An AI-assisted machine room equipment adaptive regulation and management method

Abstract

The application discloses an AI-assisted machine room equipment adaptive regulation and management method, belonging to the field of computer science equipment management. The method first collects machine room equipment power, temperature, usage time and other data and preprocesses, then evaluates the equipment state and classifies, then formulates regulation strategies for different state equipment, then optimizes resource allocation with the help of AI technology, and finally constructs an evaluation system from the equipment running state, resource utilization efficiency and business task completion condition to evaluate and feedback the system performance. Through improved discrete wavelet transform, deep belief network and other technologies, the equipment state is accurately evaluated, and potential faults are warned in advance. Reinforcement learning, genetic algorithm and other optimization resource allocation technologies are used to improve equipment utilization and business completion quality. The multi-dimensional evaluation index system and threshold adjustment strategy ensure the stable operation of the system, reduce the operation and maintenance cost, and improve the overall operation efficiency and adaptability of the machine room.

Description

Technical Field

[0001] This invention relates to the field of computer science equipment management, and in particular to an AI-assisted adaptive control and management method for computer room equipment. Background Technology

[0002] In today's rapidly developing information technology landscape, data centers, as the core locations for data storage, processing, and transmission, require stable operation and efficient management of their equipment. As data centers continue to expand in size and the number and types of equipment increase, traditional data center equipment management methods face numerous challenges:

[0003] Traditional methods are insufficient for comprehensively and accurately monitoring the operational status of a large number of different types of equipment in real time. Data center equipment has numerous operating parameters such as power, temperature, and usage time. Relying on manual inspections and simple monitoring methods makes it impossible to promptly detect potential faults or scientifically assess the overall health of the equipment. This can lead to equipment failures that may cause business interruptions and economic losses.

[0004] In terms of resource allocation, there is a lack of effective intelligent methods. The equipment load in the data center is uneven, with some equipment being overused while others are underutilized. At the same time, the real-time load of the equipment and the urgency of the tasks are not fully considered when allocating tasks, resulting in wasted resources. Furthermore, critical business tasks have slow response speeds and low completion quality, failing to meet the needs of business development.

[0005] In the face of equipment failures and potential failure risks, traditional control strategies are often reactive and simplistic, lacking specificity and dynamic adjustment capabilities. When equipment fails, the allocation of maintenance resources may be untimely, resulting in long repair times; for potentially faulty equipment, effective intervention cannot be carried out in advance, and adaptive control based on the real-time status and changing trends of the equipment is impossible.

[0006] Existing data center system performance evaluation metrics are incomplete, typically focusing only on single-dimensional indicators such as equipment operating status or resource utilization efficiency, while neglecting important aspects such as business task completion. Furthermore, the evaluation methods are not scientific enough to accurately reflect the overall system performance and cannot provide a reliable basis for system optimization.

[0007] Therefore, this invention proposes an AI-assisted adaptive control and management method for data center equipment. Summary of the Invention

[0008] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0009] In view of the problems existing in the above and / or prior art, the present invention is proposed.

[0010] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide an AI-assisted adaptive control and management method for data center equipment, which includes the following steps:

[0011] S100, Data Acquisition and Preprocessing;

[0012] S200, Equipment Condition Assessment and Classification;

[0013] S300, regulation and control strategy formulation;

[0014] S400, AI-assisted resource optimization and allocation;

[0015] S500, system performance evaluation and feedback adjustment;

[0016] Preferably, in step S100, data acquisition and preprocessing, core operating data such as power, temperature, and usage time are collected for different types of equipment to construct a device status dataset D = {x}. i y i )|i=1,2,…,|D|}. Where, Represents various operating parameters of the equipment, y i This corresponds to the device identifier or status label.

[0017] In the data acquisition phase, let the measurement error of the j-th sensor at time t be . in The average of the measurements from multiple sensors, i.e. Then the adaptive weight w j The formula for calculating (t) is:

[0018]

[0019] For temperature data T(t), a modified Discrete Wavelet Transform (DWT) is used for feature extraction and denoising. In this method, the wavelet basis function is improved, and the improved wavelet coefficient calculation formula is as follows:

[0020]

[0021] in, a m and b n These are the scaling and translation parameters in the original discrete wavelet transform.

[0022] When collecting device usage time data, an autoregressive moving average (ARMA) model is used to further process the generated data. In this method, data generated by a generative adversarial network is incorporated into the ARMA model, resulting in the following improved model:

[0023]

[0024] Where w is the weighting coefficient, used to adjust the degree of influence of the generated data in the model.

[0025] Preferably, step S200, equipment status assessment and classification, includes the following steps:

[0026] S210, Feature extraction optimization based on deep belief network (DBN);

[0027] S220, Equipment Status Assessment Incorporating Improved Grey Relational Analysis (GRA);

[0028] S230, Device status classification based on density peak clustering (DPC);

[0029] In steps S200, S210 of equipment status assessment and classification, and feature extraction optimization based on Deep Belief Network (DBN), a Deep Belief Network (DBN) model is constructed using the equipment operation data processed in S100. Hidden layer neurons h j activation probability P(h) j =1|x) is calculated using the following formula:

[0030] P(h j =1|x)=σ(b j +∑ i w ij x i )

[0031] in, It is the sigmoid activation function.

[0032] Based on traditional DBN, an adaptive learning rate η(t) is introduced to accelerate model convergence and improve feature extraction performance. The formula is:

[0033]

[0034] Where η0 is the initial learning rate, and α is the adjustment parameter. The gradient is the gradient at the k-th training iteration.

[0035] In the equipment status assessment described in S200, S220 of equipment status assessment and classification, and the improved grey relational analysis (GRA) fusion equipment status assessment, the improved correlation coefficient calculation formula is as follows:

[0036]

[0037] Where, ζ i (k) is the correlation coefficient between the i-th comparison sequence and the reference sequence at time k, and ρ∈[0,1] is the resolution coefficient.

