Productivity tool service station computing resource scheduling method, medium and system

Abstract

The invention provides a productivity tool service station computing resource scheduling method, a medium and a system, belongs to the technical field of electric digital data processing, and provides a productivity tool service station computing resource scheduling method which adopts a plurality of lightweight neural networks of a total-sub-total structure to perform resource scheduling. The method comprises the following steps: firstly, acquiring resource demand characteristics of a to-be-scheduled task set, and establishing a lightweight neural network structure comprising a multi-path selection model, a plurality of scheduling calculation models and a fusion output model; task allocation is guided through a jump index, scheme fusion is performed by using a merge index, and adaptive scheduling of diversified tasks is realized. The method further comprises the steps of model training, scheme optimization, resource allocation, index updating and the like, the scheduling effect is improved through iterative optimization, the accuracy and reliability of the scheduling scheme are ensured, and the technical problem that in the prior art, a resource scheduling model is difficult to adaptively process diversified tasks in a lightweight environment is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic digital data processing. Specifically, it relates to a method, medium, and system for scheduling computing resources of a productivity tool service station. Background Art

[0002] A productivity tool service station is a cloud service platform that provides users with various productivity tools such as industrial production data processing, scheduling, and office operation. With the expansion of the user scale and the increase in business complexity, the platform needs to efficiently process diverse tasks including text processing, image editing, audio and video transcoding, data modeling, etc. Traditional resource scheduling methods mainly adopt rule-based scheduling strategies and queue-based task allocation mechanisms, and perform task scheduling through preset priority rules and resource thresholds. These methods perform well in processing single-type tasks and have the advantages of simple implementation and low computational overhead. With the development of deep learning technology, researchers have begun to attempt to apply neural networks to the field of resource scheduling, and have proposed a scheduling framework based on deep reinforcement learning and a task allocation model based on graph neural networks.

[0003] However, the existing resource scheduling technologies face many challenges in practical applications. First, the traditional rule-based scheduling method lacks adaptability and is difficult to cope with dynamically changing task characteristics and resource states, often resulting in unbalanced resource allocation and a decline in the overall system performance. Second, although the scheduling model based on deep learning has strong adaptability, its model structure is complex and the computational overhead is large, making it difficult to be deployed in resource-constrained edge environments. In addition, most of the existing scheduling models are optimized for specific types of tasks and lack the general processing ability for diverse tasks. When facing tasks with different characteristics, the model needs to be retrained, increasing the system maintenance cost.

[0004] Currently, how to achieve adaptive scheduling of diverse tasks in a lightweight environment is still a technical problem to be solved urgently. Especially in the edge computing scenario, the computing resources are limited and the task types are diverse, which puts higher requirements on the lightweight and adaptability of the scheduling model. The existing technologies either have too large a computational overhead or insufficient adaptability, and it is difficult to achieve lightweight deployment while ensuring the scheduling effect. That is to say, there is a technical problem in the existing technology that the resource scheduling model is difficult to adaptively process diverse tasks in a lightweight environment; therefore, there is an urgent need for a new scheduling method that can efficiently process diverse tasks in a lightweight environment. Summary of the Invention

[0005] In view of this, the present invention provides a method, medium, and system for scheduling computing resources of a productivity tool service station, which can solve the technical problem in the existing technology that the resource scheduling model is difficult to adaptively process diverse tasks in a lightweight environment.

[0006] The present invention is implemented as follows: A method for scheduling computing resources of a productivity tool service station provided by the first aspect of the present invention includes the following steps: obtaining a set of tasks to be scheduled in the service station, where the set of tasks to be scheduled includes physical resource requirement parameters and information resource requirement parameters; establishing a plurality of lightweight neural networks with a total - sub - total structure, including a multi - way selection model, a plurality of scheduling calculation models, and a fusion output model; performing supervised learning training on the multi - way selection model based on task characteristics; using the multi - way selection model to classify the set of tasks to be scheduled, establishing a jump index and a merge index, and allocating the set of tasks to be scheduled to the corresponding scheduling calculation models according to task characteristics; using the scheduling calculation models to perform resource scheduling calculations on the allocated tasks respectively, and obtaining a fusion scheduling scheme by means of weighted fusion; using the fusion output model to optimize and adjust the fusion scheduling scheme; executing the scheduling scheme and recording multi - dimensional evaluation indicators; and updating and optimizing the jump index and the merge index in a reinforcement learning manner until the multi - dimensional evaluation indicators meet the preset threshold requirements.

[0007] Among them, the physical resource requirement parameters include the demand quantity of computing processors, the demand quantity of memory, the demand quantity of storage space, the demand quantity of network bandwidth, and the demand quantity of energy consumption, and the information resource requirement parameters include task priority, task timeliness, task dependency relationship, task security level, and task fault tolerance level.

[0008] Among them, the multi - way selection model adopts a three - layer fully - connected neural network structure, with the number of neurons in each layer not exceeding 64, using the ReLU activation function, the input layer receiving task feature vectors, and the output layer using the Softmax function to generate task allocation probabilities; the scheduling calculation model adopts a four - layer convolutional neural network structure, with a convolutional kernel size of 3×3, 4 convolutional layers, the number of convolutional kernels in each layer not exceeding 32, and using a max - pooling layer for feature dimensionality reduction; the fusion output model adopts a two - layer recurrent neural network structure, with the number of hidden layer nodes not exceeding 32, and using LSTM units for sequence modeling.

[0009] Among them, the jump index constructs a task migration probability matrix based on the task load feature vector and the resource utilization vector, and the merge index calculates the weight coefficient based on the historical accuracy of each scheduling calculation model and the similarity of the current task characteristics.

[0010] Among them, the training data set of the multi - way selection model includes a task feature sequence and an optimal allocation label sequence, and the multi - way selection model is trained in a supervised learning manner, with the verification set accuracy threshold set to 0.9.

[0011] Among them, the scheduling calculation model uses a genetic algorithm for optimization and solution. The population size is set to 100, the number of iterations is set to 50, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the fitness function is the evaluation function for the corresponding task type.

[0012] Among them, the weighted fusion adopts a soft voting method. For discrete scheduling decisions, weighted majority voting is used, and for continuous parameters, weighted summation is used.

[0013] Among them, the multi-dimensional evaluation indicators include resource utilization rate, task completion time, energy consumption, service quality, and system stability. The preset threshold requirements include that the resource utilization rate is not less than 80%, the average task completion time does not exceed 1.2 times the expected time, the energy utilization efficiency is not less than 70%, the service quality satisfaction is not less than 90%, and the system stability index fluctuation does not exceed 10%.

[0014] The second aspect of the present invention provides a computer-readable storage medium. Program instructions are stored in the computer-readable storage medium. When the program instructions run on a computer, they are used to execute the above-mentioned method for scheduling computing resources of a productivity tool service station.

[0015] The third aspect of the present invention provides a system for scheduling computing resources of a productivity tool service station, including the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is set inside the system, and a microprocessor for executing the program instructions stored in the computer-readable storage medium is set inside the system.

[0016] Compared with the prior art, the present invention provides a method, medium, and system for scheduling computing resources of a productivity tool service station. The present invention proposes a total-score-total structure resource scheduling method based on a lightweight neural network. Through the collaborative work of a multi-way selection model, a scheduling calculation model, and a fusion output model, adaptive scheduling of diversified tasks is achieved. This method adopts a simplified neural network structure, and the number of neurons in each layer is controlled at a relatively low level, effectively reducing the model calculation complexity and storage overhead, enabling it to operate efficiently in an edge environment.

[0017] The present invention establishes a mapping relationship among task characteristics, resource status, and scheduling strategies by introducing two key indicators, namely the jump index and the merging index. The jump index improves the adaptability of the model to new task types by dynamically evaluating the migration probability of tasks among different scheduling models; the merging index realizes the optimal integration of scheduling schemes by evaluating the credibility of the output results of different scheduling models. This design enables the present invention to achieve precise scheduling of diverse tasks while maintaining a relatively small model size. At the same time, the present invention adopts a phased training strategy, first performing single-model pre-training and then end-to-end collaborative optimization, which not only ensures the basic performance of each model but also improves the overall scheduling effect.

[0018] The present invention successfully solves the technical problem that it is difficult for resource scheduling models to adaptively process diverse tasks in lightweight environments. Through the design of lightweight neural networks and the multi-model collaboration mechanism, both the computational efficiency of the model and the ability to adaptively process diverse tasks are ensured. The introduction of the jump index and the merging index provides a quantitative basis for task migration and result fusion in the model, further improving the accuracy and robustness of scheduling. Brief Description of the Drawings

[0019] Figure 1 It is a flowchart of the method of the present invention.

