Measurement and control system adaptive adjustment method and system based on deep learning

Abstract

The invention discloses a measurement and control system adaptive adjustment method and system based on deep learning, and relates to the technical field of intelligent regulation and control, and the method comprises the steps: collecting the real-time operation data of a measurement and control system, constructing the real-time operation data into a system state vector, carrying out the feature extraction of a task in the measurement and control system based on the system state vector, and carrying out the adaptive adjustment of the measurement and control system. The method comprises the steps of classifying tasks by adopting a support vector machine, predicting resource demand quantity for different types of tasks, dynamically adjusting task priorities based on the resource demand quantity, generating task priority information, and performing global resource allocation optimization by adopting a genetic algorithm through a macroscopic scheduling layer; real-time scheduling decision making is carried out through the microcosmic scheduling layer by adopting a reinforcement learning model, and a specific resource allocation instruction is generated based on a resource allocation scheme output after optimization of the macroscopic scheduling layer. The utilization rate of system resources is improved, and task response delay is reduced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and specifically to an adaptive adjustment method and system for a measurement and control system based on deep learning. Background Art

[0002] Modern measurement and control systems are facing an increasingly complex task environment and variable workloads, and traditional static resource allocation methods can no longer meet the requirements of efficient operation. The resources of the measurement and control system are limited, while the number and types of tasks continue to increase, resulting in a common phenomenon of resource contention. Especially in some key fields, the system often faces challenges such as a sudden increase in tasks and fluctuations in resource requirements. Traditional fixed-priority scheduling strategies and pre-allocated resource methods are difficult to cope with these dynamic changes, and are prone to problems such as delayed response of high-priority tasks and low resource utilization.

[0003] Therefore, the present invention proposes an adaptive adjustment method and system for a measurement and control system based on deep learning. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. For this purpose, the present invention proposes an adaptive adjustment method and system for a measurement and control system based on deep learning, which improves the system resource utilization rate and reduces task response latency.

[0005] To achieve the above object, an adaptive adjustment system for a measurement and control system based on deep learning is proposed, including a resource status perception module, a priority adjustment module, a two-layer scheduling decision module, and a resource scheduling execution module; wherein, each module is electrically connected; The resource status perception module collects real-time operation data of the measurement and control system, preprocesses the real-time operation data, and constructs the preprocessed real-time operation data into a system state vector; and sends the system state vector to the priority adjustment module and the two-layer scheduling decision module as the input of the resource scheduling decision; The priority adjustment module receives the system state vector, extracts features of the tasks in the measurement and control system, classifies the tasks using a support vector machine, predicts the resource requirements for different types of tasks, and dynamically adjusts the task priorities based on the resource requirements; and transmits the adjusted task priority information to the two-layer scheduling decision module; The double - layer scheduling decision - making module consists of a macro - scheduling layer and a micro - scheduling layer. Among them, the macro - scheduling layer receives the system state vector and the adjusted task priority information, and uses a genetic algorithm to optimize the global resource allocation. The micro - scheduling layer, based on the optimization results of the macro - scheduling layer, uses a reinforcement learning model to make real - time scheduling decisions, and takes the resource allocation scheme of the macro - scheduling layer as a constraint condition for the micro - scheduling layer in the form of top - down guidance feedback. The double - layer scheduling decision - making module sends the optimized resource allocation scheme and the results of real - time scheduling decisions to the resource scheduling execution module. The resource scheduling execution module receives the resource allocation scheme output by the double - layer scheduling decision - making module and generates specific resource allocation instructions.

[0006] The resource status perception module collects the real - time operation data of the measurement and control system, including the following steps: Step 11: Collect real - time operation data. The real - time operation data includes CPU usage rate, memory occupancy rate, network bandwidth usage rate, storage space usage rate, and task queue status information. Step 12: Construct the real - time operation data into a system state vector through feature engineering. The priority adjustment module extracts the features of tasks in the measurement and control system, including the following steps: Step 21: Extract the task features of each task from the task queue status information in the system state vector. Step 22: Use a support vector machine to classify the extracted task features, and divide the tasks into three categories: compute - intensive, I / O - intensive, and hybrid. Step 23: For the different types of tasks after classification, use a linear regression model to predict their resource requirements.

[0007] The way for the priority adjustment module to establish a dynamic task priority adjustment mechanism is: Extract the waiting time in the task queue status information in the system state vector, and adaptively adjust the task priority according to the waiting time and the system state vector. The adaptive adjustment of task priority includes: The priority adjustment module consists of a resource competition analysis unit, a priority calculation unit, and a priority update unit.

[0008] The resource competition analysis unit calculates the current system resource competition degree based on the system resource status sub - vector and the network resource status sub - vector in the system state vector. The system resource competition degree is described by a resource competition index, which is calculated by weighted calculation of factors such as CPU usage rate, memory occupancy rate, network bandwidth usage rate, and storage space usage rate to quantify the tightness of system resources.

[0009] Among them, the priority calculation unit calculates the dynamic priority adjustment factor for each task based on the waiting time of the task, the resource competition index, and the resource demand of the task. The dynamic priority adjustment factor is calculated using an adaptive weighting method; The method for calculating the dynamic priority adjustment factor implemented by the priority calculation unit includes the following sub-steps: Step 221: Construct a waiting time compensation function. As the waiting time of the task increases, the compensation value increases non-linearly to prevent the task from starving for a long time; Step 222: Construct a resource demand penalty function. Appropriately reduce the priority of tasks with large resource demands, but set an upper limit to avoid high-resource-demand tasks from never being executed; Step 223: Construct a system load adaptive function. According to the current system load status, dynamically adjust the weights of waiting time compensation and resource demand penalty; Step 224: Weight and fuse the task priority, waiting time compensation value, resource demand penalty value, and system load adaptive factor in the queue status information to calculate the final dynamic priority adjustment factor; Among them, the priority update unit receives the dynamic priority adjustment factor and updates the priority information of the task.

[0010] The macro-scheduling layer in the two-layer scheduling decision module adopts a genetic algorithm, which includes the following steps: Step 31: Receive the system state vector and the adjusted task priority information, and design a chromosome encoding method for the global resource allocation optimization of the macro-scheduling layer to represent the resource allocation scheme; The design of the chromosome encoding method includes the following sub-steps: Step 311: Design a chromosome structure based on integer encoding to represent the mapping relationship between the predicted task resource demand and the system resources; Step 312: Based on the chromosome structure, design a chromosome segmentation strategy to encode the allocation schemes of different types of resources respectively; Step 313: Based on the segmentation strategy, design a constraint condition encoding method composed of a hard constraint encoding unit, a soft constraint encoding unit, and a constraint priority unit to ensure that resource allocation meets system limitations; Step 314: Based on the task priority information, design a priority encoding method composed of a priority mapping unit, an execution order encoding unit, and an importance degree encoding unit to reflect the execution order and importance degree of the tasks; Step 32: Design a fitness function for the global resource allocation optimization of the macro scheduling layer, which is jointly composed of an objective function construction unit, a weight allocation unit, and a normalization processing unit, and comprehensively consider the system throughput and load balancing degree; Design a corresponding fitness function for the genetic algorithm, including the following sub-steps: Step 321: Based on the chromosome coding method, construct a system throughput evaluation function jointly composed of a task execution time estimation unit, a parallelism analysis unit, and a throughput calculation unit, and calculate the number of tasks completed per unit time; Step 322: Based on the chromosome coding method, construct a load balancing degree evaluation function jointly composed of a resource utilization rate calculation unit, a variance calculation unit, and an equilibrium degree scoring unit, and calculate the variance of each resource utilization rate; Step 323: Based on the chromosome coding method, construct an energy consumption evaluation function jointly composed of a resource energy consumption model unit, a task energy consumption calculation unit, and a system energy consumption optimization unit, and calculate the energy consumption of the resource allocation scheme; Step 324: Based on the chromosome coding method, construct a delay evaluation function jointly composed of a task scheduling sequence generation unit, a critical path analysis unit, and a response time calculation unit, and calculate the average task completion time.