[0038] Grey relational degree r of the device i for:

[0039] In steps S200, S230 of equipment condition assessment and classification, and equipment condition classification based on density peak clustering (DPC), an adaptive cutoff distance adjustment factor λ is introduced. The cutoff distance is dynamically adjusted according to the distribution of data points, using the following formula:

[0040]

[0041] Preferably, step S300, the formulation of the control strategy, includes the following steps:

[0042] S310, Reinforcement learning control strategy for potentially faulty equipment;

[0043] S320, Normal equipment resource allocation strategy based on genetic algorithm optimization;

[0044] S330, Development of rapid response and repair strategies for faulty equipment;

[0045] In S300, S310 of the control strategy formulation, and the reinforcement learning control strategy for potentially faulty equipment, the Q-value update formula is as follows:

[0046] Q(s,a)←Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)]

[0047] Where α is the learning rate and γ is the discount factor.

[0048] In this method, a dual-network structure and a priority experience replay mechanism are introduced to improve algorithm performance and decision accuracy. The dual-network structure includes an evaluation network Q and a target network Q. target The target network parameters are updated periodically. A priority experience replay mechanism samples based on the importance of the experience samples, with importance determined by the TD error δ = r + γmax. a′ Q target (s′,a′)-Q(s,a) is used as a metric. The sampling probability P(i) of a sample is calculated using the following formula: Where β is a parameter that controls the sampling bias, and 0 < β < 1.

[0049] In S300, S320 (the regulation strategy formulation), and the normal equipment resource allocation strategy optimized by genetic algorithm, a real-time load balancing factor L for the equipment is introduced. i And the urgency factor E of the task j The fit is adjusted. The adjusted fit w′ ij The calculation formula is:

[0050]

[0051] Among them, L max λ1 and λ2 are the maximum values ​​of the load on all devices, and λ1 and λ2 are weighting coefficients.

[0052] In the S300, S330 (for formulating the control strategy), and the rapid response and repair strategy for faulty equipment, a resource constraint factor R is introduced. If maintenance resources are limited, it may be necessary to wait for resources to become available before repairs can proceed. Let T0 be the repair time when maintenance resources are sufficient. The formula for calculating the actual repair time T is:

[0053] T = T0·(1+λ3·(1+R)), where λ3 is the weighting coefficient.

[0054] Preferably, the S400 and AI-assisted resource optimization allocation includes the following steps:

[0055] S410. Construct a resource allocation model based on reinforcement learning;

[0056] S420, global optimization combined with simulated annealing algorithm;

[0057] S430, Dynamic resource allocation adjustment based on machine learning prediction;

[0058] In S400, S410 of AI-assisted resource optimization allocation, and the construction of a resource allocation model based on reinforcement learning, the optimal resource allocation strategy is learned by interacting with the environment (the system consisting of data center equipment and business needs).

[0059] Define the state space as s = [e1, e2, ..., e] n ,t1,t2,...,t m a1, a2, ..., a n×m ], where e i t represents the status information of the i-th device. j This represents the requirement information for the j-th task, a. ij This indicates the current allocation relationship between device i and task j (the value is 0 or 1, indicating whether allocation has occurred).

[0060] Define the action space: Action 'a' represents the decision to allocate a specific device to a specific task. For n devices and m tasks, the action space consists of n×m possible allocation decisions. The agent selects an action each time to adjust resource allocation.

[0061] Define the reward function: The reward function is designed as follows:

[0062]

[0063] Among them, t j，completed Indicates whether task j is completed (1 for completed, 0 for incomplete), w j The weight of task j reflects its importance. i,stable u represents the operational stability of device i (higher stability values ​​are larger). i It is the utilization rate of device i; a ij This represents the allocation relationship between device i and task j, c ij α is the cost of assigning device i to task j (e.g., device setup cost, transmission delay, etc.). α, β, and γ are weighting coefficients used to balance the impact of different factors on the reward.

[0064] The weights of each element in the state vector are calculated using an attention mechanism, as shown in the following formula:

[0065]

[0066] Where q(s, a) is the Q-value of taking action a in state s, and f(s) k a) is the function value of the k-th element in the state vector related to action a, w k It is the attention weight of the k-th element, calculated as follows:

[0067]

[0068] Among them, v i These are learnable parameters.

[0069] In S400, S420 of AI-assisted resource optimization allocation, and global optimization combined with simulated annealing algorithm, the probability P of accepting the new solution is determined by the Metropolis criterion:

[0070]

[0071] As the algorithm progresses, the temperature T is gradually decreased, as shown by the formula T = T0·α. k , where T0 is the initial temperature, α is the cooling coefficient (0 < α < 1), and k is the number of iterations.

[0072] In S400, S430 of AI-assisted resource optimization allocation, and dynamic resource allocation adjustment based on machine learning prediction, a Long Short-Term Memory (LSTM) network is used to perform time-series prediction of the device's operating status and business needs. The core structure of LSTM includes an input gate i t Forgotten Gate f t Output gate o t and memory unit C t At time t, the calculation formula is as follows:

[0073] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )

[0074] f i =σ(W if x t +b if +W hf h t-1 +b hf )

[0075] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0076] C t =f t ·C t-1 +i t ·tanh(W ic x t +b ic +W hc h t-1 +b hc )

[0077] h t =o t ·tanh(C t )

[0078] Where, x t It is the input at time t, h t-1 σ is the hidden state at the previous time step, W and b are the weight matrix and bias vector, and σ is the sigmoid function.

[0079] Preferably, the S500 system performance evaluation and feedback adjustment includes the following steps:

[0080] S510. Construct a multi-dimensional system performance evaluation index system;

[0081] S520, Construction of Comprehensive Performance Evaluation Model;

[0082] S530, Threshold-based adjustment strategy;

[0083] In the S510 section, a multi-dimensional system performance evaluation index system is constructed, which is based on three key dimensions: equipment operating status, resource utilization efficiency, and business task completion.

[0084] Equipment operating status dimension: A comprehensive equipment health index is introduced to measure the overall operating status of the equipment. Let the set of factors affecting equipment health be F = {f1, f2, ..., f...} n}, construct the judgment matrix A(a ij ), where a ij This indicates the importance of factor i relative to factor j. The eigenvector W = [w1, w2, ..., wj] of the judgment matrix is ​​calculated. n The weights of each factor are obtained, satisfying the following conditions: The formula for calculating the overall health index of equipment is: Among them, f i，min and f i,max These are the minimum and maximum values ​​of factor i, respectively. This normalization process makes different factors comparable.

