CN-122018323-A - Self-adaptive control method for operation parameters of artificial intelligent module

CN122018323ACN 122018323 ACN122018323 ACN 122018323ACN-122018323-A

Abstract

The invention discloses an artificial intelligent module operation parameter self-adaptive control method, which relates to the technical field of parameter control and comprises the steps of predicting future operation index values of an artificial intelligent module under different operation parameter adjustment amounts based on a state characterization vector and a preliminary control vector, constructing a multi-objective optimization problem by taking the preliminary control vector as an initial point and the future operation index values as optimization purposes, carrying out real-time search in a pareto solution set of the multi-objective optimization problem by adopting a multi-objective optimization algorithm to obtain a final control vector, and sending the final control vector to an execution mechanism of the artificial intelligent module for physical adjustment.

Inventors

MA CHAO
WU FANGMIN
ZHAO DEJIE

Assignees

苏州贯文存储科技有限公司
苏州燧火智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260402

Claims (10)

1. The self-adaptive control method for the operation parameters of the artificial intelligent module is characterized by comprising the following steps of, Acquiring multi-mode feedback data of the artificial intelligent module in real time, and constructing a space-time state model; Extracting a state characterization vector for characterizing the current running state of the artificial intelligent module based on the space-time state model; Based on the state characterization vector and the multi-mode feedback data, identifying the current operation working condition and generating a working condition feature vector; performing self-adaptive control law calculation on the state representation vector and the working condition characteristic vector to obtain a preliminary control vector for adjusting the operation parameters of the artificial intelligent module; Predicting future operation index values of the artificial intelligent module under different operation parameter adjustment amounts based on the state representation vector and the preliminary control vector; constructing a multi-objective optimization problem by taking the preliminary control vector as an initial point and the future operation index value as an optimization purpose; performing real-time search in the pareto solution set of the multi-objective optimization problem by adopting a multi-objective optimization algorithm to obtain a final control vector; and transmitting the final control vector to an actuating mechanism of the artificial intelligent module for physical adjustment.
2. The method for adaptively controlling operating parameters of an artificial intelligence module according to claim 1, wherein said multi-modal feedback data comprises time series monitoring data, event record data and topology data describing connection relations between components.
3. The method for adaptively controlling operation parameters of an artificial intelligence module according to claim 2, wherein the construction of the space-time state model is specifically as follows: constructing a basic graph structure by taking a hardware component of the artificial intelligent module as a physical node and a connection relation between the hardware components represented by topology data as an edge; And taking the time sequence monitoring data as a time sequence attribute sequence of the corresponding physical node, carrying out natural language processing on the event record data, extracting key events, and integrating the key events as attribute information into a basic diagram structure to form a space-time state model.
4. The method for adaptively controlling operation parameters of an artificial intelligent module according to claim 3, wherein said extracting a state characterization vector representing a current operation state of the artificial intelligent module is performed by performing an association analysis process on physical nodes in said space-time state model, and performing a time evolution feature extraction in combination with a time sequence attribute, so as to obtain a state characterization vector.
5. The method for adaptively controlling operation parameters of an artificial intelligence module according to claim 4, wherein the identifying the current operation condition based on the state characterization vector and the multi-mode feedback data generates a condition feature vector, specifically: carrying out statistical analysis on the state characterization vector, extracting a high-order statistical feature vector, and simultaneously carrying out trend analysis on time sequence monitoring data in multi-mode feedback data, and extracting a time sequence mode feature vector; splicing the high-order statistical feature vector and the time sequence mode feature vector to form an original working condition description feature vector; Performing linear transformation and dimension reduction on the original working condition description feature vector through principal component analysis, and selecting the first k principal components as basic components; Matching the similarity between the basic component and the template features in the working condition template library, selecting the working condition template with the highest matching degree, and extracting the corresponding template code; And carrying out weighted fusion on the template codes and the basic components to generate the working condition feature vector.
6. The method for adaptively controlling the operation parameters of the artificial intelligent module according to claim 5, wherein the adaptive control law calculation is performed on the state characterization vector and the working condition feature vector to obtain a preliminary control vector for adjusting the operation parameters of the artificial intelligent module, and specifically: Mapping the state characterization vector and the working condition feature vector into control law modulation parameters; Characteristic modulation is carried out on the state characterization vector by utilizing the control law modulation parameters, and conditional state characteristics are generated; And performing control strategy calculation on the conditional state characteristics to obtain a preliminary control vector.
7. The method for adaptively controlling the operation parameters of the artificial intelligent module according to claim 6, wherein the predicting future operation index values of the artificial intelligent module under different operation parameter adjustment amounts based on the state characterization vector and the preliminary control vector comprises the following steps: Splicing the state characterization vector and the preliminary control vector to form a joint feature vector; Performing high-order polynomial expansion on the combined feature vectors to generate nonlinear feature combinations; taking the nonlinear feature combination vector as input, and calculating a future operation index value by applying a group of predefined index prediction functions; The future operation index values include a power consumption index value, a delay index value, and a temperature index value.
8. The method for adaptively controlling operation parameters of an artificial intelligence module according to claim 7, wherein said constructing a multi-objective optimization problem is specifically: taking the preliminary control vector as a starting point of optimizing search, and respectively defining a power consumption index value, a delay index value and a temperature index value as objective functions needing to be minimized; setting a corresponding constraint boundary for each objective function; The constraint boundary comprises power consumption upper limit constraint, delay threshold constraint and temperature safety constraint, and the objective function and the constraint boundary are integrated to form a multi-objective optimization problem.
9. The method for adaptively controlling operation parameters of an artificial intelligent module according to claim 8, wherein the method is characterized in that the method adopts a multi-objective optimization algorithm to search in real time in a pareto solution set of a multi-objective optimization problem to obtain a final control vector, and specifically comprises the following steps: decomposing the multi-objective optimization problem into a plurality of scalar sub-problems by using a decomposition-based multi-objective evolutionary algorithm, and maintaining a candidate solution population in the neighborhood of each scalar sub-problem; initializing a candidate solution population by using the preliminary control vector, and then executing crossover and mutation operations to generate a new solution; calculating a power consumption index value, a delay index value and a temperature index value corresponding to each new solution by using an evaluation function, updating the candidate solution population according to the dominant relationship and the aggregation function, and iteratively executing the operations of crossing, mutation, evaluation and updating until the termination condition is met, so as to obtain a final candidate solution population; And after non-dominant sorting is carried out on the final candidate solution population, selecting an initial control vector meeting all constraint conditions at the pareto front as a final control vector.
10. The method for adaptively controlling operation parameters of an artificial intelligence module according to claim 9, wherein the step of issuing the final control vector to an actuator of the artificial intelligence module for physical adjustment comprises the steps of: Analyzing the final control vector into an adjustable parameter instruction set of the artificial intelligent module; and sequentially sending the adjustable parameter instruction set to an execution mechanism of the artificial intelligent module through a hardware communication interface to carry out parameter adjustment.