[0011] Step 33: Introduce an adaptive crossover and mutation operator jointly composed of a population diversity evaluation unit, a parameter adaptive adjustment unit, and an operation execution unit for the global resource allocation optimization of the macro scheduling layer, and dynamically adjust the parameters according to the population diversity; Introduce the adaptive crossover and the mutation operator for the global resource allocation optimization of the macro scheduling layer, including the following sub-steps: Step 331: Based on the evaluation result of the fitness function, construct a fitness difference analysis unit, a crossover probability calculation unit, and an individual selection unit to jointly and dynamically adjust the crossover probability according to the fitness difference of the population individuals; Step 332: Based on the result of the dynamically adjusted crossover probability, construct an evolution state monitoring unit, a convergence speed calculation unit, and a mutation probability adjustment unit to jointly and dynamically adjust the mutation probability according to the number of generations of evolution and the convergence speed; Step 333: Based on the chromosome coding method, design a dedicated crossover operation for the resource allocation problem jointly composed of a crossover point selection unit, an in-segment crossover unit, and a validity repair unit; Step 334: Based on the dedicated crossover operation, design a local search mutation operation for the scheduling optimization jointly composed of a mutation position selection unit, a neighborhood generation unit, and an optimal replacement unit; Step 34: Design an elite retention strategy composed of an elite selection unit, an elite preservation unit, and an elite reintroduction unit for the global resource allocation optimization of the macro-scheduling layer; The micro-scheduling layer in the double-layer scheduling decision-making module uses a reinforcement learning model for real-time scheduling decisions, including the following steps: Step 35: Receive the system state vector and the updated task queue state information, and define the state space of the micro-scheduling layer; Step 36: Based on the resource allocation scheme of the macro-scheduling layer, define the action space of the micro-scheduling layer as specific resource allocation decisions; Step 37: Based on the energy consumption evaluation function and the delay evaluation function, design the reward function of the micro-scheduling layer as the weighted sum of the task completion time and the energy consumption; Step 38: Based on the defined state space, action space, and reward function, use the Q-learning algorithm to optimize the policy of the reward function; The top-down guidance feedback includes the following steps: Step 41: Receive the resource allocation scheme optimized by the macro-scheduling layer, and construct a guidance information receiving unit, a constraint condition conversion unit, and a constraint condition application unit to use the resource allocation scheme as the constraint condition of the micro-scheduling layer; Step 42: The macro-scheduling layer regularly transmits the evaluated global optimization goal and resource usage boundary to the micro-scheduling layer; Propose an adaptive adjustment method for the measurement and control system based on deep learning, including the following steps: Step 1: Collect the real-time operation data of the measurement and control system, and construct the preprocessed real-time operation data into a system state vector; Step 2: Based on the system state vector, extract the features of the tasks in the measurement and control system, use a support vector machine to classify the tasks, predict the resource requirements for different types of tasks, and dynamically adjust the task priorities based on the resource requirements to generate task priority information; Step 3: Receive the system state vector and the adjusted task priority information through the pre-constructed macro-scheduling layer, and use a genetic algorithm to optimize the global resource allocation; based on the optimization result of the macro-scheduling layer through the pre-constructed micro-scheduling layer, use a reinforcement learning model for real-time scheduling decisions; and use top-down guidance feedback and bottom-up correction feedback to optimize the macro-scheduling layer and the micro-scheduling layer; Step 4: Generate specific resource allocation instructions based on the resource allocation scheme output after the optimization of the macro-scheduling layer.

[0012] Compared with the prior art, the beneficial effects of the present invention are: The present invention first collects the real-time operation data of the system and preprocesses it into a system state vector, providing a basis for subsequent decision-making. Then, it uses a support vector machine to classify different types of tasks, predict resource requirements, and dynamically adjust task priorities, enabling the system to intelligently allocate resources according to the current environment. At the scheduling level, the macro-scheduling layer uses an improved genetic algorithm to optimize global resource allocation and provides an optimal resource allocation strategy for the overall system; the micro-scheduling layer implements real-time scheduling decisions based on a lightweight reinforcement learning model, quickly responds to local changes, and solves technical problems such as inflexible resource allocation, untimely priority adjustment, and difficulty in balancing global optimization and local response in traditional measurement and control systems, significantly improving the system resource utilization rate, reducing task response latency, and enhancing the system's emergency handling ability. Brief Description of the Drawings

[0013] Figure 1 It is a flowchart of the adaptive adjustment method for a measurement and control system based on deep learning in Embodiment 1 of the present invention; Figure 2 It is a module connection diagram of the adaptive adjustment system for a measurement and control system based on deep learning in Embodiment 2 of the present invention. Detailed Embodiments

[0014] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention. Embodiment 1

[0015] As Figure 1 shown, the adaptive adjustment system for a measurement and control system based on deep learning includes: The adaptive adjustment system for a measurement and control system based on deep learning includes a resource status perception module, a priority adjustment module, a two-layer scheduling decision module, and a resource scheduling execution module; among them, each module is connected electrically; The resource status perception module collects the real-time operation data of the measurement and control system, preprocesses the real-time operation data, and constructs the preprocessed real-time operation data into a system state vector; and sends the system state vector to the priority adjustment module and the two-layer scheduling decision module as the input of resource scheduling decision-making; The priority adjustment module receives the system state vector, extracts features of tasks in the TT&C system, classifies the tasks using a support vector machine, predicts the resource requirements for different types of tasks, dynamically adjusts the task priorities based on the resource requirements, and transmits the adjusted task priority information to the two-layer scheduling decision module. The two-layer scheduling decision module consists of a macro-scheduling layer and a micro-scheduling layer. Among them, the macro-scheduling layer receives the system state vector and the adjusted task priority information, and uses a genetic algorithm to optimize the global resource allocation. The micro-scheduling layer, based on the optimization result of the macro-scheduling layer, uses a reinforcement learning model to make real-time scheduling decisions, and takes the resource allocation plan of the macro-scheduling layer as a constraint condition for the micro-scheduling layer in the form of top-down guidance feedback. The two-layer scheduling decision module sends the optimized resource allocation plan and the result of the real-time scheduling decision to the resource scheduling execution module. The resource scheduling execution module receives the resource allocation plan output by the two-layer scheduling decision module and generates specific resource allocation instructions.

[0016] The resource status perception module collects the real-time operation data of the TT&C system, including the following steps: Step 11: Collect real-time operation data; the real-time operation data includes CPU usage rate, memory occupancy rate, network bandwidth usage rate, storage space usage rate, and task queue status information. Specifically, in this embodiment, the real-time operation data is collected through a distributed resource monitoring network, which consists of a system resource monitor, a network resource monitor, and a task queue monitor.

[0017] Among them, the system resource monitor is deployed on each computing node of the TT&C system to monitor system resource metrics such as CPU usage rate, memory occupancy rate, and storage space usage rate in real time; the computing node refers to a combined unit of the hardware device and software environment that executes computing tasks in the TT&C system. Among them, the network resource monitor is deployed at the key network nodes of the TT&C system to monitor network performance metrics such as network bandwidth usage rate, network latency, and packet loss rate; the key network nodes can be the core switch or core router in the network where the TT&C system is located. Among them, the task queue monitor monitors the queue status information of various tasks in the TT&C system, including task queue length, task waiting time, task priority distribution, task type, computational complexity, data volume, time constraints, task resource demand characteristics, and task execution status, etc., which are task scheduling-related parameters. The task queue monitor parses the task descriptor in real time through a task descriptor parser, extracts the basic attributes and resource demand characteristics of the task, and forms structured task queue status information.

[0018] The system resource metrics, network performance metrics, and task scheduling related parameters collected each time form real-time operation data; Step 12: Construct the real-time operation data into a system state vector through feature engineering; Specifically, in the feature engineering process, first perform time window aggregation on the system resource metrics, network performance metrics, and task scheduling related parameters, and calculate statistical features such as average value, maximum value, minimum value, and change rate within several time windows; then perform correlation analysis on the statistical features after statistics, identify and retain the feature combinations that are most critical for characterizing the system state; finally, organize these feature combinations into a system state vector with a fixed dimension according to a predefined order and weight.