[0085] Resource utilization efficiency dimension: The coefficient of variation (CV) of resource utilization rate for each piece of equipment is calculated using the coefficient of variation method. i Equipment resource utilization rate U i The coefficient of variation can be calculated by the ratio of the actual equipment resources used in the task to the total equipment resources. Where, σU i Equipment resource utilization rate U i standard deviation It is their average value. The formula for calculating the resource utilization equilibrium degree is: The closer the RUBD value is to 1, the more balanced the resource utilization; the closer it is to 0, the more uneven the resource allocation.

[0086] Business task completion dimension: calculated using fuzzy comprehensive evaluation method, assuming the set of evaluation factors is U = {u1, u2, ..., u...} s The evaluation level set is V = {v1, v2, ..., v} (completion time, completion quality, urgency, etc.). t}(Very satisfied, satisfied, neutral, dissatisfied, very dissatisfied), construct a fuzzy relation matrix R = (r ij ), where r ij Indicates factor ui Evaluation level v j The membership degree. Then, the weight vector A = [a1, a2, ..., a] of each evaluation factor is determined using the analytic hierarchy process. s The fuzzy comprehensive evaluation result of business task satisfaction is: BTS = A·R, and the result is a 1×t vector.

[0087] In the construction of the S520 comprehensive performance evaluation model, firstly, the information entropy Ej of each indicator is calculated. For indicator j, the formula for calculating its information entropy is: Where n is the number of samples, x ij This represents the value of the i-th sample on index j. Then, the entropy weight w for each index is calculated. j :

[0088] Finally, the formula for calculating the overall system performance evaluation index is as follows: Among them, I1, I2, and I3 are the Comprehensive Health Index of Equipment (DCHI) in the dimension of equipment operation status, the Resource Utilization Efficiency Balance (RUBD) in the dimension of resource utilization efficiency, and the Business Task Satisfaction Index (BTS) in the dimension of business task completion, respectively.

[0089] S530, threshold-based adjustment strategy: Set different levels of SCPI thresholds, such as SCPI high SCPI medium SCPI low When SCPI ≥ SCPI high When the system is running well, the frequency of resource allocation adjustments can be appropriately reduced, but continuous monitoring is still necessary; when SCPI... medium ≤SCPI≤SCPI high At that time, make minor adjustments to the system.

[0090] This invention discloses an AI-assisted adaptive control and management method for computer room equipment, which has the following beneficial effects:

[0091] 1. By comprehensively collecting core operational data such as equipment power, temperature, and usage time, and using improved discrete wavelet transform and deep belief network technologies for data processing and feature extraction, the system can accurately assess equipment status, promptly identify potential faulty equipment, and provide early warnings. This greatly improves the reliability of equipment operation, reduces the risk of business interruption due to equipment failure, and ensures the continuity of data center services.

[0092] 2. By utilizing technologies such as reinforcement learning and genetic algorithms, combined with factors such as real-time load balancing factors and task urgency factors, the system achieves rational allocation and optimized distribution of normal equipment resources. This not only improves equipment utilization and reduces resource waste, but also prioritizes the resource needs of critical business tasks, improving the quality and response speed of task completion and enhancing the overall operational efficiency of the data center.

[0093] 3. Reinforcement learning-based control strategies, genetic algorithm-optimized resource allocation strategies, and rapid response and repair strategies were developed for potentially faulty equipment, normal equipment, and faulty equipment, respectively. These strategies can adaptively adjust according to different equipment states. When equipment malfunctions, resource constraint factors are introduced to rationally allocate maintenance resources, shorten maintenance time, and reduce operation and maintenance costs.

[0094] 4. A multi-dimensional system performance evaluation index system was constructed, comprehensively evaluating system performance from three key dimensions: equipment operating status, resource utilization efficiency, and business task completion. Threshold-based adjustment strategies can promptly adjust the system based on the evaluation results, ensuring the system is always in good operating condition, improving the stability and adaptability of the data center system, and providing strong support for continuous optimization of the data center. Attached Figure Description

[0095] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0096] Figure 1 This is a diagram of the AI-assisted adaptive control and management method for data center equipment according to the present invention;

[0097] Figure 2 This is a framework diagram of the data acquisition and preprocessing method of the present invention;

[0098] Figure 3 This is a framework diagram of the equipment condition assessment and classification method of the present invention;

[0099] Figure 4 This is a framework diagram of the regulation strategy formulation method of the present invention;

[0100] Figure 5 This is a framework diagram of the AI-assisted resource optimization allocation method of the present invention;

[0101] Figure 6 This is a framework diagram of the system performance evaluation and feedback adjustment method of the present invention;

[0102] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0103] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings of the embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0104] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the examples in the specification.

[0105] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0106] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0107] This invention provides an AI-assisted adaptive control and management method for data center equipment, comprising the following steps:

[0108] S100, Data Acquisition and Preprocessing;

[0109] S200, Equipment Condition Assessment and Classification;

[0110] S300, regulation and control strategy formulation;

[0111] S400, AI-assisted resource optimization and allocation;

[0112] S500, system performance evaluation and feedback adjustment;

[0113] In a preferred embodiment of the present invention, in S100, data acquisition and preprocessing, a comprehensive data acquisition system is constructed using multiple high-precision sensors in the data acquisition section of the computer room equipment. For different types of equipment, core operating data such as power, temperature, and usage time are collected to construct an equipment status dataset D = {(x...} i y i )|i=1,2,…,|D|}. Where, Represents various operating parameters of the equipment, y iThis corresponds to the device identifier or status label.

[0114] In the data acquisition stage, to improve the accuracy and stability of power data P(t), an adaptive weighted fusion algorithm is introduced to combine the power data P1(t), P2(t), ..., P2(t) collected by multiple sensors. n (t) is used for fusion processing. Weight w j The system is dynamically adjusted based on the real-time measurement error of the sensors. Let the measurement error of the j-th sensor at time t be... in The average of the measurements from multiple sensors, i.e. Then the adaptive weight w j The formula for calculating (t) is:

[0115]

[0116] For temperature data T(t), a modified Discrete Wavelet Transform (DWT) is used for feature extraction and denoising. In this method, the wavelet basis function is improved by introducing an adaptive scaling factor α and an adaptive translation factor β. The improved wavelet coefficient calculation formula is as follows:

[0117]

[0118] in, a m and b n These are the scaling and translation parameters in the original discrete wavelet transform.