[0020] Figure 2 It is a heat map of resource demand distribution in Embodiment 2.

[0021] Figure 3 It is a curve graph showing the relationship between task priority and timeliness in Embodiment 2.

[0022] Figure 4 It is a convergence curve graph of accuracy during the training process in Embodiment 2.

[0023] Figure 5 It is a comparison graph of indicators between the traditional method and the method of the present invention in Embodiment 2.

[0024] Figure 6 It is a time-series change graph of stability indicators in Embodiment 2. Detailed Embodiments

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] As Figure 1 shown, it is a flowchart of a method for calculating resource scheduling of a productivity tool service station provided in the first aspect of the present invention. This method includes the following steps:

[0027] S01. Obtain the set of tasks to be scheduled in the service station. The set of tasks to be scheduled includes physical resource requirement parameters and information resource requirement parameters. The physical resource requirement parameters include the demand for computing processors, memory, storage space, network bandwidth, and energy consumption. The information resource requirement parameters include task priority, task timeliness, task dependencies, task security level, and task fault tolerance level;

[0028] S02. Establish multiple lightweight neural networks with a total - sub - total structure, including a multi - way selection model, multiple scheduling calculation models, and a fusion output model;

[0029] S03. Establish a jump index for the task resource requirement characteristics. The jump index constructs a task migration probability matrix based on the task load feature vector and the resource utilization vector;

[0030] S04. Train the multi - way selection model. The training dataset includes task feature sequences and optimal allocation label sequences, and the multi - way selection model is trained in a supervised learning manner;

[0031] S05. Use the multi - way selection model to classify the set of tasks to be scheduled, and allocate the set of tasks to be scheduled to the corresponding scheduling calculation models according to task characteristics;

[0032] S06. Establish a merging index. The merging index calculates the weight coefficient based on the historical accuracy of each scheduling calculation model and the similarity of the current task characteristics;

[0033] S07. Use the scheduling calculation models to perform resource scheduling calculations on the allocated tasks respectively, and each scheduling calculation model is optimized using the evaluation function corresponding to the task type;

[0034] S08. Perform weighted fusion on the initial scheduling plan according to the merging index to obtain a fusion scheduling plan, and the weighted fusion uses a soft voting method;

[0035] S09. Use the fusion output model to optimize and adjust the fusion scheduling plan, and resolve conflicts based on resource constraint conditions and task dependencies;

[0036] S10. Execute resource allocation according to the final scheduling plan, and record multi - dimensional evaluation indicators including resource utilization rate, task completion time, energy consumption, service quality, and system stability;

[0037] S11. Update the jump index and the merging index according to the multi - dimensional evaluation indicators, and perform online optimization using reinforcement learning;

[0038] S12. Repeat steps S05 to S11 until the multi-dimensional evaluation index meets the preset threshold requirements.

[0039] The multiple lightweight neural networks with the overall structure of "total - sub - total" are described as follows:

[0040] The multi - path selection model adopts a three - layer fully - connected neural network structure. The number of neurons in each layer does not exceed 64. The ReLU activation function is used. The input layer receives the task feature vector, and the output layer uses the Softmax function to generate the task assignment probability.

[0041] The scheduling calculation model adopts a four - layer convolutional neural network structure. The size of the convolutional kernel is 3×3, the number of convolutional layers is 4, the number of convolutional kernels in each layer does not exceed 32, and the max - pooling layer is used for feature dimensionality reduction.

[0042] The fusion output model adopts a two - layer recurrent neural network structure. The number of hidden layer nodes does not exceed 32, and the LSTM unit is used for sequence modeling.

[0043] The model training process includes two stages: single - model training and collaborative training.

[0044] The first stage: Use historical scheduling data to pre - train the three types of models respectively, and optimize the model parameters in a supervised learning manner.

[0045] The second stage: Adopt an end - to - end training method for collaborative optimization, use the multi - dimensional evaluation index as the training target, and optimize the parameters of the three types of models simultaneously.

[0046] The functional relationship between the model parameters and the resource scheduling parameters is as follows:

[0047] The task load calculation function is used to calculate the resource consumption index of the task. The input includes the demand for computing processors, memory, storage space, network bandwidth, and energy consumption. The output is the task load feature vector, and the task load feature vector is input into the multi - path selection model.

[0048] The resource status evaluation function is used to evaluate the current system resource status. The input includes the utilization rate of computing resources, memory resources, storage resources, network resources, and energy utilization efficiency. The output is the resource status feature vector, and the resource status feature vector is input into the multi - path selection model.

[0049] The task feature extraction function is used to extract the business features of the task. The input includes the task priority, task timeliness, task dependency, task security level, and task fault - tolerance level. The output is the task business feature vector, and the task business feature vector is input into the scheduling calculation model.

[0050] The scheduling scheme generation function is used to generate an initial scheduling scheme. The inputs include the task load feature vector, the resource status feature vector, and the task service feature vector, and the output is the initial scheduling scheme, which is input into the fusion output model;

[0051] The evaluation index calculation function is used to calculate the scheduling effect. The inputs include resource utilization rate, task completion time, task throughput, task success rate, and system stability, and the output is the comprehensive evaluation score, which is used to update the jump index and the merging index.

[0052] Among them, the task load feature vector is used to characterize the demand characteristics of tasks for physical resources; the resource status feature vector is used to characterize the current available resource status of the system; the task service feature vector is used to characterize the service attributes and constraint conditions of tasks; the jump index is used to quantify the probability value of task migration between different scheduling models; the merging index is used to quantify the credibility of the output results of different scheduling models; soft voting adopts a weighted average method based on probability distribution; conflict resolution adopts a resource competition adjustment method based on constraint conditions; calculating the resource utilization rate represents the usage ratio of the computing processor; the memory resource utilization rate represents the occupied ratio of the memory space; the storage resource utilization rate represents the occupied ratio of the storage space; the network resource utilization rate represents the occupied ratio of the network bandwidth; the energy utilization efficiency represents the amount of computation per unit energy consumption.

[0053] The specific implementation manners of the above steps are described in detail below. The specific implementation manner of step S01 is to obtain various types of task resource requirement parameters through a task feature extraction module. First, a task feature database is established, which includes physical resource requirement parameters such as the demand quantity of computing processors, memory, storage space, network bandwidth, and energy consumption. Among them, the demand quantity of computing processors is measured by the number of CPU cores, and the value range is from 1 to 32; the memory demand quantity is in GB, and the value range is from 1 to 256; the storage space demand quantity is in GB, and the value range is from 1 to 1024; the network bandwidth demand quantity is in Mbps, and the value range is from 1 to 1000; the energy consumption demand quantity is in watts, and the value range is from 10 to 500. Then, information resource requirement parameters such as task priority, task timeliness, task dependency relationship, task security level, and task fault tolerance level are obtained. Among them, the task priority is divided into 5 levels, represented by 1 to 5, and the larger the value, the higher the priority; the task timeliness is represented by the maximum tolerable delay time of the task, and the unit is seconds; the task dependency relationship is represented by a directed acyclic graph, which records the pre-order and post-order relationships between tasks; the task security level is divided into 3 levels, represented by 1 to 3, and the larger the value, the higher the security requirement; the task fault tolerance level is divided into 3 levels, represented by 1 to 3, and the larger the value, the higher the fault tolerance requirement. Finally, these parameters are organized into a standardized task description format for subsequent processing. This step provides a data basis for subsequent resource scheduling decisions by systematically collecting and organizing task requirement parameters.

[0054] The specific implementation manner of step S02 is to construct a lightweight neural network system with a total - sub - total hierarchical structure. First, a multiplexing selection model is designed, which adopts a three - layer fully - connected neural network structure. The number of neurons in the input layer is the dimension of the task feature vector, the number of neurons in the hidden layer is 64, the number of neurons in the output layer is the number of scheduling model categories, the activation function adopts the ReLU function, and the Softmax function is used in the output layer to achieve task classification. Then, multiple scheduling calculation models are constructed. Each model adopts a four - layer convolutional neural network structure. The input layer receives the task feature matrix. The first convolutional layer contains 32 3×3 convolutional kernels, the second convolutional layer contains 24 3×3 convolutional kernels, the third convolutional layer contains 16 3×3 convolutional kernels, and the fourth convolutional layer contains 8 3×3 convolutional kernels. After each layer of convolution, a 2×2 max - pooling layer is used for dimensionality reduction, and finally, a scheduling scheme is output through a fully - connected layer. Finally, a fusion output model is implemented, which adopts a two - layer recurrent neural network structure. The number of hidden layer nodes is 32, and the LSTM unit is used to process sequence data. The input layer receives the output results of multiple scheduling calculation models, and the final scheduling scheme is generated through time - series modeling. This step realizes the functional modules of task classification, scheduling calculation, and scheme fusion by constructing a hierarchical neural network system.