[0019] For the task queue status information, the feature engineering process includes: task clustering analysis to cluster similar task features; queue status feature extraction to calculate statistical features such as queue depth, average waiting time, and priority distribution entropy; task flow feature extraction to analyze dynamic features such as task arrival rate, completion rate, and blocking rate; resource matching degree analysis to evaluate the matching degree between the current task set and available resources. These features together constitute a task queue status sub-vector, comprehensively characterizing the task scheduling status of the TT&C system.

[0020] The constructed system state vector includes three main parts: a system resource status sub-vector, which characterizes the usage of system resources such as CPU, memory, and storage; a network resource status sub-vector, which characterizes network performance metrics such as network bandwidth and latency; and a task queue status sub-vector, which characterizes task scheduling parameters such as task queue length, waiting time, task type distribution, priority distribution, and resource demand characteristics. These three sub-vectors together constitute a complete system state vector, comprehensively reflecting the real-time operation state of the TT&C system.

[0021] Further, the priority adjustment module extracts features of tasks in the TT&C system through the following steps: Step 21: Extract the task features of each task from the task queue status information in the system state vector; the task features include but are not limited to task type, computational complexity, data volume, priority, and time constraint, etc.; Step 22: Classify the extracted task features using a support vector machine, and classify the tasks into three categories: compute-intensive, IO-intensive, and hybrid; In the specific implementation process of the present invention, the main characteristics of compute-intensive tasks are high computational complexity and relatively low IO operation frequencies. Typical examples include scientific computing, image processing, and data analysis tasks. The main characteristics of IO-intensive tasks are frequent input / output operations and relatively low computational complexity. Typical examples include data transfer, file reading and writing, and network communication tasks. Hybrid tasks have high requirements for both computing and IO operations, such as database transaction processing and real-time data stream processing tasks.

[0022] Specifically, the support vector machine is a kernel function-based supervised learning classifier used to map task feature vectors to predefined task categories. The support vector machine consists of a model training unit, a feature mapping unit, and a classification decision unit.

[0023] Among them, the model training unit uses historical task data to train the support vector machine model. The historical task data contains a number of labeled task samples, and each task sample contains a vector of task features and a corresponding task category label (determined compute-intensive, IO-intensive, and hybrid). By minimizing the structural risk, the support vector machine learns the optimal classification hyperplane in the feature space to achieve accurate classification of the task categories of new tasks.

[0024] The feature mapping unit uses the radial basis function as the kernel function to map the task feature vectors to a high-dimensional feature space, enhancing the separability of the features. In the high-dimensional feature space, different types of tasks exhibit more obvious clustering characteristics, facilitating accurate classification by the support vector machine.

[0025] The classification decision unit receives the mapped high-dimensional feature vectors, calculates the distances from the high-dimensional feature vectors to each classification hyperplane, and determines the task category based on the distance relationship. The decision-making process adopts a one-versus-all strategy to construct several binary classifiers to respectively determine whether the task belongs to compute-intensive, IO-intensive, or hybrid, and finally comprehensively determines the final category of the task based on the results of each classifier.

[0026] Step 23: For the different types of classified tasks, use a linear regression model to predict their resource requirements; Specifically, the linear regression model is a statistical learning model for predicting resource requirements based on task feature vectors. The linear regression model consists of a feature selection unit, a model construction unit, and a prediction execution unit.

[0027] Among them, the feature selection unit selects the most relevant feature subset according to the task type. For example, for computationally intensive tasks, features such as computational complexity, algorithm type, and data scale are selected; for I / O intensive tasks, features such as I / O operation frequency, data transfer volume, and storage access mode are selected; for hybrid tasks, features related to both computation and I / O are considered comprehensively. The feature selection process adopts a method based on correlation analysis, retaining features highly correlated with resource requirements and removing redundant and irrelevant features. The model construction unit constructs independent linear regression models for each resource type, including a CPU demand prediction model, a memory demand prediction model, a network bandwidth demand prediction model, and a storage space demand prediction model. Each model estimates model parameters using the weighted least squares method, and the weights are dynamically adjusted according to historical prediction accuracy to improve prediction accuracy.

[0028] The prediction execution unit receives the task feature vector and task type information, and selects the corresponding linear regression model to predict the resource requirements. The prediction result includes the estimated resource requirements of the task during execution, which is output in the form of a resource demand vector and serves as an important basis for subsequent dynamic adjustment of task priorities.

[0029] The priority adjustment module establishes a dynamic task priority adjustment mechanism as follows: Extract the waiting time in the task queue status information of the system state vector, and adaptively adjust the task priority according to the waiting time and the system state vector; Specifically, the adaptive adjustment of task priority includes: The priority adjustment module consists of a resource competition analysis unit, a priority calculation unit, and a priority update unit.

[0030] The resource competition analysis unit calculates the current system resource competition degree based on the system resource status sub-vector and network resource status sub-vector in the system state vector. The system resource competition degree is described by a resource competition index, which is calculated by weighted combination of factors such as CPU usage rate, memory occupancy rate, network bandwidth usage rate, and storage space usage rate to quantify the tightness of system resources. Obviously, the higher the resource competition index, the more intense the system resource competition, and more precise task priority adjustment is required.

[0031] Among them, the priority calculation unit calculates the dynamic priority adjustment factor for each task based on the waiting time of the task, the resource competition index, and the resource requirements of the task. The dynamic priority adjustment factor is calculated using an adaptive weighting method. For example, the weights of the waiting time, resource requirements, and original priority are dynamically adjusted according to the system load condition to ensure the rationality of priority adjustment under different system states.

[0032] In the specific implementation process of the present invention, the method for calculating the dynamic priority adjustment factor implemented by the priority calculation unit includes the following sub-steps: Step 221: Construct a waiting time compensation function. As the waiting time of the task increases, the compensation value increases non-linearly to prevent the task from starving for a long time; Step 222: Construct a resource requirement penalty function. Appropriately reduce the priority of tasks with large resource requirements, but set an upper limit to avoid high-resource requirement tasks from never being executed; Step 223: Construct a system load adaptive function. According to the current system load status, dynamically adjust the weights of waiting time compensation and resource requirement penalty; Step 224: Weightedly fuse the task priority, waiting time compensation value, resource requirement penalty value, and system load adaptive factor in the queue status information to calculate the final dynamic priority adjustment factor; Among them, the priority update unit receives the dynamic priority adjustment factor and updates the priority information of the task.

[0033] Furthermore, the macro scheduling layer in the double-layer scheduling decision module adopts a genetic algorithm, which includes the following steps: Step 31: Receive the system state vector and the adjusted task priority information, and design a chromosome encoding method for the global resource allocation optimization of the macro scheduling layer to represent the resource allocation scheme; Specifically, the design of the chromosome encoding method includes the following sub-steps: Step 311: Design a chromosome structure based on integer encoding to represent the mapping relationship between the predicted task resource requirements and the system resources; Specifically, the chromosome structure based on integer encoding consists of a chromosome encoding unit and a resource mapping unit.

[0034] Among them, the chromosome encoding unit uses a multi-dimensional integer array to represent the resource allocation scheme. Each chromosome consists of several gene positions, and each gene position corresponds to an allocation relationship between a task and a resource. In the chromosome structure, the arrangement order of the gene positions is consistent with the order of the tasks in the task queue to ensure the consistency and interpretability of the encoding structure. The value of each gene position represents the resource identifier allocated to the task, and is directly mapped to a specific computing node, memory block, network channel, or storage unit through integer encoding.

[0035] The resource mapping unit converts the chromosome encoding into actual resource allocation instructions. The resource mapping unit maintains a mapping table between resource identifiers and physical resources, ensuring that the integer encoding in the chromosome can accurately correspond to the actual resources of the measurement and control system. The resource mapping process takes into account the hierarchical structure and dependencies of resources. For example, computing nodes are allocated first, then memory and storage resources are allocated within the nodes, and finally network channels are allocated to form a complete resource allocation plan.

[0036] Step 312: Based on the chromosome structure, design a chromosome segmentation strategy to encode the allocation plans of different types of resources respectively; The different types of resources may include but are not limited to computing resources, memory resources, network resources, and storage resources; Specifically, the chromosome segmentation strategy divides the chromosome into several segments, and each segment corresponds to the allocation plan of a resource type. The chromosome segmentation strategy consists of a resource type division unit, an intra-segment encoding unit, and an inter-segment association unit.