[0119] When collecting device usage time data, to address the issues of missing data and outliers, a method combining deep learning-based Generative Adversarial Networks (GANs) with time series analysis is not used. Let the time series data be U(t). First, the raw usage time data is preprocessed to extract the trend component U. trend (t) and fluctuation component U fluct (t). For missing or outlier data points, a generator G is used to generate candidate data based on known data. The generator's input is a vector z containing time information and known data features, and its output is G(z). A discriminator D judges the generated data G(z) against the real data, and the parameters of the generator and discriminator are continuously optimized through adversarial training. For time series analysis, the generated data is further processed using an autoregressive moving average (ARMA) model. In this method, the data generated by the generative adversarial network is integrated into the ARMA model, resulting in the improved model:

[0120]

[0121] Where w is the weighting coefficient, used to adjust the degree of influence of the generated data in the model.

[0122] In a preferred embodiment of the present invention, after the data collection and preprocessing of the computer room equipment is completed in S100, accurate, stable, and complete equipment operation data is obtained. In step S200, equipment status assessment and classification, based on this data, innovative algorithms and formulas are used to assess and classify the equipment status. Step S200, equipment status assessment and classification, includes the following steps:

[0123] S210, Feature extraction optimization based on deep belief network (DBN);

[0124] S220, Equipment Status Assessment Incorporating Improved Grey Relational Analysis (GRA);

[0125] S230, Device status classification based on density peak clustering (DPC);

[0126] In steps S200, S210 of equipment status assessment and classification, and feature extraction optimization based on Deep Belief Network (DBN), a Deep Belief Network (DBN) model is constructed using the equipment operation data processed in S100. The DBN is composed of multiple Restricted Boltzmann Machines (RBMs). For input layer neurons x... i and hidden layer neurons h j In RBM, the connection weights between them are w ij The biases are a i and b j Hidden layer neurons h j activation probability P(h) j =1|x) is calculated using the following formula:

[0127] P(h j =1|x)=σ(b j +∑ i w ij x i )

[0128] in, It is the sigmoid activation function.

[0129] Building upon traditional DBN, an adaptive learning rate η(t) is introduced to accelerate model convergence and improve feature extraction performance. The adaptive learning rate is dynamically adjusted based on gradient changes during training, and the formula is:

[0130]

[0131] Where η0 is the initial learning rate, and α is the adjustment parameter. The gradient is the gradient at the k-th training iteration.

[0132] In the equipment status assessment process described in S200, S220 (equipment status assessment and classification), and the improved grey relational analysis (GRA) method, after obtaining the feature vector of the equipment, an improved grey relational analysis method is used to assess the equipment status. Let the reference sequence X0 = {x0(1), x0(2), ..., x0(n)} represent the standard status characteristics of normal equipment, and compare the sequence X... i ={x i (1), x i (2), ..., x i Let (n)} be the feature vector of the device to be evaluated. Considering that different features have different degrees of influence on the device state, a feature weight vector W = [w1, w2, ..., wn] is introduced. n The improved formula for calculating the correlation coefficient is as follows:

[0133]

[0134] Where, ξ i (k) is the correlation coefficient between the i-th comparison sequence and the reference sequence at time k, and ρ∈[0,1] is the resolution coefficient.

[0135] Grey relational degree r of the device i for:

[0136] By calculating the grey relational degree between different devices and the standard state of normal devices, the higher the relational degree, the closer the device state is to the normal state; the lower the relational degree, the further the device state deviates from the normal state, which may indicate potential problems or malfunctions.

[0137] In steps S200, S230 of equipment condition assessment and classification, and equipment condition classification based on density peak clustering (DPC), the density peak clustering (DPC) algorithm is used to classify the equipment condition based on the grey relational analysis results. In the DPC algorithm, the local density ρ of each data point i... i Calculated using the following formula: Where, d ij d is the Euclidean distance between data points i and j. c It is the cutoff distance, which can be adaptively determined according to the data distribution, for example, taking the kth percentile of all distances (e.g., k=20).

[0138] The distance δ between data point i and the nearest point with a higher density. i ,for: Based on traditional DPC, to avoid the clustering results affecting the cutoff distance d c To address the over-reliance on data points, an adaptive cutoff distance adjustment factor λ is introduced. The cutoff distance is dynamically adjusted based on the distribution of data points, as shown in the formula:

[0139]

[0140] Based on local density ρ i and distance δ i A decision map is drawn, and points with high local density and distance are selected as cluster centers. Equipment is divided into different state categories, such as normal operation, potential failure, and failure. Equipment falling within the normal operation cluster is considered to be in good operating condition; equipment in the potential failure cluster, although not yet showing obvious failures, exhibits certain abnormal trends and requires close monitoring; and equipment in the failure category requires immediate repair or replacement.

[0141] In a preferred embodiment of the present invention, after completing S200, the equipment status assessment and classification, the different states of the equipment in the computer room have been clearly identified. This step will, based on these results, use innovative algorithms and formulas to formulate targeted control strategies. Step S300, control strategy formulation, includes the following steps:

[0142] S310, Reinforcement learning control strategy for potentially faulty equipment;

[0143] S320, Normal equipment resource allocation strategy based on genetic algorithm optimization;

[0144] S330, Development of rapid response and repair strategies for faulty equipment;

[0145] In steps S300, S310 (formulation of the control strategy), and the reinforcement learning control strategy for potentially faulty devices, for devices identified as potentially faulty in S200, a control strategy is formulated using a reinforcement learning algorithm based on a Deep Q-Network (DQN). In traditional DQN, the agent interacts with the environment, selects action a based on state s, obtains reward r, and transitions to a new state s′. The Q-value update formula is:

[0146] Q(s,a)←Q(s,a)+α[r+γmax a′ Q(s′,a')-Q(s,a)]

[0147] Where α is the learning rate and γ is the discount factor.

[0148] In this method, a dual-network structure and a priority experience replay mechanism are introduced to improve algorithm performance and decision accuracy. The dual-network structure includes an evaluation network Q and a target network Q. target The target network parameters are updated periodically. A priority experience replay mechanism samples based on the importance of the experience samples, with importance determined by the TD error δ = r + γmax. a′ Q target(s′,a′)-Q(s,a) is used as a metric. The sampling probability P(i) of a sample is calculated using the following formula: Where β is a parameter that controls the sampling bias, and 0 < β < 1.