[0055] The specific implementation of step S03 is to construct a task migration probability matrix based on task characteristics and resource status. First, calculate the task load feature vector, which includes indicators such as CPU utilization rate, memory occupancy rate, storage space occupancy rate, network bandwidth occupancy rate, and energy consumption rate. The min-max normalization method is used to map each indicator to the interval from 0 to 1. Then, calculate the resource utilization vector, which includes the current usage of various resources, and also perform normalization processing. Next, construct the task migration probability matrix. The dimension of the matrix is the square of the number of scheduling models. The matrix elements represent the probability of a task migrating from one scheduling model to another. The initial probability value is calculated based on the similarity of task characteristics. Finally, define the jump index update rule to dynamically adjust the probability matrix based on the task execution effect. The adjustment step size is set to 0.1 to ensure that the probability value is between 0 and 1. This step realizes the dynamic switching between scheduling models by establishing a task migration mechanism.

[0056] The specific implementation of step S04 is to train a multi-way selection model using the supervised learning method. First, collect historical scheduling data, which includes task feature sequences and corresponding optimal scheduling model labels. The size of the data set is not less than 10,000 records. Then, preprocess the data, including feature normalization, missing value filling, and outlier handling. The missing values are filled with the mean value, and the outliers are detected and processed using the 3-sigma method. Next, set the training parameters, including the learning rate set to 0.001, the batch size set to 64, the number of training epochs set to 100, and use the cross-entropy loss function and the Adam optimizer. Finally, perform model training and use the 5-fold cross-validation method to evaluate the model performance. The accuracy threshold of the validation set is set to 0.9. If the threshold is not reached, increase the number of training epochs. This step enables the multi-way selection model to accurately classify different types of tasks through supervised learning training.

[0057] The specific implementation of step S05 is to use the trained multi-way selection model to classify the tasks to be scheduled. First, extract the feature vector of the tasks to be scheduled, including physical resource requirement parameters and information resource requirement parameters, and process the features using the same normalization method as the training data. Then, input the feature vector into the multi-way selection model to obtain the selection probability of each scheduling calculation model. The probability value represents the degree to which the task is suitable to be processed by this model. Next, assign the task to the corresponding scheduling calculation model according to the probability value. If the maximum probability value is lower than the threshold of 0.6, randomly assign the task to one of the top 3 models with the highest probability values. Finally, record the assignment results, including the task identifier, the selected scheduling model, and the selection probability, for subsequent evaluation and optimization. This step realizes the preliminary allocation of scheduling resources through task classification.

[0058] The specific implementation of step S06 is to construct a merging index for evaluating the credibility of the scheduling model. First, the historical accuracy rates of each scheduling calculation model are statistically analyzed, including indicators such as the rationality of resource allocation, the timeliness of task completion, and the system stability. The average accuracy rate of the last 1000 scheduling operations is calculated using the sliding window method. Then, the similarity between the current task and the historical tasks is calculated. Using the cosine similarity method, the task feature vectors are mapped to a high-dimensional space to calculate the cosine value of the included angle. Next, the weight coefficients are calculated based on the accuracy rate and the similarity. The accuracy rate weight is set to 0.6, and the similarity weight is set to 0.4, obtaining a merging index that reflects the credibility of each model. Finally, an update mechanism for the merging index is designed to dynamically adjust the weight coefficients according to the scheduling effect, and the adjustment step size is set to 0.05. Through constructing the merging index, this step provides a weight basis for the fusion of scheduling schemes.

[0059] The specific implementation of step S07 is to use multiple scheduling calculation models to process the assigned tasks in parallel. First, a specific evaluation function is defined for each type of task. For compute-intensive tasks, the processor utilization rate and computing efficiency are concerned; for storage-intensive tasks, the space utilization rate and read / write speed are concerned; for network-intensive tasks, the bandwidth utilization rate and transmission delay are concerned. Then, each scheduling calculation model uses the genetic algorithm for optimization and solution. The population size is set to 100, the number of iterations is set to 50, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the fitness function is the corresponding evaluation function. Next, an initial scheduling scheme is generated, including resource allocation strategies, task execution orders, and load balancing schemes. Finally, the feasibility of the scheduling scheme is verified to check whether it meets the resource constraints and task dependencies. Through parallel computing, this step generates multiple targeted scheduling schemes.

[0060] The specific implementation of step S08 is to use the soft voting method to fuse the output results of multiple scheduling models. First, the initial scheduling schemes generated by each model are obtained, including resource allocation schemes and task execution plans. Then, the voting weights of each model are calculated according to the merging index, and the weight values reflect the credibility of the model for the current task type. Next, the weighted average method is used to fuse the resource allocation schemes. For discrete scheduling decisions, the weighted majority voting is used, and for continuous parameters, the weighted summation is used. Finally, the fused scheme is normalized to ensure that the resource allocation meets the system constraint conditions. Through scheme fusion, this step generates a better scheduling strategy by integrating the advantages of multiple models.

[0061] The specific implementation of step S09 is to optimize and adjust the scheduling scheme using a fusion output model. First, check for resource conflicts in the fusion scheduling scheme, including computing resource competition, storage space conflicts, network bandwidth competition, etc. Then, formulate a conflict resolution strategy based on task priorities and dependencies. High-priority tasks are given priority to obtain resources, and tasks with dependencies are allocated resources according to the topological order. Next, use an LSTM network to model the task sequence, predict resource usage trends and potential conflicts, and set the prediction window to 10 time steps. Finally, make a forward-looking adjustment to the scheduling scheme based on the prediction results, and solve the resource competition problem through load balancing and task migration. This step ensures the executability of the scheduling scheme through conflict resolution.

[0062] The specific implementation of step S10 is to execute the final scheduling scheme and collect evaluation metrics. First, allocate system resources according to the scheduling scheme, including CPU core allocation, memory space allocation, storage device allocation, and network bandwidth allocation. Then, monitor the task execution process and record metrics such as resource utilization, task completion time, and energy consumption, with a sampling period of 1 second. Next, calculate the quality-of-service metrics, including task response time, throughput, success rate, etc., with a statistical period of 1 hour. Finally, evaluate the system stability, including metrics such as resource utilization fluctuations, task queue length, and system load, with an evaluation period of 1 day. This step provides a basis for model optimization by executing the scheduling scheme and collecting feedback data.

[0063] The specific implementation of step S11 is to optimize the scheduling model using a reinforcement learning method. First, construct the state space and action space. The state includes the system resource state and the task queue state, and the action includes scheduling decisions and resource allocation strategies. Then, design a reward function, comprehensively considering resource utilization, task completion time, energy consumption, quality of service, and system stability, with the weights of each metric being 0.3, 0.2, 0.15, 0.2, and 0.15 respectively. Next, use the deep Q-learning algorithm for online optimization, set the size of the experience replay pool to 10000, the target network update period to 100 steps, and the discount factor to 0.9. Finally, update the jump index and merge index according to the learning results, with an update period of 1000 steps. This step realizes the continuous optimization of the scheduling strategy through reinforcement learning.

[0064] The specific implementation of step S12 is to repeatedly execute the scheduling optimization process until the termination condition is met. First, set the target thresholds for evaluation metrics, including a resource utilization rate of no less than 80%, an average task completion time of no more than 1.2 times the expected time, an energy utilization efficiency of no less than 70%, a service quality satisfaction level of no less than 90%, and a system stability index fluctuation of no more than 10%. Then, repeatedly execute steps such as task classification, scheduling calculation, solution integration, conflict resolution, solution execution, and metric update. Next, evaluate whether the multi-dimensional metrics reach the target thresholds at the end of each cycle, with the evaluation cycle set to 1 hour. Finally, when all metrics meet the threshold requirements for 3 consecutive cycles, complete the current scheduling optimization process. This step ensures that the scheduling performance meets the system requirements through iterative optimization.