[0037] Among them, the resource type division unit divides the chromosome into a computing resource segment, a memory resource segment, a network resource segment, and a storage resource segment according to the resource types of the measurement and control system. The length of each resource segment is related to the number of resources of the corresponding type and the number of tasks in the system, ensuring that the allocation plan of this type of resource can be completely represented.

[0038] The intra-segment encoding unit designs a dedicated encoding method for each resource type. For the computing resource segment, a combined encoding of the processor identifier and the core identifier is used; for the memory resource segment, a combined encoding of the memory block identifier and the allocation size is used; for the network resource segment, a combined encoding of the channel identifier and the bandwidth allocation ratio is used; for the storage resource segment, a combined encoding of the storage unit identifier and the access permission is used.

[0039] The inter-segment association unit maintains the association relationship between different resource segments to ensure the consistency of resource allocation. For example, when a certain task is allocated to a specific computing node, the allocation of its memory and storage resources should give priority to resources that are physically close to the computing node to reduce data transmission overhead. The inter-segment association unit quantifies the association degree between different resources by establishing a resource affinity matrix to guide the coordinated allocation of multi-type resources.

[0040] Step 313: Based on the segmentation strategy, design a constraint condition encoding method consisting of a hard constraint encoding unit, a soft constraint encoding unit, and a constraint priority unit to constrain the resource allocation to meet the system limitations; Specifically, the hard constraint encoding unit encodes the system constraint conditions that must be satisfied, including resource capacity constraints, task dependency constraints, and resource mutual exclusion constraints. The resource capacity constraint ensures that the total amount of resources allocated to all tasks does not exceed the available resources of the system; the task dependency constraint ensures that tasks with dependencies are allocated resources in the correct order; the resource mutual exclusion constraint ensures that exclusive resources are not occupied by multiple tasks simultaneously. The hard constraint conditions are encoded through a feasibility check function to verify each chromosome, and chromosomes that do not meet the hard constraints will be marked as invalid solutions.

[0041] The soft constraint encoding unit encodes the constraint conditions that affect system performance but can be appropriately relaxed, such as load balancing constraints, energy consumption control constraints, and response time constraints. The soft constraint conditions are encoded through a penalty function to quantify the degree of violation of the soft constraints and incorporate the penalty value into the calculation of the fitness function, guiding the genetic algorithm to evolve in the direction of satisfying the soft constraints.

[0042] The constraint priority unit sets the priorities of different constraint conditions and makes trade-offs in case of multi-constraint conflicts. The constraint priority is encoded through weight coefficients, and violations of high-priority constraints will result in larger penalty values. The constraint priority can be dynamically adjusted according to the current state of the measurement and control system. For example, when the system load is low, the priority of the energy consumption control constraint can be increased; when the task urgency is high, the priority of the response time constraint can be increased.

[0043] Step 314: Based on the task priority information, design a priority encoding method consisting of a priority mapping unit, an execution order encoding unit, and an importance encoding unit to reflect the execution order and importance of tasks; Specifically, the priority mapping unit maps the task priority information output by the priority adjustment module into the chromosome encoding. The mapping process uses a priority conversion function to convert the task priority value into adjustment factors for the selection probability, crossover probability, and mutation probability in the genetic algorithm, enabling high-priority tasks to receive more attention during the resource allocation process.

[0044] The execution order encoding unit encodes the execution order information of tasks in the chromosome. The execution order is encoded through the relative positions of gene loci and special marker genes in the chromosome to ensure that tasks with dependencies are executed in the correct order. For a task with a predecessor task, its gene locus contains a reference to the completion flag of the predecessor task, and the task can only be scheduled for execution after the predecessor task is completed.

[0045] The importance coding unit encodes the importance information of tasks in the chromosome. The importance is encoded by adding a weight field in the gene locus, which affects the priority order and proportion of resource allocation. The gene locus of a high-importance task obtains a higher weight value, has priority in resource competition, and may obtain more resource allocation. The importance coding also affects the calculation of the fitness function, making the completion of important tasks contribute more to the fitness and guiding the genetic algorithm to preferentially meet the resource requirements of important tasks.

[0046] Step 32: Design a fitness function composed of an objective function construction unit, a weight allocation unit, and a normalization processing unit for the global resource allocation optimization of the macro-scheduling layer, comprehensively considering system throughput and load balance. Specifically, the objective function construction unit constructs a mathematical model for evaluating the quality of the resource allocation scheme. Based on multiple evaluation indicators such as system throughput, load balance, energy consumption efficiency, and task response time, the unit constructs a multi-objective optimization function. The system throughput objective function calculates the number of tasks completed per unit time; the load balance objective function calculates the variance of the resource utilization rates of each computing node, memory unit, network channel, and storage unit, and the smaller the variance, the more balanced the load; the energy consumption efficiency objective function calculates the energy consumption required to complete all tasks; the task response time objective function calculates the average time from task submission to completion.

[0047] The weight allocation unit assigns weight coefficients to each objective in the multi-objective optimization function. The weight allocation is based on the current state and operation requirements of the measurement and control system, and dynamically adjusts the importance of each objective. For example, in a high-load situation, the weight of system throughput can be increased; in an energy-constrained situation, the weight of energy consumption efficiency can be increased; in a situation with high real-time requirements, the weight of task response time can be increased. The weight allocation adopts an adaptive mechanism and automatically adjusts according to the indicators in the system state vector to ensure that the fitness function can reflect the current system's optimization requirements.

[0048] The normalization processing unit normalizes the calculation results of each objective function so that indicators with different dimensions can be weighted and summed. The normalization adopts the maximum-minimum normalization method, mapping the values of each objective function to the interval [0,1]. The normalized objective function values are multiplied by the corresponding weight coefficients and then summed to obtain the final fitness value. The higher the fitness value, the better the resource allocation scheme, and the greater the chance of being retained and reproduced in the selection operation of the genetic algorithm.

[0049] Designing a corresponding fitness function for the genetic algorithm includes the following sub-steps: Step 321: Based on the chromosome encoding method, construct a system throughput evaluation function composed of a task execution time estimation unit, a parallelism analysis unit, and a throughput calculation unit, and calculate the number of tasks completed per unit time. Specifically, the task execution time estimation unit estimates the execution time of each task based on the resource requirement characteristics of the task and the allocated resources. The estimation process considers the computational complexity, data volume, resource utilization efficiency, and resource performance parameters of the task, and calculates the expected execution time of the task under the given resource configuration through a performance model.

[0050] The parallelism analysis unit analyzes the parallel execution of tasks in the resource allocation plan. The analysis process considers the dependency relationship between tasks, the independence of resource allocation, and the parallel processing ability of the system, and determines the maximum number of tasks that the system can execute simultaneously under the given resource allocation plan.

[0051] The throughput calculation unit calculates the number of tasks that the system can complete per unit time based on the task execution time and the system parallelism. The calculation process considers the arrival pattern of tasks, the execution time distribution, and the system scheduling strategy, and obtains the expected throughput of the system through a queuing theory model or a discrete event simulation method. The throughput calculation result is an important part of the fitness function, guiding the genetic algorithm to optimize the processing ability of the system.

[0052] Step 322: Based on the chromosome encoding method, construct a load balancing degree evaluation function composed of a resource utilization rate calculation unit, a variance calculation unit, and an equilibrium degree scoring unit, and calculate the variance of each resource utilization rate. Specifically, the resource utilization rate calculation unit calculates the expected utilization rate of each resource unit in the system based on the resource allocation plan encoded by the chromosome. The calculation process considers the number of tasks allocated to each resource, the resource requirement intensity of the tasks, and the processing ability of the resources, and obtains the utilization rate prediction value of each resource unit.

[0053] The variance calculation unit calculates the statistical variance of the utilization rates of the same type of resources. The calculation process groups the utilization rates of computing resources, memory resources, network resources, and storage resources respectively, and calculates the variance of the resource utilization rates within each group. The smaller the variance value, the more balanced the load distribution of this type of resources.