[0149] In practical applications, state s consists of the device feature vector extracted by S200 and historical control actions. Action a includes adjusting device operating parameters (such as reducing power, increasing heat dissipation, etc.) and arranging maintenance plans. The reward r is set according to the degree of improvement of the device state. If the device state tends to stabilize, the reward is positive; conversely, if the state deteriorates, the reward is negative.

[0150] In S300, S320 (the regulation strategy formulation), and the normal equipment resource allocation strategy optimized by genetic algorithm, for normally operating equipment, a resource allocation strategy optimized by genetic algorithm is adopted to improve resource utilization. The equipment set is M = {m1, m2, ..., m}. n The task set is N = {n1, n2, ..., n}. m Equipment m i For task n j The compatibility is w ij Traditional resource allocation models aim to maximize matching value. Where x ij As a decision variable, when device m i Assigned to task n j time x ij =1, otherwise x ij =0. In this method, a real-time load balancing factor L for the device is introduced. i And the urgency factor E of the task j The fit is adjusted. The adjusted fit w′ ij The calculation formula is:

[0151]

[0152] Among them, L max λ1 and λ2 are the maximum values ​​of the load on all devices, and λ1 and λ2 are weighting coefficients.

[0153] Genetic algorithms find the optimal resource allocation strategy by performing selection, crossover, and mutation operations on chromosomes (resource allocation schemes). The selection operation uses a tournament selection method, randomly selecting k individuals from the population, and choosing the individual with the highest fitness to advance to the next generation. The crossover operation uses a crossover probability P... c Gene exchange is performed on two parent chromosomes to generate offspring chromosomes. The mutation operation has a mutation probability P. m Genes on chromosomes are randomly altered. The fitness function F is constructed based on the modified fitness:

[0154]

[0155] Through continuous evolution of the genetic algorithm, the optimal equipment resource allocation scheme is obtained, enabling normally operating equipment to meet task requirements while achieving efficient resource utilization and load balancing.

[0156] In S300, S330 of the control strategy formulation, and the formulation of the rapid response and repair strategy for faulty equipment, a rapid response and repair strategy based on time series prediction and resource constraints is formulated for equipment that has been identified as faulty. First, the equipment repair time T is predicted using time series analysis methods. Assume the equipment fault time series is t1, t2, ..., t... k The Autoregressive Integral Moving Average (ARIMA) model is used for prediction. The general form of the ARIMA(p, d, q) model is: Φ(B)(1-B) d X t =Θ(B)ε t in, It is the autoregressive part. Here, (1-B) is the autoregressive coefficient, and B is the shift operator. d It is a difference operator used to stationary time series, where d is the difference order, and Θ(B) = 1 + θ1B + θ2B 2 +...+θ q B q It is the moving average part, θ i It is the moving average coefficient, ε t It is a white noise sequence.

[0157] Considering the constraints of maintenance resources (such as maintenance personnel and spare parts), a resource constraint factor R is introduced. If maintenance resources are limited, it may be necessary to wait for resources to become available before maintenance can proceed. Let the repair time when maintenance resources are sufficient be T0, and the formula for calculating the actual repair time T is:

[0158] T = T0·(1+λ3·(1+R))

[0159] Where λ3 is the weighting coefficient.

[0160] Based on the predicted repair time and resource constraints, formulate a reasonable equipment replacement or backup equipment activation strategy. If the repair time is long and has a significant impact on business, activate backup equipment in a timely manner; if there is no suitable backup equipment, arrange the repair sequence reasonably according to resource availability, prioritizing the repair of equipment with the greatest impact on business, so as to reduce the impact of faulty equipment on the overall operation of the data center.

[0161] In a preferred embodiment of the present invention, after the control strategy is formulated in S300, the operating status of the equipment is effectively managed and maintained. However, in order to further improve the overall operating efficiency and resource utilization of the data center, this step S400, AI-assisted resource optimization allocation, uses AI technology to optimize resource allocation, ensuring that equipment resources can accurately match business needs, while coordinating with the previous control strategy. This step S400, AI-assisted resource optimization allocation, includes the following steps:

[0162] S410. Construct a resource allocation model based on reinforcement learning;

[0163] S420, global optimization combined with simulated annealing algorithm;

[0164] S430, Dynamic resource allocation adjustment based on machine learning prediction;

[0165] In steps S400, S410 (AI-assisted resource optimization and allocation), and the construction of a reinforcement learning-based resource allocation model, the resource allocation model is built based on the equipment state control results in S300, using a deep reinforcement learning algorithm. In this model, the intelligent agent (AI system) learns the optimal resource allocation strategy by interacting with the environment (the system consisting of data center equipment and business requirements).

[0166] Define the state space: The state vector s consists of real-time device status information (including device type, current operating status, remaining service life, etc., which are acquired and updated during S200 device status assessment and classification and S300 control strategy formulation), business task requirement information (task type, urgency, resource requirements, etc.), and the current resource allocation. That is, s = [e1, e2, ..., e...]. n ,t1, t2, ...,t m a1, a2, ..., a n×m ], where e i t represents the status information of the i-th device. j This represents the requirement information for the j-th task, a. ij This indicates the current allocation relationship between device i and task j (the value is 0 or 1, indicating whether allocation has occurred).

[0167] Define the action space: Action 'a' represents the decision to allocate a specific device to a specific task. For n devices and m tasks, the action space consists of n×m possible allocation decisions. The agent selects an action each time to adjust resource allocation.

[0168] Define the reward function: The reward function *r* measures the impact of the agent's actions on system performance. Drawing inspiration from the S300's approach to improving device state and resource utilization efficiency, the reward function is designed as follows:

[0169]

[0170] Among them, t j，completed Indicates whether task j is completed (1 for completed, 0 for incomplete), w j The weight of task j reflects its importance. i,stable u represents the operational stability of device i (higher stability values ​​are larger). i It is the utilization rate of device i; a ij This represents the allocation relationship between device i and task j, c ij α is the cost of assigning device i to task j (e.g., device setup cost, transmission delay, etc.). α, β, and γ are weighting coefficients used to balance the impact of different factors on the reward.

[0171] The model is trained using the Deep Q-Network (DQN) algorithm. An innovation is made to the traditional DQN by introducing an attention mechanism to enhance the agent's focus on key information. When calculating the Q-value, the weights of each element in the state vector are calculated using the attention mechanism, as shown in the following formula:

[0172]

[0173] Where q(s, a) is the Q-value of taking action a in state s, and f(s) k a) is the function value of the k-th element in the state vector related to action a, w k It is the attention weight of the k-th element, calculated as follows:

[0174]

[0175] Where, v i These are learnable parameters. Through attention mechanisms, agents can focus more on information that significantly impacts resource allocation decisions, thereby improving the accuracy and efficiency of resource allocation.