[0065] The overall structure contains the following specific model designs:

[0066] 1. Multi-way selection model (allocation model):

[0067] Principle: Use a three-layer fully connected neural network to implement task classification and allocation. The input layer receives task feature vectors, including parameters such as computing resource requirements, memory requirements, storage requirements, network requirements, and energy requirements; the hidden layer uses the ReLU activation function for feature transformation and combination; the output layer uses the Softmax function to calculate the probability distribution of tasks being suitable for allocation to each scheduling model. The allocation principle is based on pattern matching between task features and historical optimal allocation experiences, and the best processing models for different types of tasks are learned through supervised learning.

[0068] 2. Scheduling calculation model:

[0069] Includes the following specialized scheduling models:

[0070] Computation-intensive scheduling model: For scheduling computing resources such as CPUs and GPUs, optimize computing load balancing and processor utilization; Memory-intensive scheduling model: For scheduling memory resources, optimize memory allocation and recycling strategies; Storage-intensive scheduling model: For scheduling storage resources, optimize data access and caching strategies; Network-intensive scheduling model: For scheduling network resources, optimize bandwidth allocation and transmission routing; Hybrid scheduling model: Handle tasks with balanced multiple resource requirements;

[0071] These models all adopt a four-layer convolutional neural network structure and optimize the performance metrics they focus on through different evaluation functions.

[0072] 3. Fusion output model:

[0073] Adopt a two-layer LSTM recurrent neural network structure to serialize the output results of multiple scheduling models, and generate the final scheduling plan by considering temporal dependencies and resource constraints;

[0074] The specific implementation steps of the two stages of the training process are as follows:

[0075] The first stage: Single model training

[0076] 1. Multi-way selection model pre-training: Collect historical task data, establish the mapping relationship between task features and the optimal scheduling model; Standardize the task features to generate the training dataset; Set the training parameters: learning rate 0.001, batch size 64, number of training epochs 100; Use the cross-entropy loss function and the Adam optimizer for supervised learning; Adopt 5-fold cross-validation to evaluate the model performance.

[0077] 2. Scheduling calculation model pre-training: Classify historical data by task type, and prepare a dedicated training set for each scheduling model; Define the evaluation functions and performance metrics for various tasks; Use the genetic algorithm to optimize the parameter configuration of each scheduling model; Adopt the gradient descent method to minimize the resource scheduling loss function; Verify the scheduling effect of the model on specific task types;

[0078] 3. Fusion output model pre-training: Collect the historical output sequences of multiple scheduling models; Construct a time-series training dataset, including scheduling schemes and execution effects; Train the LSTM network to capture the time-series dependencies of scheduling decisions; Optimize the conflict detection and resolution strategies; Evaluate the overall performance of the fusion scheme;

[0079] The second stage: Collaborative training

[0080] 1. Initialization stage: Load the pre-trained model parameters; Establish the data interaction interface between models; Configure the end-to-end training environment;

[0081] 2. Joint optimization stage: Construct a unified multi-dimensional evaluation index system; Design an end-to-end loss function, comprehensively considering resource utilization rate, task completion time, energy consumption, service quality, and system stability; Establish a gradient backpropagation channel to achieve cross-model parameter updates; Adopt the reinforcement learning method to optimize the overall scheduling strategy;

[0082] 3. Dynamic adjustment stage: Monitor the system performance indicators in real time; Dynamically update the jump index and merge index; Adjust the model parameters according to the actual operation effect; Optimize the model collaboration mechanism;

[0083] 4. Convergence verification stage: Evaluate whether the overall scheduling performance reaches the preset goal; Test the stability of the model collaboration effect; Verify the system's processing ability for different types of tasks; Confirm the convergence of the training process.

[0084] Through this phased training method, the professionalism of each model in its respective field is ensured first, and then the overall scheduling effect is improved through collaborative training. This training strategy can effectively balance the professionalism and collaboration of the models, and improve the performance and stability of the entire scheduling system.

[0085] In a preferred specific embodiment of the present invention, for the industrial productivity service station scenario, the specific design of the overall - part - overall structure in this scenario is elaborated in detail below.

[0086] The multi - way selection model is implemented by a three - layer fully - connected neural network to achieve intelligent classification and allocation of production tasks. The input layer of this model receives multi - dimensional task feature vectors from the industrial site, including key production parameters such as workpiece machining accuracy requirements, material type codes, process parameter configurations, production batch sizes, delivery time windows, etc. These parameters are input into the neural network after being standardized. The hidden layer is set with 64 neurons, and the ReLU activation function is used for feature transformation and extraction. Through the optimization learning of the weight matrix, the feature patterns of different types of production tasks are gradually mastered. The output layer uses the Softmax function to calculate the probability distribution of tasks suitable for allocation to different professional scheduling models, including multiple directions such as precision machining scheduling, 3D printing scheduling, automated assembly scheduling, and robot control scheduling, so as to achieve intelligent diversion of production tasks. The allocation principle of this model is based on the deep learning of historical production data. By mining the mapping relationship between task features and the optimal scheduling mode, the accuracy and efficiency of task allocation are continuously improved.

[0087] The scheduling calculation model designs multiple specialized scheduling strategy mechanisms for different scenarios of industrial production. The precision machining scheduling model is mainly for the machining tasks of CNC equipment groups. This model uses a four - layer convolutional network structure, takes the machining process feature map as the input, and establishes a feature extraction layer containing 32 3×3 convolutional kernels, focusing on optimizing the balance between the utilization efficiency of machining equipment and machining accuracy; the 3D printing scheduling model focuses on processing additive manufacturing tasks, also using a four - layer convolutional structure, but its convolutional kernel quantity configuration is a shrinking structure of 24 - 16 - 8 - 4, focusing on optimizing the utilization rate of printing materials and printing quality indicators; the automated assembly scheduling model designs a convolutional structure based on timing constraints for the characteristics of pipeline operations, extracts assembly sequence features through 16 3×3 convolutional kernels, and optimizes the line balance rate; the robot control scheduling model uses a dynamically responsive convolutional architecture, configures 8 3×3 convolutional kernels, and focuses on optimizing the robot action sequence and collaboration efficiency. Each professional scheduling model is equipped with a corresponding evaluation function and optimization goal, and feature dimensionality reduction is performed through the max - pooling layer, and finally a targeted scheduling plan is generated.

[0088] The fusion output model adopts a two-layer LSTM recurrent neural network structure to achieve the intelligent integration of multiple professional scheduling schemes. The first layer of the LSTM network in this model contains 32 memory units, which are mainly responsible for processing the temporal dependencies between various production tasks, and capturing the correlation features of long-term and short-term production tasks through the gating mechanism. The second layer of the LSTM network is also configured with 32 memory units, which focus on processing resource competition and scheduling conflicts, and realizing the real-time optimization of resource allocation through the dynamic memory mechanism. The model training adopts a two-stage strategy: the first stage is the pre-training stage, where the historical scheduling sequences of various professional scheduling models are collected respectively, a training data set containing multi-dimensional features such as production plans, equipment status, and quality indicators is constructed, and the LSTM network parameters are optimized using supervised learning methods. The second stage is the co-training stage, where the outputs of multiple professional scheduling models are input into the LSTM network in real time, and the network parameters are continuously adjusted through reinforcement learning methods to optimize the overall scheduling effect. By comprehensively considering multiple dimensions such as production efficiency, quality requirements, and energy consumption, the model finally outputs a balanced and executable scheduling scheme.

[0089] Two extreme training scenarios are respectively optimized for the production peak period and the trough period. In the high-load scenario, first, the historical data of the peak production season is collected, and the extreme working condition data with equipment utilization rate exceeding 90%, serious order backlog, and tight delivery dates is screened out to construct a high-intensity training data set. During the training process, a relatively small learning rate of 0.0001 is adopted, and through the mini-batch training method with a batch size of 128, the model parameters are gradually optimized. At the same time, noise data is introduced to enhance the diversity of training samples, and the early stopping strategy is adopted to avoid overfitting. The key indicators for model verification focus on key indicators such as scheduling response time, resource utilization rate, and task completion rate. Through multiple rounds of iterative optimization, the stability and efficiency of the system under high-load conditions are ensured. The low-load scenario focuses on simulating the special situations in the off-peak production season and equipment maintenance period, and constructs a training data set with equipment utilization rate lower than 30%. A relatively large learning rate of 0.01 is adopted, and through the active learning method and the dynamic batch adjustment strategy, the sensitivity of the model to sparse tasks is improved. The training process focuses on optimizing energy efficiency and maintenance costs, and through gradient clipping and regularization constraints, the economy and reliability of the model under low-load conditions are ensured.

