Intelligent operation and maintenance resource allocation method and device

Abstract

The invention discloses an intelligent operation and maintenance resource allocation method and device. According to the method, full-link multi-source data is collected and pre-processed to form multi-dimensional pre-processing features, the system state can be comprehensively described, a mixed deep learning model is called to capture a nonlinear relation and time sequence dependence, the prediction precision is improved, an interpretable result containing a confidence interval and a feature contribution degree is generated, and the prediction accuracy is improved. The black box characteristic of deep learning is broken; a resource scheduling strategy is generated by combining a dynamic threshold value with a multi-target optimization algorithm, adaptive adjustment can be performed according to a real-time load, multiple targets such as resource cost and service quality are balanced at the same time, and the defects of a traditional fixed threshold value and single target optimization are avoided; besides, according to the method, a prediction result is directly converted into a resource scheduling strategy, dynamic allocation is executed, a prediction-decision-execution closed loop is formed, reaction delay caused by separation of a prediction system and an allocation system in the prior art is reduced, the resource utilization rate is increased, and the method can better adapt to dynamic changes of services.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of server management, and in particular to an intelligent operation and maintenance resource allocation method and device. BACKGROUND

[0002] With the dynamic development of business and the continuous change of system load, reasonable prediction of server resource usage and effective allocation can ensure stable system performance, avoid resource waste and reduce operating costs.

[0003] In the prior art, capacity prediction methods based on machine learning have achieved certain applications. For example, traditional machine learning algorithms such as random forest and support vector machine (SVM) are used to model multi-dimensional historical data, which can realize resource utilization rate prediction; time series prediction models constructed by means of recurrent neural networks such as LSTM can capture periodic and trend changes in load; and multi-factor fusion prediction combining business data, system state and external factors (such as promotional activities and holidays) further improves the comprehensiveness of the prediction.

[0004] However, the prior art still has obvious deficiencies. Single-dimensional analysis cannot depict the complex correlation of multiple resource indicators, the "black box" characteristics of deep learning models result in poor interpretability, the processing capacity for abnormal situations such as sudden traffic is limited, and the model lacks self-adaptation ability, requiring frequent manual intervention. In addition, the fragmentation of the prediction and execution systems leads to reaction delay, and single-target optimization cannot balance multiple goals such as business demand and cost control.

[0005] Therefore, how to realize multi-dimensional fusion prediction and dynamic allocation is a problem that those skilled in the art urgently need to solve. SUMMARY

[0006] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide an intelligent operation and maintenance resource allocation method, device, server, computer readable storage medium and computer program product, which can realize multi-dimensional fusion prediction and dynamic allocation.

[0007] In a first aspect, an embodiment of the present application provides an intelligent operation and maintenance resource allocation method, comprising:

[0008] Collecting multi-source data of a full-link business scenario and performing data preprocessing to obtain multi-dimensional preprocessed features;

[0009] Calling a hybrid deep learning model to perform multi-dimensional resource future usage prediction according to the multi-dimensional preprocessed features, and generating a prediction result containing a confidence interval and interpretability information;

[0010] Based on the prediction result, generating a resource scheduling strategy through a multi-objective optimization algorithm combined with a dynamic threshold;

[0011] According to the resource scheduling strategy, resources are dynamically allocated.

[0012] In an embodiment, the calling hybrid deep learning model performs multi-dimensional resource future usage prediction according to the multi-dimensional preprocessed features, including:

[0013] The feature extraction network calling the multi-branch architecture separately extracts features of type-matched features through each branch and then fuses them to obtain cross-modal fusion features;

[0014] The cross-modal fusion features are input into a time series modeling network for time series dependency modeling to obtain a time series feature sequence;

[0015] The time series feature sequence is processed by an attention mechanism network to calculate feature weights and generate an attention weight matrix;

[0016] The calling hybrid prediction network generates a plurality of groups of prediction sample distributions based on the attention weight matrix and the time series feature sequence, combined with the feature contribution degree quantified based on the attention weight matrix, including the prediction results of the confidence interval and the interpretability information.

[0017] In an embodiment, the feature extraction network calling the multi-branch architecture separately extracts features of type-matched features through each branch and then fuses them, including:

[0018] The multi-dimensional preprocessed features are split by data type into multi-modal data including time series data, system index data, event data, and business data;

[0019] The multi-modal data is processed according to the corresponding feature processing procedure according to the data type to obtain time series features, system features, event features, and business features;

[0020] The time series features, system features, event features, and business features are fused to generate cross-modal fusion features.