[0054] The equilibrium degree scoring unit comprehensively evaluates the load balancing degree of various types of resources as an equilibrium degree score. The scoring process considers the importance and sensitivity of different types of resources, and performs a weighted average on the variance values of various types of resources to obtain the overall load balancing degree score of the system. The equilibrium degree score is an integral part of the fitness function, guiding the genetic algorithm to optimize the balanced use of resources and avoid resource hotspots and resource starvation phenomena.

[0055] Step 323: Based on the chromosome coding method, construct an energy consumption evaluation function composed of a resource energy consumption model unit, a task energy consumption calculation unit, and a system energy consumption optimization unit to calculate the energy consumption of the resource allocation plan; Specifically, the resource energy consumption model unit establishes an energy consumption model for various resources of the measurement and control system. The model considers the relationship between the static power consumption, dynamic power consumption, and resource utilization rate of the resources, and describes the energy consumption performance of the resources under different loads through an energy consumption characteristic curve. The energy consumption model covers the energy consumption characteristics of computing resources, memory resources, network resources, and storage resources, providing basic data for energy consumption evaluation.

[0056] The task energy consumption calculation unit calculates the energy consumption during the execution of each task based on the resource energy consumption model and the resource usage of the task. The calculation process considers the execution time of the task, the resource usage intensity, and the energy consumption characteristics of the resources, and obtains the total energy consumption prediction value of the task through the energy consumption integration method.

[0057] The system energy consumption optimization unit comprehensively considers the energy consumption of all tasks and the energy consumption management strategy of the system to calculate the energy efficiency of the entire resource allocation plan. The optimization process considers the dynamic power management of the resources, the energy consumption-aware scheduling of the tasks, and the heat dissipation constraints of the system, and evaluates the energy utilization efficiency of the resource allocation plan through the energy efficiency ratio index. The energy consumption evaluation result is used as a component of the fitness function to guide the genetic algorithm to optimize energy use while ensuring performance.

[0058] Step 324: Based on the chromosome coding method, construct a delay evaluation function composed of a task scheduling sequence generation unit, a critical path analysis unit, and a response time calculation unit to calculate the average task completion time; Specifically, the task scheduling sequence generation unit generates a scheduling execution sequence of the tasks based on the resource allocation plan encoded by the chromosome and the task dependencies. The generation process considers the priorities of the tasks, the dependency constraints, and the available time of the resources, and constructs a Gantt chart of task execution through a list scheduling algorithm or a critical path method to clarify the start time and completion time of each task.

[0059] The critical path analysis unit identifies the critical path in the task execution sequence. The analysis process considers the dependencies and execution times between tasks, and traverses the longest execution path from the first task to the last task. The execution time of the tasks on the critical path determines the shortest completion time of the entire task set and is the key object of optimization scheduling.

[0060] The response time calculation unit calculates the average response time of a task from submission to completion. The calculation process takes into account the waiting time, execution time, and resource competition of the task, and obtains the average response time index of the system through a weighted average method. The response time evaluation result is used as a component of the fitness function to guide the genetic algorithm to optimize the timely processing of tasks and improve the real-time performance of the system.

[0061] Step 33: Introduce an adaptive crossover and mutation operator composed of a population diversity evaluation unit, a parameter adaptive adjustment unit, and an operation execution unit for the global resource allocation optimization of the macro-scheduling layer, and dynamically adjust the parameters according to the population diversity. Specifically, the population diversity evaluation unit evaluates the diversity level of the current population. The evaluation process uses a method that combines the gene diversity index and the phenotype diversity index to calculate the average difference degree between individuals in the population. The gene diversity measures the difference in chromosome encoding through the Hamming distance or the edit distance; the phenotype diversity measures the difference in individual performance through the distribution range and variance of fitness values.

[0062] The parameter adaptive adjustment unit dynamically adjusts the crossover probability and the mutation probability according to the population diversity evaluation result. When the population diversity is low, increase the mutation probability and decrease the crossover probability to promote the population to explore new solution spaces; when the population diversity is high, increase the crossover probability and decrease the mutation probability to accelerate the population's convergence to the optimal solution. The parameter adjustment uses fuzzy control rules or adaptive functions to achieve a smooth mapping between the parameters and the diversity.

[0063] The operation execution unit performs the crossover and mutation operations. The crossover operation uses a crossover method dedicated to the resource allocation problem, such as partially mapped crossover or order crossover, to ensure that the generated offspring chromosomes still represent valid resource allocation schemes; the mutation operation uses a local search strategy, such as neighborhood search or simulated annealing, to explore better resource allocation schemes on the premise of maintaining the validity of the chromosome structure.

[0064] Introducing the adaptive crossover and the mutation operator for the global resource allocation optimization of the macro-scheduling layer includes the following sub-steps: Step 331: Based on the evaluation result of the fitness function, construct a fitness difference analysis unit, a crossover probability calculation unit, and an individual selection unit to jointly and dynamically adjust the crossover probability according to the fitness difference of population individuals. Among them, the fitness difference analysis unit calculates the fitness statistical characteristics of individuals in the population, including the maximum fitness, the minimum fitness, the average fitness, and the fitness standard deviation. The analysis result reflects the evolutionary state and convergence trend of the population, and provides a basis for the dynamic adjustment of the crossover probability.

[0065] The crossover probability calculation unit calculates the crossover probability for each pair of individuals to be crossed based on the fitness difference analysis result. The calculation process takes into account the gap between the fitness value of an individual and the average fitness of the population. For individuals with fitness higher than the average, a lower crossover probability is assigned to protect their excellent genes; for individuals with fitness lower than the average, a higher crossover probability is assigned to promote gene recombination. The calculation of the crossover probability uses a non-linear mapping function, and different adjustment strategies are adopted at the initial and late stages of population convergence.

[0066] The individual selection unit selects pairs of individuals participating in the crossover operation according to the calculated crossover probability. The selection process uses the roulette wheel method or the tournament method, and combines the crossover probability for random selection to ensure that the selection process has a certain degree of randomness and tends to select individuals with higher fitness. The selected pairs of individuals will undergo a crossover operation to generate new offspring individuals, enriching the genetic diversity of the population.

[0067] Step 332: Based on the result of dynamically adjusting the crossover probability, construct an evolution state monitoring unit, a convergence speed calculation unit, and a mutation probability adjustment unit to jointly and dynamically adjust the mutation probability according to the generation number of evolution and the convergence speed; Among them, the evolution state monitoring unit tracks and records the generation number of evolution and the change of population state of the genetic algorithm. The monitoring process records the optimal fitness, average fitness, and fitness variance of each generation, constructs the historical trajectory of population evolution, and provides time-series data for the dynamic adjustment of the mutation probability.

[0068] The convergence speed calculation unit calculates the convergence speed of the population based on the evolution state monitoring data. The calculation process takes into account the change rate of the optimal fitness and the change trend of population diversity in several consecutive generations, and obtains the convergence speed index of the current population through the sliding window method or the exponential smoothing method.

[0069] The mutation probability adjustment unit dynamically adjusts the mutation probability according to the generation number of evolution and the convergence speed. The adjustment strategy follows the principle of "low mutation in the early stage, high mutation in the middle stage, and fine mutation in the late stage". A lower mutation probability is maintained in the initial stage of the algorithm to avoid destroying excellent genes; the mutation probability is increased in the middle stage to enhance the exploration ability of the population; the mutation probability is decreased in the late stage for fine search to accelerate convergence to the optimal solution. The specific adjustment of the mutation probability uses a piecewise function or an adaptive function to achieve the dynamic matching of the mutation probability and the evolution state.

[0070] Step 333: Based on the chromosome encoding method, design a dedicated crossover operation for the resource allocation problem jointly composed of a crossover point selection unit, an intra-segment crossover unit, and a validity repair unit; Specifically, the crossover point selection unit selects appropriate crossover points in the parental chromosomes. The selection process takes into account the segmented structure of the chromosomes and the integrity of resource allocation, and preferentially selects crossover points at the boundaries of resource segments to ensure that the crossover operation does not disrupt the internal consistency of resource allocation. For multi-point crossover, the selection of crossover points also considers the dependencies between resources to avoid splitting strongly related resource allocation schemes into different crossover segments.