[0090] This overall structure design based on the industrial productivity service station, through a multi-level neural network architecture and targeted training strategies, realizes the intelligent and refined management of production scheduling. The system shows strong adaptability and robustness in practical applications, can effectively meet the scheduling requirements under different load conditions, and significantly improves production efficiency and resource utilization rate. Through continuous online learning and optimization, the system performance is continuously improved, providing strong support for the intelligent transformation of industrial production.

[0091] The following is a detailed description of the calculation process involved in the present invention.

[0092] 1. The calculation process of the task feature vector in step S01 is specifically represented as follows:

[0093]

[0094] In the formula, is the physical resource requirement vector; p 1 is the demand for computing processors, ranging from 1 to 32; p 2 is the memory demand, with a range of 1256; p 3 is the storage space demand, ranging from 1 to 1024; p 4 is the network bandwidth demand, and the range

[0095] is from 1 to 1000; p 5 is the energy consumption demand, ranging from 10 to 500; is the information resource requirement vector; i 1 is the task priority, ranging from 1 to 5; i 2 is the task timeliness; i 3 is the task dependency; i 4 is the task security level, ranging from 1 to 3; i 5 is the task fault tolerance level, ranging from 1 to 3.

[0096] 2. The neural network calculation process in step S02 is specifically represented as follows:

[0097] Forward propagation calculation of the multi-way selection model:

[0098] h 1 = ReLU(W 1 x + b 1 );

[0099] h 2 = ReLU(W 2 h 1 + b 2 );

[0100] y = Softmax(W 3 h 2 + b 3 );

[0101] In the formula, x is the input feature vector; h 1 , h 2 are the hidden layer outputs; y is the output layer probability distribution; W 1 , W 2 , W 3 are the weight matrices; b 1 , b 2 , b 3is the bias vector; ReLU(x) = max(0, x) is the activation function.

[0102] 3. The calculation of the jump index in step S03 is specifically expressed as follows:

[0103]

[0104] In the formula, M ij is the migration probability from model i to model j; f i , f j is the task feature vector; sim(f i , f j ) is the feature similarity; d ij is the resource load difference; t ij is the historical migration time interval; α, β, γ are weight coefficients and satisfy α + β + γ = 1; ∈ is the smoothing factor, with a value of 0.01; λ is the time decay coefficient, with a value of 0.1.

[0105] 4. The calculation of the merge index in step S06 is specifically expressed as follows:

[0106]

[0107] In the formula, w i is the weight of the i-th model; η i is the historical accuracy; s i is the task similarity; n is the number of models.

[0108] 5. The calculation of the evaluation function in step S07 is specifically expressed as follows:

[0109] Calculation-intensive task evaluation function:

[0110]

[0111] In the formula, f c is the calculated evaluation score; c used is the computing resources used; c total is the total computing resources; t exp is the expected execution time; t act is the actual execution time; e base is the baseline energy consumption; e act is the actual energy consumption; ω 1 , ω 2 , ω 3 are weight coefficients and satisfy ω 1 + ω 2 + ω 3 = 1.

[0112] Storage-intensive task evaluation function:

[0113]

[0114] In the formula, f s is the stored evaluation score; s free is the remaining storage space; s total is the total storage space; v act is the actual read / write speed; v max is the maximum read / write speed; l act is the actual access latency; l max is the maximum allowable latency; μ 1 , μ 2 , μ 3 are the weight coefficients and satisfy μ 1 +μ 2 +μ 3 = 1.

[0115] Network-intensive task evaluation function:

[0116]

[0117] In the formula, f n is the network evaluation score; b free is the remaining bandwidth; b total is the total bandwidth; r act is the actual transmission rate; r max is the maximum transmission rate; d act is the actual transmission latency; θ 1 , θ 2 , θ 3 are the weight coefficients and satisfy θ 1 +θ 2 +θ 3 = 1; σ is the delay sensitivity coefficient, with a value of 0.5.

[0118] 6. The soft voting fusion calculation in step S08 is specifically expressed as follows:

[0119]

[0120] In the formula, S final is the fused scheduling scheme; w i is the weight of the i-th model; S i is the scheduling scheme of the i-th model; n is the number of models; δ is a random perturbation term, following a normal distribution N(0, 0.1).

[0121] 7. The calculation of the reinforcement learning reward function in step S11 is specifically expressed as follows:

[0122]

[0123] Wherein, R is the total reward value; U is the resource utilization rate; T is the task completion time; E is the energy consumption; Q is the service quality index; S is the system stability index; φ 1 , φ 2 , φ 3 , φ 4 , φ 5 is the weight coefficient and satisfies φ 1 + φ 2 + φ 3 + φ 4 + φ 5 = 1.

[0124] The design principle and meaning of the above equation are as follows:

[0125] 1. The task feature vector is represented in vector form, which is convenient for subsequent neural network processing and feature extraction;

[0126] 2. The neural network model adopts a multi-layer structure, extracts features through non-linear transformation, and realizes task classification;

[0127] 3. The jump index considers feature similarity, resource load and time factors, and introduces exponential decay to reflect the time effect;

[0128] 4. The merging index adopts the Softmax form to ensure that the weights sum to 1 and are all positive values;

[0129] 5. The evaluation function considers the characteristics of different types of tasks and comprehensively considers factors such as resource utilization efficiency, time and energy consumption;

[0130] 6. The soft voting fusion introduces random perturbations to increase exploration and avoid local optima;

[0131] 7. The reward function adopts the weighted summation form to balance multiple optimization objectives.

[0132] The functional relationship between the model parameters and the resource scheduling parameters is described in detail below.

[0133] 1. The task load calculation function is specifically expressed as follows:

[0134]

[0135] Wherein, is the task load feature vector; is the physical resource requirement vector; is the normalization function; p 1 is the CPU demand; p 2 is the memory demand; p 3 is the storage demand; p 4 is the bandwidth demand; p 5 is the energy consumption demand; xmin , x max are the minimum and maximum values of the corresponding parameters, respectively.

[0136] 2. The resource status evaluation function is specifically expressed as follows:

[0137]

[0138] In the formula, is the resource status feature vector; is the resource utilization rate vector; u i is the current utilization rate of the i-th type of resource; d i is the resource load fluctuation degree; α i is the fluctuation sensitivity coefficient, and its value range is 0.1 to 1.0.

[0139] 3. The task feature extraction function is specifically expressed as follows:

[0140]

[0141] In the formula, is the task service feature vector; is the information resource requirement vector; θ i (x) is the feature transformation function; ε i is the random perturbation term, which follows the normal distribution N(0, 0.01); i 1 is the priority; i 2 is the timeliness; i 3 is the dependency; i 4 is the security level; i 5 is the fault tolerance level.

[0142] 4. The scheduling scheme generation function is specifically expressed as follows:

[0143]

[0144] In the formula, S is the scheduling scheme matrix; ω ij represents the allocation ratio of the i-th task using the j-th type of resource; m is the number of tasks; n is the number of resource types; is the load feature vector; is the status feature vector; is the service feature vector.

[0145] 5. The evaluation index calculation function is specifically expressed as follows:

[0146]

[0147] In the formula, E is the comprehensive evaluation score; U is the resource utilization rate; T is the task completion time; V is the task throughput; Q is the task success rate; S is the system stability; Umax For the maximum resource utilization rate; T min For the minimum completion time; V max For the maximum throughput; λ 1 , λ 2 , λ 3 , λ 4 , λ 5 is the weight coefficient and satisfies ξ is a random error term, following the normal distribution N(0, 0.05).

[0148] The design principles and meanings of these functional relationships are as follows:

[0149] 1. The task load calculation function adopts normalization processing to eliminate the influence of dimensions and facilitate the unified measurement of different resource requirements;

[0150] 2. The resource status evaluation function introduces an exponential decay term to reflect the impact of resource fluctuations on system stability;

[0151] 3. The task feature extraction function adds random perturbations to increase the robustness of the model and avoid overfitting;

[0152] 4. The scheduling scheme generation function adopts a matrix form to intuitively express the resource allocation relationship and facilitate subsequent optimization and adjustment;

[0153] 5. The evaluation index calculation function comprehensively considers multiple performance indexes and balances the importance of each index through weighted summation.

[0154] The second aspect of the present invention provides a computer-readable storage medium, in which program instructions are stored. When the program instructions run on a computer, they are used to execute the above-mentioned method for scheduling computing resources of a productivity tool service station.