[0021] In an embodiment, the multi-modal data is processed according to the corresponding feature processing procedure according to the data type, including:

[0022] The time series data is processed by a multi-layer 1D-CNN to extract multi-scale features to obtain time series features;

[0023] The system index data is processed by a fully connected network for feature conversion and dimension reduction to obtain system features;

[0024] The event data is processed by an Embedding layer to convert discrete events into dense vector representations to obtain event features;

[0025] The Transformer layer is called for the business data to perform internal correlation feature extraction through a self-attention mechanism to obtain business features.

[0026] In an embodiment, the calling mixed prediction network generates a prediction result containing confidence intervals and interpretability information based on the attention weight matrix and the time series feature sequence, a plurality of groups of prediction sample distributions generated by random inactivation, and feature contributions quantified based on the attention weight matrix, including:

[0027] The calling gating network generates fusion weights based on the importance scores of the cross-modal features.

[0028] According to the fusion weights, the attention weight matrix and the time series feature sequence are dynamically weighted and fused to obtain mixed features.

[0029] The mixed features are nonlinearly transformed and dimensionally mapped by a multilayer perceptron to generate a prediction result of multi-dimensional resources.

[0030] In an embodiment, the method for generating the interpretability information includes:

[0031] The SHAP library is called to calculate the global contribution of each type of feature in the time series feature sequence to the prediction result.

[0032] Key information located by the mixed prediction network in the generation of the prediction result is extracted.

[0033] After gradient perturbation of the data corresponding to the key information, the input value is input into the mixed prediction model for feature sensitivity analysis to obtain a sensitivity curve.

[0034] The global contribution value, the key information, and the sensitivity curve are packaged as a structured information package as the interpretability information.

[0035] In an embodiment, based on the prediction result, a resource scheduling strategy is generated by a multi-objective optimization algorithm combined with a dynamic threshold, including:

[0036] Usage trend and fluctuation range features are extracted from the prediction result to construct an initial vector containing resource types and time granularities.

[0037] The initial vector is input into a multi-objective optimization model to solve by a multi-objective optimization algorithm to generate resource scheduling strategy data containing resource allocation priorities and scheduling opportunities; wherein the multi-objective optimization model is based on prediction features to set resource cost and service quality optimization targets, and is constructed and generated in combination with historical scheduling rules.

[0038] In an embodiment, according to the resource scheduling strategy, resource dynamic allocation is performed, including:

[0039] The resource scheduling strategy is automatically disassembled by using a strategy analysis engine to generate a resource allocation instruction set with priorities;

[0040] Based on the intelligent resource scheduling middleware, the container orchestration interface is called according to the priority of the resource allocation instruction set to perform elastic allocation of computing resources and dynamic mounting of storage resources.

[0041] In an embodiment, the intelligent operation and maintenance resource allocation method further comprises:

[0042] Real-time monitoring of system operation indicators and running states to generate system monitoring data;

[0043] Effect evaluation according to the changes in the system monitoring data before and after the allocation to generate an evaluation result;

[0044] According to the evaluation result, the parameters of the hybrid deep learning model and the dynamic threshold are fed back and optimized.

[0045] In a second aspect, the embodiments of the present application provide an intelligent operation and maintenance resource allocation device, comprising:

[0046] A data acquisition and preprocessing module is configured to acquire multi-source data of a full-link business scenario and perform data preprocessing to obtain multi-dimensional preprocessing features;

[0047] A deep learning prediction module is configured to call a hybrid deep learning model to perform multi-dimensional resource future usage prediction according to the multi-dimensional preprocessing features and generate a prediction result containing a confidence interval and interpretability information;

[0048] An intelligent decision-making module is configured to generate a resource scheduling strategy based on the prediction result by using a multi-objective optimization algorithm combined with a dynamic threshold;

[0049] A dynamic resource allocation execution module is configured to perform dynamic resource allocation according to the resource scheduling strategy.

[0050] The intelligent operation and maintenance resource allocation method provided by the application can comprehensively depict system states by collecting full-link multi-source data and preprocessing to form multi-dimensional preprocessing features, solve the limitations of single-dimensional analysis of the prior art, capture nonlinear relationships and time sequence dependencies by calling a hybrid deep learning model, improve prediction accuracy, and generate interpretable results containing confidence intervals and feature contribution degrees, breaking the black box characteristics of deep learning; a resource scheduling strategy is generated by combining a dynamic threshold with a multi-objective optimization algorithm, which can adaptively adjust according to real-time load, balance multiple objectives such as resource cost and service quality, and avoid the shortcomings of traditional fixed thresholds and the defects of single-objective optimization; in addition, the method directly converts the prediction results into a resource scheduling strategy and performs dynamic allocation, forming a prediction-decision-execution closed loop, reducing the reaction delay caused by the separation of prediction and allocation systems in the prior art, improving resource utilization, and better adapting to business dynamics.