[0176] In the S400, S420 of AI-assisted resource optimization allocation, and global optimization combined with simulated annealing algorithm, although reinforcement learning can optimize resource allocation to a certain extent, in order to avoid getting trapped in local optima, the resource allocation scheme is globally optimized by combining simulated annealing algorithm.

[0177] In simulated annealing, an initial resource allocation scheme x0 is first randomly generated, and its objective function value E(x0) is calculated. The objective function can be the negative of the reward function r. Then, at the current temperature T, the current scheme is randomly perturbed to generate a new scheme x1, and the objective function value E(x1) of the new scheme is calculated. The probability P of accepting the new scheme is determined by the Metropolis criterion.

[0178]

[0179] As the algorithm progresses, the temperature T is gradually decreased, as shown by the formula T = T0·α. k Where T0 is the initial temperature, α is the cooling coefficient (0 < α < 1), and k is the number of iterations. In each iteration, the decision to accept a new solution is based on the acceptance probability. When the temperature drops to a certain level, the algorithm stops, and the solution obtained at this point is the globally optimized resource allocation scheme.

[0180] In S400, S430 (AI-assisted resource optimization allocation), and the dynamic resource allocation adjustment based on machine learning prediction, machine learning algorithms are used to predict the future demand and status of data center equipment, so as to adjust the resource allocation strategy in advance and achieve dynamic resource allocation. A Long Short-Term Memory (LSTM) network is used to perform time-series prediction of equipment operating status and business needs. The core structure of LSTM includes an input gate i t Forgotten Gate f t Output gate o t and memory unit C t At time t, the calculation formula is as follows:

[0181] i t =σ(W ii x t +b ii +W h ih t-1 +b hi )

[0182] f i =σ(W if x t +b if +W hf h t-1 +b hf )

[0183] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0184] C t =f t ·C t-1 +i t ·tanh(W ic x t +b ic +W hc h t-1 +b hc)

[0185] h t =o t ·tanh(C t )

[0186] Where, x t It is the input at time t, h t-1 σ is the hidden state at the previous time step, W and b are the weight matrix and bias vector, and σ is the sigmoid function.

[0187] Based on the forecast results, resource allocation strategies are adjusted in advance. If it is predicted that a certain device will experience excessive load in the future, some tasks will be allocated to other idle devices in advance; if it is predicted that the workload of a certain business task will increase significantly, sufficient equipment resources will be prepared in advance. Through this prediction-based dynamic resource allocation adjustment, resource allocation becomes more flexible and intelligent, better adapting to the dynamic changes of data center equipment and business. This echoes the control strategies in S300 for changes in equipment status, further optimizing the overall configuration of data center resources.

[0188] In a preferred embodiment of the present invention, after completing the AI-assisted resource optimization allocation in S400, it is necessary to conduct a comprehensive performance evaluation of the entire data center equipment management system and make feedback adjustments based on the evaluation results. This step S500, system performance evaluation and feedback adjustment, includes the following steps:

[0189] S510. Construct a multi-dimensional system performance evaluation index system;

[0190] S520, Construction of Comprehensive Performance Evaluation Model;

[0191] S530, Adjustment strategies based on trend analysis;

[0192] In the S510 section, a multi-dimensional system performance evaluation index system is constructed, which is based on three key dimensions: equipment operating status, resource utilization efficiency, and business task completion.

[0193] Equipment operating status dimension: A comprehensive equipment health index is introduced to measure the overall operating status of the equipment. This index comprehensively considers factors such as the current status of the equipment (e.g., temperature, power, etc.), historical failure counts, maintenance records, and the equipment status category assessed in S200. The weight of each factor is determined using the analytic hierarchy process (AHP). Let the set of factors affecting equipment health be F = {f1, f2, ..., f...} n}, construct the judgment matrix A(a ij ), where a ij This indicates the importance of factor i relative to factor j. The eigenvector W = [w1, w2, ..., wj] of the judgment matrix is ​​calculated.n The weights of each factor are obtained, satisfying the following conditions: The formula for calculating the overall health index of equipment is: Among them, f i，min and f i,max These are the minimum and maximum values ​​of factor i, respectively. This normalization process makes different factors comparable.

[0194] Resource utilization efficiency dimension: This dimension defines resource utilization balance to assess the degree of equal distribution of equipment resources across different tasks and time periods. The coefficient of variation (CV) is calculated for the resource utilization of each piece of equipment using the coefficient of variation method. i Equipment resource utilization rate U i The coefficient of variation can be calculated by the ratio of the actual equipment resources used in the task to the total equipment resources. Where, σU i Equipment resource utilization rate U i standard deviation It is their average value. The formula for calculating the resource utilization equilibrium degree is: The closer the RUBD value is to 1, the more balanced the resource utilization; the closer it is to 0, the more uneven the resource allocation.

[0195] Business Task Completion Dimension: Business Task Satisfaction (BTS) is used to measure the level of satisfaction with resource allocation and equipment operation for business tasks. This indicator comprehensively considers factors such as task completion time, completion quality, and task urgency. It is calculated using the fuzzy comprehensive evaluation method, with the set of evaluation factors denoted as U = {u1, u2, ..., u...}. s (e.g., completion time, completion quality, urgency, etc.), the evaluation level set is V = {v1, v2, ..., v...} t} (e.g., very satisfied, satisfied, neutral, dissatisfied, very dissatisfied), construct a fuzzy relation matrix R = (r ij ), where r ij Indicates factor u i Evaluation level v j The membership degree. Then, the weight vector A = [a1, a2, ..., a] of each evaluation factor is determined using the analytic hierarchy process. s The fuzzy comprehensive evaluation result of business task satisfaction is: BTS = A·R, and the result is a 1×t vector. The level of business task satisfaction is determined by the principle of maximum membership.