[0155] The third aspect of the present invention provides a system for scheduling computing resources of a productivity tool service station, including the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged inside the system, and a microprocessor for executing the program instructions stored in the computer-readable storage medium is arranged inside the system.

[0156] Specifically, the principle of the present invention is as follows: The core technical principle of the present invention lies in constructing a lightweight neural network total - sub - total structure, and realizing hierarchical decision - making of resource scheduling through multiple neural network models with complementary functions. In model design, the multi - path selection model is responsible for task classification and distribution. It adopts a three - layer fully - connected network structure, with the number of neurons in each layer limited to within 64. Through the simplified network structure, it can achieve rapid task feature extraction and classification. The scheduling calculation model adopts a four - layer convolutional neural network, using small - size convolutional kernels and a limited number of feature maps, which can reduce the computational complexity while maintaining the effective ability to extract task features. The fusion output model adopts a two - layer recurrent neural network, and realizes the temporal integration of multiple scheduling schemes through streamlined LSTM units.

[0157] The present invention innovatively proposes two key indicators, namely the jump index and the merge index. The jump index is constructed based on the task load feature vector and the resource status feature vector. By calculating the similarity between the task features and each scheduling model, it evaluates the rationality of task migration. This design enables the system to adaptively adjust the task allocation strategy according to the dynamic changes of task features. The merge index evaluates the historical performance of each scheduling calculation model and the relevance of the current task, and assigns reasonable weights to the output results of different models to achieve the optimal fusion of scheduling schemes.

[0158] Through the phased training strategy, the present invention solves the problem of balancing the efficiency and effect of model training. In the pre - training stage, historical data is used to optimize each model respectively to establish basic feature extraction and decision - making capabilities. In the collaborative training stage, the entire network is optimized in an end - to - end manner to enhance the collaborative effect between models. This training method not only ensures the professionalism of each model but also realizes the improvement of the overall scheduling performance. Especially in terms of model parameter update, the present invention designs a series of feature extraction functions and evaluation index calculation functions, and establishes a clear mapping relationship between model parameters and resource scheduling parameters, providing reliable data support for the adaptive optimization of the model.

[0159] The following provides a specific Embodiment 1 of the present invention. The specific implementation of each step in this Embodiment 1 is described in detail as follows.

[0160] The specific implementation of step S01 is to obtain various task resource requirement parameters through the task feature extraction module. First, a task feature database is established using a feature extractor, which includes physical resource requirement parameters such as the demand for a computing processor, memory, storage space, network bandwidth, and energy consumption. These parameters constitute a physical resource demand vector:

[0161]

[0162] In the formula, p 1To calculate the processor demand, measured by the number of CPU cores, with a value range from 1 to 32; p 2 To calculate the memory demand, measured in GB, with a value range from 1 to 256; p 3 To calculate the storage space demand, measured in GB, with a value range from 1 to 1024; p 4 To calculate the network bandwidth demand, measured in Mbps, with a value range from 1 to 1000; p 5 To calculate the energy consumption demand, measured in watts, with a value range from 10 to 500. Then obtain the information resource demand parameters to form an information resource demand vector:

[0163]

[0164] In the formula, i 1 is the task priority, divided into 5 levels, represented by 1 to 5, and the larger the value, the higher the priority; i 2 is the task timeliness, represented by the maximum tolerable delay time of the task, with the unit of seconds; i 3 is the task dependency, represented by a directed acyclic graph, recording the pre-order and post-order relationships between tasks; i 4 is the task security level, divided into 3 levels, represented by 1 to 3, and the larger the value, the higher the security requirement; i 5 is the task fault tolerance level, divided into 3 levels, represented by 1 to 3, and the larger the value, the higher the fault tolerance requirement. Then use the task load calculation function to perform feature transformation on the physical resource demand vector:

[0165]

[0166] In the formula, is the task load feature vector; is the normalization function; x min , x max are the minimum and maximum values of the corresponding parameters respectively. Finally, use the task feature extraction function to perform feature transformation on the information resource demand vector:

[0167]

[0168] In the formula, is the task service feature vector; θ i (x) is the feature transformation function; ε i is the random perturbation term, following the normal distribution N(0, 0.01). This step uses feature extraction and normalization techniques to realize the standardized representation of task features, providing a data basis for subsequent resource scheduling decisions.

[0169] The specific implementation of step S02 is to construct a lightweight neural network system with a hierarchical structure of general - total - general. First, a multiplexer model is designed, which adopts a three - layer fully - connected neural network structure. The number of neurons in the input layer is equal to the dimension of the task feature vector, the number of neurons in the hidden layer is 64, and the number of neurons in the output layer is equal to the number of scheduling model categories. The calculation process of each layer is as follows:

[0170] h 1 =ReLU(W 1 x + b 1 ),

[0171] h 2 =ReLU(W 2 h 1 + b 2 ),

[0172] y=Softmax(W 3 h 2 + b 3 ),

[0173] where x is the input feature vector; h 1 , h 2 are the outputs of the hidden layer; y is the probability distribution of the output layer; W 1 , W 2 , W 3 are weight matrices; b 1 , b 2 , b 3 are bias vectors; ReLU(x)=max(0, x) is the activation function. Then, multiple scheduling calculation models are constructed. Each model adopts a four - layer convolutional neural network structure. The input layer receives the task feature matrix. The first convolutional layer contains 32 3×3 convolutional kernels, the second convolutional layer contains 24 3×3 convolutional kernels, the third convolutional layer contains 16 3×3 convolutional kernels, and the fourth convolutional layer contains 8 3×3 convolutional kernels. After each layer of convolution, a 2×2 max - pooling layer is used for dimensionality reduction. The calculation process of each layer of convolution is as follows:

[0174] C l =Pool(ReLU(Conv(X l-1 , K l )+ b l ))

[0175] where C l is the output of the l - th layer of convolution; X l-1 is the feature map of the previous layer; K l is the convolutional kernel parameter; b l is the bias term; Conv is the convolution operation; Pool is the max - pooling operation. Finally, the scheduling scheme is output through the fully - connected layer. This step realizes the task classification and scheduling calculation functions based on deep - learning technology.

[0176] The specific implementation of step S03 is to construct a task migration probability matrix based on task characteristics and resource status. First, obtain the current resource utilization status, and use a resource status evaluation function to calculate the resource status feature vector:

[0177]

[0178] In the formula, is the resource status feature vector; is the resource utilization rate vector; u i is the current utilization rate of the i-th type of resource; d i is the resource load fluctuation degree; α i is the fluctuation sensitivity coefficient, and its value range is 0.1 to 1.0. Then, construct a task migration probability matrix according to the task feature similarity and resource load difference:

[0179]

[0180] In the formula, M ij is the migration probability from model i to model j; f i , f j is the task feature vector; sim(f i , f j ) is the feature similarity; d ij is the resource load difference; t ij is the historical migration time interval; α, β, γ are weight coefficients and satisfy α + β + γ = 1; ∈ is the smoothing factor, and its value is 0.01; λ is the time decay coefficient, and its value is 0.1. This step realizes the dynamic switching between scheduling models through the task migration mechanism.

[0181] The specific implementation of step S04 is to train a multi-way selection model using the supervised learning method. First, collect historical scheduling data, including task feature sequences and corresponding optimal scheduling model labels. The size of the data set is not less than 10,000 records, and use the cross-entropy loss function to calculate the model prediction error:

[0182]

[0183] In the formula, L is the loss function value; N is the number of samples; M is the number of model categories; y ij is the true label; is the predicted probability. Then use the backpropagation algorithm based on gradient descent to update the model parameters:

[0184]

[0185] In the formula, are the weights and biases of the l-th layer at the t-th iteration; η is the learning rate, set to 0.001. Finally, 5-fold cross-validation is used to evaluate the model performance, and the validation set accuracy threshold is set to 0.9. This step is trained through supervised learning to enable the multi-way selection model to accurately classify different types of tasks.

[0186] The specific implementation of step S05 is to use the trained multi-way selection model to classify the tasks to be scheduled. First, obtain the feature vector of the task to be scheduled, and normalize the physical resource demand vector to obtain the load feature vector:

[0187]

[0188] where is the task load feature vector; φ i (x) is the normalization function; p i is the physical resource demand parameter. Then, perform forward propagation calculation through the multi-way selection model:

[0189] h 1 = ReLU(W 1 x + b 1 ),

[0190] h 2 = ReLU(W 2 h 1 + b 2 ),

[0191] y = Softmax(W 3 h 2 + b 3 ),

[0192] where h 1 , h 2 are the hidden layer outputs; y is the model selection probability distribution; W l , b l are the weight and bias parameters. Then, select the scheduling calculation model according to the probability distribution:

[0193] k = argmax i {y i | y i ≥ θ th},

[0194] where k is the selected model index; y i is the selection probability of the i-th model; θ th is the probability threshold, set to 0.6. This step realizes the preliminary allocation of scheduling resources through task classification.