[0051] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0052] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments thereof, when read in conjunction with the accompanying drawings:

[0053] Figure 1 A flowchart of an intelligent operation and maintenance resource allocation method provided by an embodiment of the application is shown;

[0054] Figure 2 A feature processing and fusion diagram for multiple data modalities provided by an embodiment of the application is shown;

[0055] Figure 3 A whole flowchart of operation and maintenance resource allocation provided by an embodiment of the application is shown;

[0056] Figure 4 An exemplary structural block diagram of an intelligent operation and maintenance resource allocation device provided by an embodiment of the application is shown. DETAILED DESCRIPTION

[0057] The application will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related application, and not to limit the application. In addition, it should be noted that, for the sake of description, only the parts related to the application are shown in the drawings.

[0058] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments. Although the embodiments of the present application provide the method operation instruction steps as shown in the following embodiments or drawings, more or fewer operation instruction steps can be included in the method based on conventional or non-creative labor. The execution order of the steps is not limited to the execution order provided by the embodiments of the present application in the case of no necessary causal relationship in logic. The method can be executed in sequence or in parallel when the method is executed in actual processing or by the device, as shown in the embodiments or drawings.

[0059] Embodiment one:

[0060] The present embodiment proposes an intelligent operation and maintenance resource allocation method, please refer to Figure 1 , Figure 1 The present embodiment shows a flowchart of the intelligent operation and maintenance resource allocation method. As shown in Figure 1 , the method comprises:

[0061] S101, collect multi-source data of full-link business scenarios, and perform data preprocessing to obtain multi-dimensional preprocessed features;

[0062] Collect various types of data of each link in the business process, covering each link from the initiation, processing to the end of the business. Then, the data is processed by data cleaning (for example, using Z-Score method to identify outliers, threshold setting is 3σ), anomaly detection and processing, missing value filling (time series interpolation method, historical pattern filling method and machine learning prediction filling method can be used), etc. In the present embodiment, the specific means of preprocessing is not limited, valuable features are extracted from the original data, including statistical features (mean, variance, quantile, etc.), time series features (trend, periodicity, autocorrelation, etc.) and derived features (growth rate, resource utilization efficiency, etc.), forming a feature set containing multiple dimensions of information. In order to capture the time series characteristics, a sliding window technique can be used to extract features, typical window sizes include 5 minutes, 1 hour, 24 hours and 7 days, etc., which are not limited in the present embodiment.

[0063] Multi-source data refers to data including multiple types and multiple sources, such as system monitoring data (CPU, memory, network, storage usage, etc.), application layer data (request volume, response time, error rate, etc.), business data (transaction volume, user activity, etc.) and external environment data (holidays, marketing activities, etc.). When collecting multi-source data, a distributed collection architecture can be used, and each collection node collects raw data through a lightweight Agent; the data collection frequency can be dynamically adjusted according to the importance of the monitored object, for example, the minimum collection interval of the key indicators is 10 seconds.

[0064] Among them, the collection of full-link multi-source data can cover various links and influencing factors of business operation, avoid one-sidedness of single-dimensional data, and more accurately reflect the real state of the system; multi-dimensional preprocessed features can capture the complex correlation between different types of data, provide a richer information foundation for subsequent prediction and decision-making, and solve the limitations of single-dimensional analysis in the prior art.

[0065] S102, calling a hybrid deep learning model to perform multi-dimensional resource future usage prediction according to the multi-dimensional preprocessed features, and generating a prediction result containing a confidence interval and explainability information;

[0066] The hybrid deep learning model refers to a pre-trained resource prediction model composed of multiple hierarchical architectures. In this embodiment, the model structure and specific hierarchical parameter settings are not limited, and corresponding feature extraction and analysis levels can be configured according to different data types, which will not be described here.