[0196] In the construction of the S520 comprehensive performance evaluation model, a linear weighted comprehensive model based on the entropy weight method is adopted to obtain the comprehensive performance evaluation results of the system. The entropy weight method is used to determine the objective weights of each dimension's evaluation indicators, avoiding the influence of subjective factors. First, the information entropy E of each indicator is calculated. j For index j, its information entropy is calculated using the following formula: Where n is the number of samples, x ij This represents the value of the i-th sample on index j. Then, the entropy weight w for each index is calculated. j :

[0197] Finally, the formula for calculating the overall system performance evaluation index is as follows: Among them, I1, I2, and I3 are the Comprehensive Health Index of Equipment (DCHI) in the dimension of equipment operation status, the Resource Utilization Efficiency Balance (RUBD) in the dimension of resource utilization efficiency, and the Business Task Satisfaction Index (BTS) in the dimension of business task completion, respectively.

[0198] Feedback Adjustment Mechanism: Based on the results of the System Comprehensive Performance Index (SCPI), formulate corresponding feedback adjustment strategies.

[0199] S530, threshold-based adjustment strategy: Set different levels of SCP1 thresholds, such as SCPI high SCPI medium SCPI low When SCPI ≥ SCPI high When the system is running well, the frequency of resource allocation adjustments can be appropriately reduced, but continuous monitoring is still necessary; when SCPI... medim ≤SCPI≤SCPI high At that time, make minor adjustments to the system.

[0200] Through the aforementioned system performance evaluation and anti-coupled adjustment mechanisms, problems existing in the system operation can be identified in a timely manner, and resource allocation strategies and equipment control strategies can be optimized, enabling the entire data center equipment management system to continuously adapt to the changing environment.

[0201] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-assisted adaptive control and management method for computer room equipment, comprising the following steps: S100, Data Acquisition and Preprocessing; S200, Equipment Condition Assessment and Classification; S300, regulation and control strategy formulation; S400, AI-assisted resource optimization and allocation; S500, system performance evaluation and feedback adjustment.

2. The adaptive control and management method for computer room equipment according to claim 1, characterized in that, In step S100, data acquisition and preprocessing, core operating data such as power, temperature, and usage time are collected for different types of equipment to construct an equipment status dataset D = {(x i y i )|i=1,2,…,|D|}. Where, Represents various operating parameters of the equipment, y i This corresponds to the device identifier or status label. In the data acquisition phase, let the measurement error of the j-th sensor at time t be . in The average of the measurements from multiple sensors, i.e. Then the adaptive weight w j The formula for calculating (t) is: For temperature data T(t), a modified Discrete Wavelet Transform (DWT) is used for feature extraction and denoising. In this method, the wavelet basis function is improved, and the improved wavelet coefficient calculation formula is as follows: in, a m and b n These are the scaling and translation parameters in the original discrete wavelet transform. When collecting device usage time data, an autoregressive moving average (ARMA) model is used to further process the generated data. In this method, data generated by a generative adversarial network is incorporated into the ARMA model, resulting in the following improved model: Where w is the weighting coefficient, used to adjust the degree of influence of the generated data in the model.

3. The adaptive control and management method for computer room equipment according to claim 1, characterized in that, Step S200, equipment status assessment and classification, includes the following steps: S210, Feature extraction optimization based on deep belief network (DBN); S220, Equipment Status Assessment Incorporating Improved Grey Relational Analysis (GRA); S230, Device status classification based on density peak clustering (DPC); In steps S200, S210 of equipment status assessment and classification, and feature extraction optimization based on Deep Belief Network (DBN), a Deep Belief Network (DBN) model is constructed using the equipment operation data processed in S100. Hidden layer neurons h j activation probability P(h) j =1|x) is calculated using the following formula: P(h j =1|x)=σ(b j +∑ i w ij x i ) in, It is the sigmoid activation function. Based on traditional DBN, an adaptive learning rate η(t) is introduced to accelerate model convergence and improve feature extraction performance. The formula is: Where η0 is the initial learning rate and α is the adjustment parameter. The gradient is the gradient at the k-th training iteration. In the equipment status assessment described in S200, S220 of equipment status assessment and classification, and the improved grey relational analysis (GRA) fusion equipment status assessment, the improved correlation coefficient calculation formula is as follows: Where, ζ i (k) is the correlation coefficient between the i-th comparison sequence and the reference sequence at time k, and ρ∈[0,1] is the resolution coefficient. Grey relational degree r of the device i for: In steps S200, S230 of equipment condition assessment and classification, and equipment condition classification based on density peak clustering (DPC), an adaptive cutoff distance adjustment factor λ is introduced. The cutoff distance is dynamically adjusted according to the distribution of data points, using the following formula:

4. The adaptive control and management method for computer room equipment according to claim 1, characterized in that, Step S300, the formulation of the control strategy, includes the following steps: S310, Reinforcement learning control strategy for potentially faulty equipment; S320, Normal equipment resource allocation strategy based on genetic algorithm optimization; S330, Development of rapid response and repair strategies for faulty equipment; In S300, S310 of the control strategy formulation, and the reinforcement learning control strategy for potentially faulty equipment, the Q-value update formula is as follows: Q(s,a)←Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)] Where α is the learning rate and γ is the discount factor. In this method, a dual-network structure and a priority experience replay mechanism are introduced to improve algorithm performance and decision accuracy. The dual-network structure includes an evaluation network Q and a target network Q. target The target network parameters are updated periodically. A priority experience replay mechanism samples based on the importance of the experience samples, with importance determined by the TD error δ = r + γmax. a′ Q target (s′,a′)-Q(s,a) is used as a metric. The sampling probability P(i) of a sample is calculated using the following formula: Where β is a parameter that controls the sampling bias, 0 < β < 1. In S300, S320 (the regulation strategy formulation), and the normal equipment resource allocation strategy optimized by genetic algorithm, a real-time load balancing factor L for the equipment is introduced. i And the urgency factor E of the task j The fit is adjusted. The adjusted fit w′ ij The calculation formula is: Among them, L max λ1 and λ2 are the maximum values ​​of the load on all devices, and λ1 and λ2 are weighting coefficients. In the S300, S330 (for formulating the control strategy), and the rapid response and repair strategy for faulty equipment, a resource constraint factor R is introduced. If maintenance resources are limited, it may be necessary to wait for resources to become available before repairs can proceed. Let T0 be the repair time when maintenance resources are sufficient. The formula for calculating the actual repair time T is: T = T0·(1+λ3·(1+R)), where λ3 is the weighting coefficient.