Description

Self-adaptive control method for operation parameters of artificial intelligent module Technical Field The invention relates to the technical field of parameter control, in particular to an artificial intelligent module operation parameter self-adaptive control method. Background With the rapid development of artificial intelligence technology, the deployment of the AI module in a data center, edge computing equipment and a special acceleration platform is increasingly wide, so that the dynamic regulation and control of operation parameters becomes a key link for guaranteeing the performance and reliability of the module in order to meet the severe requirements of diversified task loads on energy efficiency, delay and thermal management, in recent years, a parameter regulation method based on feedback control, reinforcement learning and heuristic strategies is sequentially proposed, and part of schemes try to introduce operation monitoring data to realize a certain degree of self-adaptability, thereby not only pushing the AI module to progress towards higher-level autonomy and intelligent evolution, but also laying an important foundation for constructing an efficient and reliable artificial intelligent operation environment. Nevertheless, the existing intelligent module operation parameter control method still has improvement, firstly, the space-time coupling relation among the modules cannot be effectively modeled, and especially key event semantic information contained in logs is ignored, so that the working condition identification capability is limited, secondly, the control strategy lacks a conditioning mechanism, and decision logic cannot be dynamically adjusted according to the operation working condition, so that the strategy generalization capability is weak, and performance jitter is easy to occur in a new load mode. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides an artificial intelligent module operation parameter self-adaptive control method which solves the problems of limited working condition identification capability and weak strategy generalization capability. In order to solve the technical problems, the invention provides the following technical scheme: the invention provides an artificial intelligent module operation parameter self-adaptive control method, which comprises, Acquiring multi-mode feedback data of the artificial intelligent module in real time, and constructing a space-time state model; Extracting a state characterization vector for characterizing the current running state of the artificial intelligent module based on the space-time state model; Based on the state characterization vector and the multi-mode feedback data, identifying the current operation working condition and generating a working condition feature vector; performing self-adaptive control law calculation on the state representation vector and the working condition characteristic vector to obtain a preliminary control vector for adjusting the operation parameters of the artificial intelligent module; Predicting future operation index values of the artificial intelligent module under different operation parameter adjustment amounts based on the state representation vector and the preliminary control vector; constructing a multi-objective optimization problem by taking the preliminary control vector as an initial point and the future operation index value as an optimization purpose; performing real-time search in the pareto solution set of the multi-objective optimization problem by adopting a multi-objective optimization algorithm to obtain a final control vector; and transmitting the final control vector to an actuating mechanism of the artificial intelligent module for physical adjustment. As a preferable scheme of the self-adaptive control method for the operation parameters of the artificial intelligent module, the multi-mode feedback data comprises time sequence monitoring data, event recording data and topology data describing the connection relation among components. As a preferable scheme of the self-adaptive control method for the operation parameters of the artificial intelligent module, the invention comprises the following steps of constructing a space-time state model: constructing a basic graph structure by taking a hardware component of the artificial intelligent module as a physical node and a connection relation between the hardware components represented by topology data as an edge; And taking the time sequence monitoring data as a time sequence attribute sequence of the corresponding physical node, carrying out natural language processing on the event record data, extracting key events, and integrating the key events as attribute information into a basic diagram structure to form a space-time state model. The invention is used as a preferred scheme of the self-adaptive control m