[0071] The intra-segment crossover unit performs dedicated crossover operations for different resource segments. For the computing resource segment, the partially mapped crossover method is adopted to maintain the one-to-one mapping relationship between tasks and processors; for the memory resource segment, the uniform crossover method is adopted to independently exchange the memory allocations of each task; for the network resource segment, the order crossover method is adopted to maintain the relative order of network channel allocations; for the storage resource segment, the single-point crossover method is adopted to exchange large chunks of storage allocation schemes.

[0072] The validity repair unit checks whether the offspring chromosomes generated after crossover satisfy the constraint conditions of resource allocation. The checking process verifies hard constraint conditions such as resource capacity constraints, task dependency constraints, and resource mutual exclusion constraints. For chromosomes that violate the constraints, they are corrected through local adjustment or greedy repair methods to ensure that the offspring chromosomes represent valid resource allocation schemes.

[0073] Step 334: Based on the dedicated crossover operation, design a local search mutation operation for scheduling optimization jointly composed of a mutation position selection unit, a neighborhood generation unit, and an optimal replacement unit; Specifically, the mutation position selection unit selects the positions in the chromosomes that need to mutate. The selection process takes into account the imbalance of resource utilization and the criticality of tasks, and preferentially selects resource units with high loads or tasks on the critical path for mutation to improve the pertinence and effectiveness of the mutation operation.

[0074] The neighborhood generation unit generates multiple candidate mutation schemes based on the selected mutation positions. The generation process adopts problem-specific neighborhood definitions, such as resource reallocation neighborhoods, task exchange neighborhoods, or priority adjustment neighborhoods, to construct the local search space of the current solution. The neighborhood generation takes into account the constraint conditions of resource allocation and only generates valid mutation schemes that satisfy the constraints.

[0075] The optimal replacement unit evaluates the fitness of all candidate mutation schemes and selects the optimal scheme to replace the original solution. The evaluation process adopts an incremental calculation method, only calculating the impact of the mutated part on the fitness to improve the evaluation efficiency. The replacement strategy can be greedy selection (always selecting the best scheme) or probabilistic acceptance (accepting sub-optimal schemes with a certain probability) to balance the convergence speed of the algorithm and the ability to jump out of local optima.

[0076] Step 34: Design an elite retention strategy composed of an elite selection unit, an elite preservation unit, and an elite reintroduction unit for the global resource allocation optimization of the macro-scheduling layer to ensure that the optimal solution evaluated is not destroyed; Specifically, in each generation of the evolutionary process, the elite selection unit selects a number of individuals with the highest fitness as elite individuals from the current population according to the evaluation results of the fitness function. The number of elite individuals is dynamically determined according to the population size and usually remains at a certain proportion of the total population. The elite selection process adopts a sorting-based selection method. First, all individuals are sorted in descending order of fitness value, and then the individuals ranked at the front are selected as elite individuals.

[0077] The elite preservation unit directly copies the selected elite individuals into the next-generation population without going through the crossover and mutation operations to ensure that these high-quality solutions are not destroyed during the evolutionary process. The elite preservation unit maintains an elite pool for storing the optimal solutions found in historical iterations and regularly updates the individuals in the elite pool to ensure that the elite pool always stores high-quality solutions representing different optimization directions.

[0078] The elite reintroduction unit monitors the population diversity during the evolutionary process. When it detects that the population diversity drops below a preset threshold, it selects elite individuals with a large difference from the current population from the elite pool and reintroduces them into the current population to increase the population diversity and avoid the algorithm falling into a local optimal solution. The reintroduction process adopts a diversity evaluation method based on distance measurement to ensure that the reintroduced elite individuals can maximize the population diversity.

[0079] The micro-scheduling layer in the double-layer scheduling decision module uses a reinforcement learning model to make real-time scheduling decisions, including the following steps: Step 35: Receive the system state vector and the updated task queue state information and define the state space of the micro-scheduling layer; Specifically, the state space of the micro-scheduling layer consists of a state representation unit, a state compression unit, and a state update unit.

[0080] Among them, the state representation unit combines the system state vector and the updated task queue state information into a complete state description. The state description includes three parts: the current available resource state, the state of tasks to be scheduled, and the system environment state. The available resource state describes the current available quantity and utilization rate of various resources; the state of tasks to be scheduled describes the types, priorities, resource requirements, and waiting times of tasks in the task queue; the system environment state describes environmental factors such as the system load trend, resource competition degree, and task arrival pattern.

[0081] The state compression unit performs dimensionality reduction on the complete state description, reducing the dimension of the state space and improving the convergence speed of the reinforcement learning algorithm. The state compression adopts the principal component analysis method, retaining the state features that are most critical for decision-making and eliminating redundant and noise features. The compressed state representation retains the main features of the original state information while reducing the dimension of the state space.

[0082] The state update unit updates the system state at each decision time point. The state update process includes: obtaining the latest system state vector from the resource state perception module; obtaining the updated task queue state information from the priority adjustment module; fusing the newly obtained information with the historical state information to form a temporal state representation; and compressing the fused state through the state compression unit to generate the state representation at the current decision moment.

[0083] Step 36: Based on the resource allocation scheme of the macro scheduling layer, define the action space of the micro scheduling layer as specific resource allocation decisions; Specifically, the action space of the micro scheduling layer consists of an action definition unit, an action constraint unit, and an action execution unit.

[0084] Among them, the action definition unit represents the resource allocation decision as a multi-dimensional vector, and each dimension corresponds to the allocation decision of a resource type. For each task to be scheduled, the action vector describes the specific quantity of various resources allocated to the task. The design of the action space takes into account the divisibility of resources and the parallelism of tasks, supporting fine-grained resource allocation and concurrent execution of tasks.

[0085] The action constraint unit sets constraint conditions for the action space of the micro scheduling layer based on the resource allocation scheme of the macro scheduling layer. The constraint conditions include total resource constraint, task priority constraint, and system policy constraint. The total resource constraint ensures that the total amount of resources allocated by the micro scheduling layer does not exceed the total amount of resources allocated by the macro scheduling layer; the task priority constraint ensures that high-priority tasks are preferentially allocated resources; and the system policy constraint sets specific resource allocation rules according to the system management policy, such as reserved resources, resource isolation, and resource reservation.

[0086] The action execution unit converts the selected action into specific resource allocation instructions and implements resource allocation through the resource scheduling execution module. The action execution process includes: parsing the selected action vector to determine the resource allocation scheme for each task; checking whether the resource allocation scheme meets all constraint conditions; generating a sequence of resource allocation instructions; and passing the instruction sequence to the resource scheduling execution module for execution.

[0087] Step 37: Based on the energy consumption evaluation function and the delay evaluation function, design the reward function of the micro scheduling layer as the weighted sum of the task completion time and the energy consumption; Specifically, the reward function of the micro-scheduling layer consists of a reward calculation unit, a reward normalization unit, and a reward adjustment unit.

[0088] Among them, the reward calculation unit comprehensively considers two key indicators, the task completion time and the energy consumption, and constructs a weighted reward function. The task completion time is calculated through a delay evaluation function, which reflects the impact of the resource allocation decision on the system response performance; the energy consumption is calculated through an energy consumption evaluation function, which reflects the impact of the resource allocation decision on the system energy efficiency. The reward calculation uses negative values, and the shorter the completion time and the lower the energy consumption, the higher the reward value (the smaller the absolute value of the negative value).

[0089] The reward normalization unit normalizes the calculated original reward value, making the reward values generated by tasks of different scales and types comparable. The normalization process considers the complexity and resource requirements of the tasks, adjusts the scale of the reward values of different tasks, and ensures that the reward signal can accurately reflect the decision-making quality and is not affected by the task scale and type.

[0090] The reward adjustment unit dynamically adjusts the weights of the completion time and the energy consumption according to the system state and the optimization goal. When the system load is light, the weight of the completion time is increased to give priority to improving the system response performance; when the system load is heavy, the weight of the energy consumption is increased to give priority to improving the system energy efficiency. The weight adjustment process uses an adaptive method to dynamically calculate the optimal weight combination according to factors such as the system load status, the task queue length, and the resource utilization rate.