[0067] The confidence interval refers to the prediction result reliability range calculated based on sample statistical characteristics by generating multiple sets of prediction sample distributions through random inactivation and other techniques, which is used to quantitatively show the uncertainty of the prediction. The confidence interval provides a quantitative reference for the uncertainty of the prediction for the operation and maintenance personnel, facilitating the development of flexible resource strategies. The operation and maintenance personnel can reserve redundant resources according to the upper limit of the confidence interval to avoid resource shortages or excessive configuration caused by relying solely on point prediction, which can solve the problem that the static threshold in the traditional method cannot adapt to load fluctuations. At the same time, the confidence interval can intuitively reflect the reliability of the prediction result. When the business load is at a high position in the interval, the system can trigger an expansion warning in advance, which can respond more timely to potential resource bottlenecks compared to the lag response of the fixed threshold in the traditional method.

[0068] The explainability information refers to quantifying the contribution of each feature to the prediction result based on the attention weight matrix. The explainability information breaks the black box nature of deep learning, allowing operation and maintenance personnel to understand the basis for the prediction, such as a promotion activity data contributing 30% of the memory usage prediction weight. The operation and maintenance personnel can verify the rationality of the prediction logic, avoid questioning the result due to unexplainable models, and solve the lack of transparency problem in existing machine learning methods. At the same time, when the prediction result deviates greatly from the actual value, the explainability information can be used to quickly locate the key features affecting the prediction, such as abnormal fluctuations in certain types of business data, so as to optimize data collection or model parameters accordingly and achieve precise tracing.

[0069] S103, generating a resource scheduling strategy based on the prediction result through a multi-objective optimization algorithm combined with a dynamic threshold;

[0070] Based on the multi-dimensional resource usage prediction results output by the hybrid deep learning model, an algorithm capable of simultaneously optimizing multiple targets is used, and a specific resource allocation and scheduling scheme is generated in combination with a threshold that can be dynamically adjusted according to business load.

[0071] The multi-target optimization algorithm takes multiple targets such as minimization of resource cost and optimization of service quality (such as response time compliance rate) as the optimization direction, and constructs a target function. In this embodiment, the optimal resource scheduling scheme that takes multiple targets into account is solved by the multi-target optimization algorithm based on the resource usage trend and fluctuation range in the prediction results, such as determining the CPU and memory allocation priority and specific values in different time periods. Compared with the single-target optimization method in the traditional method, the resource efficiency and service quality are simultaneously optimized by the multi-target algorithm in this embodiment, which avoids the problem of service degradation caused by excessive downsizing to reduce cost or resource waste caused by excessive expansion to ensure performance.

[0072] In addition, in this embodiment, the traditional fixed threshold is abandoned, such as triggering expansion when the CPU utilization rate is 80%, and the threshold is dynamically adjusted according to real-time prediction results and business scenarios, for example, temporarily increasing the memory usage expansion threshold from 75% to 85% during the promotion activity to adapt to the burst traffic. The dynamic threshold can be adjusted in real time according to the prediction results, which improves the accuracy of the scheduling strategy

[0073] It should be noted that the generation of the dynamic threshold in this embodiment needs to be combined with historical scheduling rules and current prediction features, such as automatically calculating a reasonable trigger threshold based on the predicted load fluctuation range. The specific dynamic threshold generation algorithm is not limited in this embodiment, and the corresponding algorithm can be selected according to the needs of the actual application scenario.

[0074] S104, dynamically allocate resources according to the resource scheduling strategy.

[0075] The resource scheduling strategy generated by the multi-target optimization algorithm in combination with the dynamic threshold is converted into specific resource allocation operations to realize the elastic adjustment of computing resources and storage resources. The supported resource types include but are not limited to virtual machines, containers, function calculation resources, physical servers, etc.; the allocation operation types include but are not limited to adding resources, releasing resources, resource configuration adjustment, load migration, etc.; and the dynamic allocation execution strategy includes but is not limited to batch execution, gray release, and automatic rollback in case of failure.

[0076] Compared with the reaction delay caused by the separation of the prediction system and the resource allocation system in the traditional method, the closed-loop linkage of prediction and execution in the embodiment realizes the full-process automation of prediction-strategy generation-dynamic allocation, without the need for manual participation in resource allocation decision-making and execution, and can handle multi-resource collaborative scheduling in complex scenarios, significantly improves the response speed, and can avoid the service interruption risk caused by manual intervention in the traditional method.

[0077] Before dynamic resource allocation, safety checks and risk assessments can be further performed. Specifically, the safety checks include but are not limited to verifying resource availability, confirming operation permissions, and assessing potential impacts on the business; the risk assessments include but are not limited to single-resource change amplitude limitation and additional approval processes for high-risk operations. Of course, the steps of safety checks and risk assessments can also not be performed, which are not limited in the embodiment.

