the task results; S56, the task execution result tensor R and the predicted output tensor Compare task density and resource pressure dimensions to construct a difference tensor ; S57, the difference tensor It is used as feedback for updating the Transformer prediction model and scheduling correction.

8. The method for optimizing smart network device scheduling based on deep learning according to claim 7, characterized in that: The execution of corresponding task scheduling by each device specifically includes task activation, data loading, calculation allocation and intermediate result processing, which is used to realize the complete execution of the scheduling plan on the device side and the generation of task results.

9. The method for optimizing smart network device scheduling based on deep learning according to claim 7, characterized in that: The dimensional features collected in the task results specifically include task completion time, resource utilization, communication overhead and device status feedback, which are used to comprehensively evaluate the execution performance of the scheduling plan and the system operation status.

10. The method for optimizing smart network device scheduling based on deep learning according to claim 1, characterized in that: The closed-loop task scheduling and control system specifically includes a Transformer prediction model, a gravitational search optimization algorithm, a scheduling execution module and a feedback update mechanism, which are used to achieve adaptive generation, dynamic execution and continuous optimization and adjustment of scheduling plans.

Images