[0195] The specific implementation of step S06 is to construct a merging index for evaluating the credibility of the scheduling model. First, the historical accuracy of each scheduling calculation model is statistically analyzed, and the average accuracy of the last 1000 schedules is calculated using the sliding window method:

[0196]

[0197] In the formula, η i is the historical accuracy of the i-th model; W is the window size, set to 1000; a ij is the accuracy index of the j-th schedule. Then, the similarity between the current task and the historical tasks is calculated:

[0198]

[0199] In the formula, s i is the similarity value; is the current task feature vector; is the historical task feature vector. Then, the merging index is calculated based on the accuracy and similarity:

[0200]

[0201] In the formula, w i is the weight of the i-th model; n is the number of models. This step provides a weight basis for the fusion of the scheduling scheme by constructing the merging index.

[0202] The specific implementation of step S07 is to use multiple scheduling calculation models to process the assigned tasks in parallel. First, the corresponding evaluation function is selected according to the task type. For compute-intensive tasks, use:

[0203]

[0204] For storage-intensive tasks, use:

[0205]

[0206] For network-intensive tasks, use:

[0207]

[0208] The meanings of the parameters in the formula are the same as those described above. Then, the genetic algorithm is used to optimize and solve the scheduling scheme. The population size is set to 100, the number of iterations is set to 50, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. This step generates multiple targeted scheduling schemes through parallel computing.

[0209] The specific implementation of step S08 is to use the soft voting method to fuse the output results of multiple scheduling models. First, obtain the initial scheduling schemes generated by each scheduling calculation model:

[0210]

[0211] Wherein, S is the scheduling scheme matrix; ω ij represents the allocation ratio of the i-th task using the j-th type of resource; m is the number of tasks; n is the number of resource types. Then, weighted fusion is performed based on the merging index:

[0212]

[0213] Wherein, S final is the fused scheduling scheme; w i is the weight of the i-th model; S i is the scheduling scheme of the i-th model; δ is a random perturbation term, following a normal distribution N(0, 0.1). Through scheme fusion in this step, the advantages of multiple models are integrated to generate a better scheduling strategy.

[0214] The specific implementation manner of step S09 is to optimize and adjust the scheduling scheme using the fused output model. First, a resource conflict detection matrix is constructed to check the resource competition situation in the fused scheduling scheme:

[0215]

[0216] Wherein, C ij is an element of the conflict detection matrix. Then, a task priority matrix is constructed according to the task priority and dependency relationship:

[0217]

[0218] Wherein, P ij is an element of the priority matrix. Next, a recurrent neural network is used to predict the resource usage trend:

[0219] h t = f(W xh x t + W hh h t-1 + b h ),

[0220] y t = g(W hy h t + b y ),

[0221] Wherein, h t is the hidden state; x t is the input sequence; y t is the predicted output; W xh , W hh , W hy are weight matrices; b h , by is the bias vector; f and g are activation functions. Finally, adjust the resource allocation strategy based on the prediction results:

[0222] S opt = S final ⊙ (I - C)+ΔS,

[0223] In the formula, S opt is the optimized scheduling scheme; S final is the integrated scheduling scheme; I is the identity matrix; C is the conflict matrix; ΔS is the adjustment amount. This step ensures the executability of the scheduling scheme through conflict resolution.

[0224] The specific implementation of step S10 is to execute the final scheduling scheme and collect evaluation indicators. First, calculate the resource utilization rate indicator:

[0225]

[0226] In the formula, U is the average resource utilization rate; T is the statistical period; n is the number of resource types; r i，t is the usage amount of the i-th type of resource at the t-th moment; R i is the total amount of the i-th type of resource. Then, calculate the task completion time indicator:

[0227]

[0228] In the formula, T comp is the average completion time; M is the number of tasks; are the start and end times of the task respectively. Then, calculate the quality of service indicator:

[0229]

[0230] In the formula, Q is the quality of service score; N succ , N total are the number of successfully completed and total tasks respectively; V act , V exp are the actual and expected throughputs respectively; D act , D max are the actual and maximum delays respectively; α 1 , α 2 , α 3 are the weight coefficients. Finally, calculate the system stability indicator:

[0231]

[0232] In the formula, S is the stability score; σ U is the standard deviation of resource utilization; L is the task queue length; F is the failure rate; β 1 , β 2 , β3 is the weight coefficient. This step provides feedback data for model optimization through performance evaluation.

[0233] The specific implementation of step S11 is to optimize the scheduling model using reinforcement learning. First, define the state space and action space. The state vector is:

[0234] s t =[U t ,Q t ,L t ,E t ,V t ) T ,

[0235] where s t is the system state vector; U t is the resource utilization rate; Q t is the quality of service; L t is the load level; E t is the energy consumption level; V t is the task queue length. The action vector is:

[0236] a t =[d 1 ,d 2 ,…,d K ) T ,

[0237] where a t is the scheduling decision vector; d k is the k-th scheduling parameter. Then calculate the reward value:

[0238]

[0239] The meanings of the parameters in the formula are the same as those described above. Then use the deep Q-learning algorithm to update the value function:

[0240] Q(s t ,a t )=Q(s t ,a t )+α[r t +γmax a′ Q(s t+1 ,a′)-Q(s t ,a t )],

[0241] where Q(s t ,a t ) is the state-action value function; α is the learning rate; γ is the discount factor; r t is the immediate reward. This step realizes the continuous optimization of the scheduling strategy through reinforcement learning.

[0242] The specific implementation method of step S12 is to execute the scheduling optimization process cyclically until the termination condition is met. First, define the target value of the evaluation index:

[0243] G=[g 1 , g 2 , g 3 , g 4 , g 5 ] T ,

[0244] In the formula, g 1 is the resource utilization rate target, no less than 0.8; g 2 The time target for task completion shall not exceed 1.2 times the expected time; g 3 The energy utilization rate target is no less than 0.7; g 4 The service quality target is no less than 0.9; g 5 The system stability target is set, and the fluctuation does not exceed 0.1. Then calculate the difference between the current performance index and the target value:

[0245] D=||XG|| 2 ,

[0246] In the formula, D is the performance difference; X is the current indicator vector; G is the target value vector. Finally, the termination condition is determined:

[0247] If D≤∈ is satisfied for k consecutive cycles, the optimization is terminated.

[0248] Where k is the number of cycles, which is set to 3; ∈ is the error threshold, which is set to 0.05. This step ensures that the scheduling performance meets the system requirements through iterative optimization.

[0249] In order to better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A cloud computing data center uses the productivity tool service station computing resource scheduling method of the present invention to manage resources. The data center has 1,000 servers, each of which is equipped with a 32-core CPU, 256GB memory, 4TB storage space, and 10GBps network bandwidth. In the actual operation process, it faces problems such as uneven resource utilization, delayed task response, and excessive energy consumption. The specific process of implementing this method is described as follows.

[0250] First, we counted the resource requirement parameters of various tasks in the past month, as shown in Table 1:

[0251] Table 1 Task resource requirement statistics

[0252]

[0253] Figure 2It is a heatmap of resource demand distribution, showing the demand levels of different types of tasks for various resources. The horizontal axis represents 5 different resource metrics: CPU utilization rate, memory usage rate, storage occupancy rate, bandwidth usage rate, and energy efficiency; the vertical axis represents 4 different task types: compute-intensive, storage-intensive, network-intensive, and hybrid. The values in the heatmap represent the normalized resource demand intensity, and the darker the color, the stronger the demand. It can be seen from the figure that compute-intensive tasks have the highest demand for CPU and energy, storage-intensive tasks have prominent demands for memory and storage, network-intensive tasks have relatively high demands for bandwidth, and the resource demands of hybrid tasks are relatively balanced.