[0091] Step 38: Based on the defined state space, action space, and reward function, use the Q-learning algorithm to optimize the policy of the reward function; Specifically, the Q-learning algorithm consists of a Q-value update unit, an exploration strategy unit, and a policy extraction unit.

[0092] Among them, the Q-value update unit maintains a Q-table or a Q-network to store the value estimates of state-action pairs. After each decision, according to the observed reward and the next state, use the temporal difference learning method to update the Q-value. The Q-value update adopts an experience replay mechanism, stores the historical decision-making experiences in the replay buffer, and randomly samples several experiences from the buffer for batch update each time, improving the learning efficiency and stability.

[0093] The exploration strategy unit uses an ε-greedy strategy to balance exploration and exploitation. During the decision-making process, randomly select actions with a probability of ε for exploration, and select the action with the highest current Q-value with a probability of 1-ε for exploitation. The ε value gradually decays with the learning process, mainly for exploration in the initial stage and mainly for exploitation in the later stage, ensuring that the algorithm can fully explore the state-action space and gradually converge to the optimal policy.

[0094] The policy extraction unit extracts the decision-making policy based on the learned Q values. At each decision-making moment, it queries the Q-table or Q-network according to the current state and selects the action with the highest Q value as the decision result. The policy extraction process takes into account the action constraint conditions to ensure that the extracted policy meets all constraint requirements. For complex state spaces, the policy extraction unit adopts function approximation methods, such as neural network models, to approximately represent the Q function and achieve effective processing of continuous state spaces.

[0095] Furthermore, the top-down guidance feedback includes the following steps: Step 41: Receive the resource allocation plan optimized by the macro-scheduling layer. The guidance information receiving unit, the constraint condition conversion unit, and the constraint condition application unit together use the resource allocation plan as the constraint condition for the micro-scheduling layer. Among them, the guidance information receiving unit receives the resource allocation plan output by the macro-scheduling layer. The resource allocation plan includes global optimization information such as the total system resource allocation, task priority ranking, and resource usage boundaries. The guidance information receiving unit establishes a connection with the macro-scheduling layer through a high-speed data channel to ensure that the resource allocation plan can be transmitted to the micro-scheduling layer in real time, providing global guidance for real-time scheduling decisions.

[0096] Among them, the constraint condition conversion unit converts the resource allocation plan into a constraint condition format that can be directly used by the micro-scheduling layer. The conversion process includes resource boundary extraction, priority rule conversion, and time window division. Resource boundary extraction extracts the upper limits of various resources from the resource allocation plan; priority rule conversion maps the macro-task priorities to the priority constraints of micro-scheduling; time window division divides the global resource allocation plan into several scheduling periods according to the time dimension, providing time-segmented resource constraints for the micro-scheduling layer.

[0097] Among them, the constraint condition application unit applies the converted constraint conditions to the decision-making process of the micro-scheduling layer. The application methods include action space constraint, reward function adjustment, and state space reconstruction. Action space constraint restricts the micro-scheduling layer to make decisions only within the action subspace that meets the constraint conditions; reward function adjustment adjusts the reward signal of the micro-scheduling layer according to the optimization goal of the macro-scheduling layer; state space reconstruction adds features related to the constraint conditions to the state representation of the micro-scheduling layer, enabling the micro-scheduling layer to perceive global constraints.

[0098] Step 42: The macro-scheduling layer regularly transmits the evaluated global optimization goal and resource usage boundaries to the micro-scheduling layer. Specifically, the macro-scheduling layer consists of a target transmission unit, a boundary update unit, and a transmission cycle control unit.

[0099] Among them, the target transfer unit is responsible for transferring the global optimization objectives of the macro-scheduling layer to the micro-scheduling layer. The global optimization objectives include system throughput objective, load balancing degree objective, energy consumption efficiency objective, and task response time objective. The target transfer unit not only transfers the target values, but also transfers the weight coefficients of each target to guide the micro-scheduling layer to make trade-offs among multiple objectives. The target transfer adopts a structured target description format, including information such as target type, target value, weight coefficient, and timeliness mark.

[0100] Among them, the boundary update unit is responsible for updating and transferring the resource usage boundaries. The resource usage boundaries define the usage limits of the micro-scheduling layer on various resources, including computing resource boundaries, memory resource boundaries, network resource boundaries, and storage resource boundaries. The boundary update process takes into account the dynamic changes in the system resource status, and dynamically adjusts the resource usage boundaries according to the resource utilization in the system state vector, ensuring that the boundary settings can not only meet the global optimization objectives, but also adapt to the real-time fluctuations of the system resources.

[0101] Among them, the transfer period control unit controls the timing and frequency of target and boundary transfers. The transfer period control adopts an adaptive mechanism, and dynamically adjusts the transfer period according to the system load status and task change frequency. When the system load is stable, a longer transfer period is adopted to reduce communication overhead; when the system load fluctuates greatly, the transfer period is shortened to ensure that the micro-scheduling layer can obtain the latest global guidance information in a timely manner. The transfer period control also includes a trigger mechanism. When a major change occurs in the system state, a target and boundary transfer is immediately triggered to ensure that the micro-scheduling layer can quickly respond to system changes.

[0102] The resource scheduling execution module generates specific resource allocation instructions in the following way: According to the total system resource allocation, task priority ranking, and resource usage boundaries in the resource allocation plan, corresponding total system resource allocation instructions, task priority ranking instructions, and resource usage convenience instructions are generated for the measurement and control system, constituting the resource allocation instructions. Embodiment 2

[0103] As Figure 2 shown, the adaptive adjustment method for the measurement and control system based on deep learning includes the following steps: Step 1: Collect the real-time operation data of the measurement and control system, and construct the preprocessed real-time operation data into a system state vector; Step 2: Based on the system state vector, extract the features of the tasks in the measurement and control system, classify the tasks using a support vector machine, predict the resource requirements for different types of tasks, and dynamically adjust the task priorities based on the resource requirements to generate task priority information; Step 3: Receive the system state vector and the adjusted task priority information through a pre-constructed macro-scheduling layer, and use a genetic algorithm to optimize the global resource allocation; based on the optimization results of the macro-scheduling layer through a pre-constructed micro-scheduling layer, use a reinforcement learning model to make real-time scheduling decisions; and optimize the macro-scheduling layer and the micro-scheduling layer using top-down guidance feedback and bottom-up correction feedback; Step 4: Generate specific resource allocation instructions based on the resource allocation plan output after the optimization of the macro-scheduling layer.

[0104] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An adaptive adjustment system for a measurement and control system based on deep learning, characterized in that, It includes a resource status perception module, a priority adjustment module, a two-layer scheduling decision-making module, and a resource scheduling execution module; among them, each module is connected electrically; The resource status perception module collects the real-time operation data of the measurement and control system, preprocesses the real-time operation data, and constructs the preprocessed real-time operation data into a system state vector; and sends the system state vector to the priority adjustment module and the two-layer scheduling decision-making module; The priority adjustment module receives the system state vector, extracts the characteristics of the tasks in the measurement and control system, classifies the tasks using a support vector machine, predicts the resource requirements for different types of tasks, and dynamically adjusts the task priorities based on the resource requirements; and transmits the adjusted task priority information to the two-layer scheduling decision-making module; The two-layer scheduling decision-making module consists of a macro scheduling layer and a micro scheduling layer; among them, the macro scheduling layer receives the system state vector and the adjusted task priority information, and uses a genetic algorithm to optimize the global resource allocation; the micro scheduling layer based on the optimization result of the macro scheduling layer, uses a reinforcement learning model to make real-time scheduling decisions; and takes the resource allocation plan of the macro scheduling layer as the constraint condition for the micro scheduling layer in the form of top-down guidance feedback; the two-layer scheduling decision-making module sends the optimized resource allocation plan and the result of the real-time scheduling decision to the resource scheduling execution module; The resource scheduling execution module receives the resource allocation plan output by the two-layer scheduling decision-making module and generates specific resource allocation instructions.

2. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 1, wherein The steps for the resource status perception module to collect the real-time operation data of the measurement and control system are as follows: Step 11: Collect real-time operation data; the real-time operation data includes CPU usage rate, memory occupancy rate, network bandwidth usage rate, storage space usage rate, and task queue status information; Step 12: Construct the real-time operation data into a system state vector through feature engineering.

3. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 2, characterized in that The steps for the priority adjustment module to extract the characteristics of the tasks in the measurement and control system are as follows: Step 21: Extract the task characteristics of each task from the task queue status information in the system state vector; Step 22: Classify the extracted task characteristics using a support vector machine, and divide the tasks into three categories: compute-intensive, I / O-intensive, and hybrid; Step 23: For the different types of tasks after classification, use a linear regression model to predict their resource requirements.

4. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 3, wherein The method for the priority adjustment module to establish a dynamic task priority adjustment mechanism is: Extract the waiting time in the task queue status information of the system state vector, and adaptively adjust the task priorities according to the waiting time and the system state vector; The adaptive adjustment of task priorities includes: The priority adjustment module consists of a resource competition analysis unit, a priority calculation unit, and a priority update unit; The resource competition analysis unit calculates the current system resource competition degree based on the system resource status sub-vector and the network resource status sub-vector in the system status vector; the system resource competition degree is described by a resource competition index, and this resource competition index is calculated by weighted calculation considering factors such as CPU usage rate, memory occupancy rate, network bandwidth usage rate, and storage space usage rate, so as to quantify the tightness of system resources; Among them, the priority calculation unit calculates the dynamic priority adjustment factor of each task based on the waiting time of the task, the resource competition index, and the resource demand of the task; the dynamic priority adjustment factor is calculated by an adaptive weighting method; The calculation method of the dynamic priority adjustment factor implemented by the priority calculation unit includes the following sub-steps: Step 221: Construct a waiting time compensation function. As the waiting time of the task increases, the compensation value increases non-linearly to prevent the task from starving for a long time; Step 222: Construct a resource demand penalty function. Appropriately reduce the priority of tasks with large resource demands, but set an upper limit to avoid high-resource-demand tasks from never being executed; Step 223: Construct a system load adaptive function to dynamically adjust the weights of waiting time compensation and resource demand penalty according to the current system load status; Step 224: Weight and fuse the task priority, waiting time compensation value, resource demand penalty value, and system load adaptive factor in the queue status information to calculate the final dynamic priority adjustment factor.

5. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 4, characterized in that, The genetic algorithm adopted by the macro scheduling layer in the double-layer scheduling decision module includes the following steps: Step 31: Receive the system status vector and the adjusted task priority information, and design a chromosome coding method for the global resource allocation optimization of the macro scheduling layer to represent the resource allocation scheme; Step 32: Design a fitness function composed of a target function construction unit, a weight allocation unit, and a normalization processing unit for the global resource allocation optimization of the macro scheduling layer, considering system throughput and load balance; The design of the corresponding fitness function for the genetic algorithm includes the following sub-steps: Step 321: Based on the chromosome coding method, construct a system throughput evaluation function composed of a task execution time estimation unit, a parallelism analysis unit, and a throughput calculation unit to calculate the number of tasks completed per unit time; Step 322: Based on the chromosome coding method, construct a load balance degree evaluation function composed of a resource utilization rate calculation unit, a variance calculation unit, and a balance degree scoring unit to calculate the variance of each resource utilization rate; Step 323: Based on the chromosome coding method, construct an energy consumption evaluation function composed of a resource energy consumption model unit, a task energy consumption calculation unit, and a system energy consumption optimization unit to calculate the energy consumption of the resource allocation scheme; Step 324: Based on the chromosome coding method, construct a delay evaluation function composed of a task scheduling sequence generation unit, a critical path analysis unit, and a response time calculation unit to calculate the average task completion time; Step 33: Introduce an adaptive crossover and mutation operator composed of a population diversity evaluation unit, a parameter adaptive adjustment unit, and an operation execution unit for the global resource allocation optimization of the macro-scheduling layer, and dynamically adjust the parameters according to the population diversity; Step 34: Design an elite retention strategy composed of an elite selection unit, an elite preservation unit, and an elite re-introduction unit for the global resource allocation optimization of the macro-scheduling layer.

6. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 5, wherein The micro-scheduling layer in the double-layer scheduling decision module uses a reinforcement learning model for real-time scheduling decisions, including the following steps: Step 35: Receive the system state vector and the updated task queue state information, and define the state space of the micro-scheduling layer; Step 36: Based on the resource allocation scheme of the macro-scheduling layer, define the action space of the micro-scheduling layer as specific resource allocation decisions; Step 37: Based on the energy consumption evaluation function and the delay evaluation function, design the reward function of the micro-scheduling layer as the weighted sum of the task completion time and the energy consumption; Step 38: Based on the defined state space, action space, and reward function, use the Q-learning algorithm to optimize the strategy of the reward function.

7. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 6, characterized in that The design of the chromosome coding method includes the following sub-steps: Step 311: Design a chromosome structure based on integer coding to represent the mapping relationship between the predicted task resource requirements and the system resources; Step 312: Based on the chromosome structure, design a chromosome segmentation strategy to encode the allocation schemes of different types of resources respectively; Step 313: Based on the segmentation strategy, design a constraint condition coding method composed of a hard constraint coding unit, a soft constraint coding unit, and a constraint priority unit to ensure that the resource allocation meets the system limitations; Step 314: Based on the task priority information, design a priority coding method composed of a priority mapping unit, an execution order coding unit, and an importance coding unit to reflect the execution order and importance of the tasks.

8. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 7, characterized in that Introducing the adaptive crossover and the mutation operator for the global resource allocation optimization of the macro-scheduling layer includes the following sub-steps: Step 331: Based on the evaluation results of the fitness function, construct a fitness difference analysis unit, a crossover probability calculation unit, and an individual selection unit to dynamically adjust the crossover probability according to the fitness differences of the population individuals; Step 332: Based on the results of the dynamically adjusted crossover probability, construct an evolution state monitoring unit, a convergence speed calculation unit, and a mutation probability adjustment unit to dynamically adjust the mutation probability according to the number of generations of evolution and the convergence speed; Step 333: Based on the chromosome coding method, design a dedicated crossover operation for the resource allocation problem composed of a crossover point selection unit, an intra-segment crossover unit, and a validity repair unit; Step 334: Based on the dedicated crossover operation, design a local search mutation operation for scheduling optimization composed of a mutation position selection unit, a neighborhood generation unit, and an optimal replacement unit.

9. The adaptive adjustment system of the measurement and control system based on deep learning according to claim 8, wherein The top-down guidance feedback includes the following steps: Step 41: Receive the resource allocation scheme optimized by the macro-scheduling layer. The guidance information receiving unit, constraint condition conversion unit, and constraint condition application unit together use the resource allocation scheme as the constraint condition for the micro-scheduling layer; Step 42: The macro-scheduling layer regularly transmits the evaluated global optimization objective and resource usage boundary to the micro-scheduling layer.

10. The adaptive adjustment method for the measurement and control system based on deep learning is implemented based on the adaptive adjustment system for the measurement and control system based on deep learning described in any one of claims 1-9, and is characterized in that, It includes the following steps: Step 1: Collect the real-time operation data of the measurement and control system, and construct the preprocessed real-time operation data into a system state vector; Step 2: Based on the system state vector, extract the features of the tasks in the measurement and control system, classify the tasks using a support vector machine, predict the resource requirements for different types of tasks, and dynamically adjust the task priorities based on the resource requirements to generate task priority information; Step 3: Receive the system state vector and the adjusted task priority information through the pre-constructed macro-scheduling layer, and use a genetic algorithm to optimize the global resource allocation; based on the optimization results of the macro-scheduling layer through the pre-constructed micro-scheduling layer, use a reinforcement learning model to make real-time scheduling decisions; and optimize the macro-scheduling layer and the micro-scheduling layer using top-down guidance feedback and bottom-up correction feedback; Step 4: Generate specific resource allocation instructions based on the resource allocation scheme output after the optimization of the macro-scheduling layer.

Images