5. The adaptive control and management method for computer room equipment according to claim 1, characterized in that, The S400 and AI-assisted resource optimization allocation includes the following steps: S410. Construct a resource allocation model based on reinforcement learning; S420, global optimization combined with simulated annealing algorithm; S430, Dynamic resource allocation adjustment based on machine learning prediction; In S400, S410 of AI-assisted resource optimization allocation, and the construction of a resource allocation model based on reinforcement learning, the optimal resource allocation strategy is learned by interacting with the environment (the system consisting of data center equipment and business needs). Define the state space as s = [e1, e2, ..., e] n ,t1,t2,...,t m a1, a2, ..., a n×m ], where e i t represents the status information of the i-th device. j This represents the requirement information for the j-th task, a. ij This indicates the current allocation relationship between device i and task j (the value is 0 or 1, indicating whether allocation has occurred). Define the action space: Action 'a' represents the decision to allocate a specific device to a specific task. For n devices and m tasks, the action space consists of n×m possible allocation decisions. The agent selects an action each time to adjust resource allocation. Define the reward function: The reward function is designed as follows: Among them, t j，completed Indicates whether task j is completed (1 for completed, 0 for incomplete), w j The weight of task j reflects its importance. i,stable u represents the operational stability of device i (higher stability values ​​are larger). i It is the utilization rate of device i; a ij This represents the allocation relationship between device i and task j, c ij α is the cost of assigning device i to task j (e.g., device setup cost, transmission delay, etc.). α, β, and γ are weighting coefficients used to balance the impact of different factors on the reward. The weights of each element in the state vector are calculated using an attention mechanism, as shown in the following formula: Where q(s, a) is the Q-value of taking action a in state s, and f(s) k a) is the function value of the k-th element in the state vector related to action a, w k It is the attention weight of the k-th element, calculated as follows: Among them, v i These are learnable parameters. In S400, S420 of AI-assisted resource optimization allocation, and global optimization combined with simulated annealing algorithm, the probability P of accepting the new solution is determined by the Metropolis criterion: As the algorithm progresses, the temperature T is gradually decreased, as shown by the formula T = T0·α. k , where T0 is the initial temperature, α is the cooling coefficient (0 < α < 1), and k is the number of iterations. In S400, S430 of AI-assisted resource optimization allocation, and dynamic resource allocation adjustment based on machine learning prediction, a Long Short-Term Memory (LSTM) network is used to perform time-series prediction of the device's operating status and business needs. The core structure of LSTM includes an input gate i t Forgotten Gate f t Output gate o t and memory unit C t At time t, the calculation formula is as follows: i t =σ(W ii x t +b ii +W hi h t-1 +b hi ) f t =σ(W if x t +b if +W hf h t-1 +b hf ) o t =σ(W io x t +b io +W ho h t-1 +b ho ) C t =f t ·C t-1 +i t ·tanh(W ic x t +b ic +W hc h t-1 +b hc ) h t = no t fish (C) t ) Where, x t It is the input at time t, h t-1 σ is the hidden state at the previous time step, W and b are the weight matrix and bias vector, and σ is the sigmoid function.

6. The adaptive control and management method for computer room equipment according to claim 1, characterized in that, The S500 system performance evaluation and feedback adjustment includes the following steps: S510. Construct a multi-dimensional system performance evaluation index system; S520, Construction of Comprehensive Performance Evaluation Model; S530, Threshold-based adjustment strategy; In the S510 section, a multi-dimensional system performance evaluation index system is constructed, which is based on three key dimensions: equipment operating status, resource utilization efficiency, and business task completion. Equipment operating status dimension: A comprehensive equipment health index is introduced to measure the overall operating status of the equipment. Let the set of factors affecting equipment health be F = {f1, f2, ..., f...} n }, construct the judgment matrix A(a ij ), where a ij This indicates the importance of factor i relative to factor j. The eigenvector W = [w1, w2, ..., wj] of the judgment matrix is ​​calculated. n The weights of each factor are obtained, satisfying the following conditions: The formula for calculating the overall health index of equipment is as follows: Among them, f i，min and f i,max These are the minimum and maximum values ​​of factor i, respectively. This normalization process makes different factors comparable. Resource utilization efficiency dimension: The coefficient of variation (CV) of resource utilization rate for each piece of equipment is calculated using the coefficient of variation method. i Equipment resource utilization rate U i The coefficient of variation can be calculated by the ratio of the actual equipment resources used in the task to the total equipment resources. The formula for the coefficient of variation is: Where, σU i Equipment resource utilization rate U i standard deviation It is their average value. The formula for calculating the resource utilization equilibrium degree is: The closer the RUBD value is to 1, the more balanced the resource utilization; the closer it is to 0, the more uneven the resource allocation. Business task completion dimension: calculated using fuzzy comprehensive evaluation method, assuming the set of evaluation factors is U = {u1, u2, ..., u...} s The evaluation level set is V = {v1, v2, ..., v} (completion time, completion quality, urgency, etc.). t }(Very satisfied, satisfied, neutral, dissatisfied, very dissatisfied), construct a fuzzy relation matrix R = (r ij ), where r ij Indicates factor u i Evaluation level v j The membership degree. Then, the weight vector A = [a1, a2, ..., a] of each evaluation factor is determined using the analytic hierarchy process. s The fuzzy comprehensive evaluation result of business task satisfaction is: BTS = A·R, and the result is a 1×t vector. In the construction of the S520 comprehensive performance evaluation model, the information entropy E of each indicator is first calculated. j For index j, its information entropy is calculated using the following formula: Where n is the number of samples, x ij This represents the value of the i-th sample on index j. Then, the entropy weight w for each index is calculated. j : Finally, the formula for calculating the overall system performance evaluation index is as follows: Among them, I1, I2, and I3 are the Comprehensive Health Index (DCHI) of the equipment operation status dimension, the Resource Utilization Equilibrium (RUBD) of the resource utilization efficiency dimension, and the Business Task Satisfaction (BTS) of the business task completion dimension, respectively. S530, threshold-based adjustment strategy: Set different levels of SCPI thresholds, such as SCPI high SCPI medium， SCP low When SCPI ≥ SCPI high When the system is running well, the frequency of resource allocation adjustments can be appropriately reduced, but continuous monitoring is still necessary; when SCPI... medium ≤SCPI≤SCPI high At that time, make minor adjustments to the system.

Images