[0254] Construct a physical resource demand vector according to the task characteristics and an information resource demand vector For a typical compute-intensive task, its physical resource demand vector is:

[0255] Next, establish a task priority evaluation system, as shown in Table 2:

[0256] Table 2 Task Priority Evaluation System

[0257] Priority Timeliness (seconds) Security level Fault tolerance level Usage scenario 5 <60 3 3 Real-time processing of critical services 4 60-300 2-3 2-3 Processing of important services 3 300-1800 2 2 Processing of regular services 2 1800-7200 1-2 1-2 Batch processing tasks 1 >7200 1 1 Background tasks

[0258] Figure 3 Shows the relationship curve between task priority and timeliness. The horizontal axis represents the timeliness requirement of the task (in seconds), and the vertical axis represents the corresponding priority score (P score ). There are three curves in the figure, representing high, medium, and low priority tasks respectively, distinguished by different line types. Each curve shows an exponential decay characteristic, indicating that as the timeliness requirement decreases, the priority score gradually decreases. Among them, the decay rate of high-priority tasks is the fastest, indicating that they are the most sensitive to timeliness requirements.

[0259] Based on the above evaluation system, the information resource demand vector of compute-intensive tasks is:

[0260] According to the resource demand characteristics, use a multiplexer model to classify tasks. After training with 10,000 historical data, the accuracy rate on the validation set reached 0.92. Figure 4 Depicts the convergence curves of the accuracy rates of four classifiers during the training process. The horizontal axis represents the number of training rounds, and the vertical axis represents the classification accuracy rate (Acc rate)。Curves of different colors and line types represent classifiers for different types of tasks. As can be seen from the figure, as the number of training rounds increases, the accuracy of each classifier shows a convergent trend, and among them, the final accuracy of the compute-intensive classifier is the highest, reaching 92%. For the above compute-intensive task, the classification probability distribution output by the model is: y = [0.85, 0.08, 0.05, 0.02] T , corresponding to the compute-intensive, storage-intensive, network-intensive, and hybrid schedulers respectively.

[0261] Then, the historical performance metrics of each scheduling model are counted, as shown in Table 3:

[0262] Table 3 Historical Performance Statistics of Scheduling Models

[0263] Scheduling model Resource utilization rate Task completion rate Energy efficiency Average latency (seconds) Stability index Computation-intensive 0.85 0.95 0.78 120 0.92 Storage-intensive 0.82 0.93 0.82 180 0.90 Network-intensive 0.80 0.92 0.85 150 0.88 Hybrid 0.78 0.90 0.80 200 0.85

[0264] Based on the historical performance, the merging index of each model is calculated: w = [0.45, 0.25, 0.20, 0.10] T .

[0265] For this compute-intensive task, the resource allocation plan generated by its scheduling calculation model is:

[0266]

[0267] After resource conflict detection and optimization adjustment, the finally executed scheduling plan achieved the performance metrics shown in Table 4 within a one-hour operation cycle:

[0268] Table 4 Execution Effect of Scheduling Plan

[0269] Evaluation metrics Before optimization After optimization Improvement ratio Resource utilization rate 0.65 0.85 30.8％ Task completion time (seconds) 240 150 37.5％ Energy utilization efficiency 0.60 0.75 25.0％ Service quality score 0.75 0.92 22.7％ System stability 0.70 0.88 25.7％

[0270] Figure 5 The performance of the traditional method and the method of the present invention in five key performance metrics is compared through a bar chart. The horizontal axis represents different performance metrics, and the vertical axis represents the normalized score. It can be intuitively seen from the juxtaposed bar charts that the method of the present invention is superior to the traditional method in each metric, especially in terms of resource utilization and service quality. Through continuous optimization by reinforcement learning, the system achieved the preset performance goals after running for one week: the resource utilization rate was maintained above 85%, the average task completion time was shortened by 20% compared with the expectation, the energy utilization efficiency was increased to 75%, the service quality satisfaction reached 92%, and the system stability fluctuation was controlled within 8%.

[0271] Figure 6 shows the temporal variation of the stability metric of the system during 24 hours of operation. The horizontal axis represents the running time (hours), and the vertical axis represents the system stability metric (S index)。The figure contains two curves, representing the method of the present invention and the traditional method respectively, and the fluctuation range is shown through a translucent area. It can be seen from the figure that the stability index of the method of the present invention is not only higher in the overall level, but also has a smaller fluctuation range, indicating that the system runs more stably. Compared with the traditional resource scheduling method, the present invention has the following advantages: The traditional method usually adopts fixed scheduling strategies, such as first-come-first-served, priority scheduling, etc., and cannot dynamically adjust the resource allocation strategy according to the task characteristics and system status. While the present invention realizes the automatic extraction and classification of task characteristics by constructing a lightweight neural network system, and can dynamically optimize the scheduling strategy according to the historical performance. In practical applications, the resource utilization rate of the traditional method is generally between 50% and 60%, the task response time fluctuates greatly, and the energy utilization efficiency is generally low. While the present invention improves the resource utilization rate to more than 85%, shortens the task response time by 37.5%, and improves the energy utilization efficiency by 25%, significantly improving the operation efficiency of the data center. At the same time, the present invention introduces a jump index and a merge index, and effectively improves the accuracy and robustness of the scheduling decision through multi-model collaboration and soft voting fusion. In addition, the present invention adopts a reinforcement learning method to continuously optimize the scheduling strategy, can adapt to the dynamic changes of the business load, and maintain the efficient and stable operation of the system.

[0272] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 5 below.

[0273] Table 5 Variable Explanation Table

[0274]

[0275]

[0276] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered by the protection scope of the present invention.

Claims

1. A method for scheduling computing resources of a productivity tool service station, characterized in that: The following steps are involved: Acquire a set of tasks to be scheduled in the service station, wherein the set of tasks to be scheduled includes physical resource requirement parameters and information resource requirement parameters; Establish multiple lightweight neural networks of the overall-division-overall structure, including a multi-path selection model, multiple scheduling calculation models, and a fusion output model; Performing supervised learning training on the multi-path selection model based on task characteristics; The multi-path selection model is used to classify the set of tasks to be scheduled, a jump index and a merge index are established, and the set of tasks to be scheduled is assigned to a corresponding scheduling calculation model according to the task characteristics; the scheduling calculation model is used to perform resource scheduling calculations on the assigned tasks respectively, and a fusion scheduling scheme is obtained by a weighted fusion method; the fusion output model is used to optimize and adjust the fusion scheduling scheme; Execute the scheduling plan and record the multi-dimensional evaluation index; use reinforcement learning to update and optimize the jump index and the merge index until the multi-dimensional evaluation index reaches a preset threshold requirement.

2. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The physical resource requirement parameters include computing processor requirements, memory requirements, storage space requirements, network bandwidth requirements, and energy consumption requirements; the information resource requirement parameters include task priority, task timeliness, task dependency, task safety level, and task fault tolerance level.

3. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The multi-path selection model adopts a three-layer fully connected neural network structure, with the number of neurons in each layer not exceeding 64, using the ReLU activation function, the input layer receiving the task feature vector, and the output layer using the Softmax function to generate the task assignment probability; the scheduling calculation model adopts a four-layer convolutional neural network structure, the convolution kernel size is 3×3, the number of convolution layers is 4, the number of convolution kernels in each layer is not exceeding 32, and the maximum pooling layer is used for feature dimensionality reduction; the fusion output model adopts a two-layer recurrent neural network structure, the number of hidden layer nodes is not exceeding 32, and LSTM units are used for sequence modeling.

4. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The jump index constructs a task migration probability matrix based on the task load feature vector and the resource utilization vector, and the merge index calculates the weight coefficient based on the historical accuracy of each scheduling calculation model and the similarity of the current task feature.

5. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The training data set of the multi-path selection model includes a task feature sequence and an optimal allocation label sequence. The multi-path selection model is trained using a supervised learning method, and the validation set accuracy threshold is set to 0.

9.

6. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The scheduling calculation model is optimized and solved by a genetic algorithm, with the population size set to 100, the number of iterations set to 50, the crossover probability set to 0.8, the mutation probability set to 0.1, and the fitness function being the evaluation function of the corresponding task type.

7. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The weighted fusion adopts a soft voting method, using weighted majority voting for discrete scheduling decisions and weighted summation for continuous parameters.

8. The method for scheduling computing resources of a productivity tool service station according to claim 1, characterized in that: The multi-dimensional evaluation indicators include resource utilization, task completion time, energy consumption, service quality, and system stability. The preset threshold requirements include resource utilization not less than 80%, average task completion time not more than 1.2 times the expected time, energy utilization efficiency not less than 70%, service quality satisfaction not less than 90%, and system stability indicator fluctuation not more than 10%.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the productivity tool service station computing resource scheduling method described in any one of claims 1-8.

10. A productivity tool service station computing resource scheduling system, characterized in that: The system comprises the computer-readable storage medium as claimed in claim 9, wherein the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images