[0078] After dynamic resource allocation, decision abnormalities and execution abnormalities can be further identified and solved. Decision abnormalities refer to abnormal states caused by the decision process or results of multi-objective optimization that violate system preset rules, business objectives or safety boundaries, resulting in unreasonable resource allocation, rising system risks or damaged business. Safety threshold constraints (such as setting upper and lower limits for resource adjustment to prevent excessive adjustment) can be used to detect decision conflicts and automatically trigger manual review processes or other conflict resolution methods. Execution abnormalities refer to abnormal states caused by hardware failures, network interruptions, parameter configuration errors, program vulnerabilities, resource conflicts, etc., which prevent the preset operation processes (such as resource allocation, strategy execution, model optimization, etc.) from being completed as expected or cause errors that do not match the expected results. Retest strategies (number of times, interval) can be intelligently set according to the error types, large-scale adjustments can be performed in batches, the effects can be verified after each batch, and the system can automatically roll back to the state before the adjustment when serious negative effects are detected.

[0079] Based on the above introduction, the intelligent operation and maintenance resource allocation method provided in the embodiment can collect full-link multi-source data and pre-process them to form multi-dimensional pre-processed features, which can comprehensively depict the system state, solve the limitations of single-dimensional analysis in the prior art, capture nonlinear relationships and time series dependencies by calling a hybrid deep learning model, improve prediction accuracy, and generate interpretable results including confidence intervals and feature contribution degrees, breaking the black box characteristics of deep learning. By using dynamic thresholds in combination with multi-objective optimization algorithms to generate resource scheduling strategies, the system can adaptively adjust according to real-time loads while balancing multiple objectives such as resource cost and service quality, avoiding the shortcomings of traditional fixed thresholds and the defects of single-objective optimization. In addition, the method directly converts the prediction results into resource scheduling strategies and performs dynamic allocation, forming a prediction-decision-execution closed loop, reducing the reaction delay caused by the separation of the prediction system and the resource allocation system in the prior art, improving resource utilization, and better adapting to business dynamics.

[0080] Embodiment Two:

[0081] In order to more systematically capture the correlation between data, improve the input quality of the prediction model and the reliability of the prediction result, an implementation step of a hybrid deep learning model for predicting future usage of multi-dimensional resources according to multi-dimensional preprocessed features is proposed in this embodiment.

[0082] Step S102 calls the hybrid deep learning model to predict the future usage of multi-dimensional resources according to multi-dimensional preprocessed features, which can specifically include the following steps:

[0083] Step S21, calling the feature extraction network of the multi-branch architecture to separately extract features through each branch for type-matched features, and then fusing them to obtain cross-modal fusion features;

[0084] The feature extraction network sets up independent branches according to data types (time series, system indicators, events, and business data), and each branch extracts features through a dedicated network (such as 1D-CNN, fully connected layer, and Transformer), and then concatenates them by dimension to form cross-modal fusion features. For example, the time series branch uses 1D-CNN to capture the multi-scale fluctuations of CPU usage (3-minute short-term and 7-day long-term), and the business branch uses Transformer to analyze the association features between large promotion activities and order volume.

[0085] Step S22, inputting the cross-modal fusion features into the time series modeling network to model the time series dependence and obtain the time series feature sequence;

[0086] The cross-modal fusion features are input into the time series modeling network (e.g., a 2-layer stacked structure containing a bidirectional LSTM network with 256 hidden units in each layer), which learns the temporal correlation of historical features through a gating mechanism and outputs a time series feature sequence containing long-term trends (such as resource demand rising due to business growth) and short-term fluctuations (such as daily traffic peaks).

[0087] Further, a Dropout layer (with a rate of 0.3) can be inserted between the LSTM layers to prevent overfitting.

[0088] Step S23, calculating the feature weight matrix of the time series feature sequence through the attention mechanism network to generate the attention weight matrix;

[0089] The time series feature sequence output by the LSTM is calculated through a multi-head self-attention mechanism to generate a feature weight matrix, focusing on time points or data patterns that are more important for prediction (such as resource peak moments during large promotions), and filtering redundant information. In one configuration, the number of attention heads can be set to 8, and the attention dimension can be set to 64.

[0090] Step S24, calling the hybrid prediction network to generate a prediction result based on the attention weight matrix and the time sequence feature sequence, the prediction sample distribution generated by random inactivation, and the feature contribution degree quantified based on the attention weight matrix, which contains confidence intervals and explainability information.

[0091] The LSTM output and the attention weight matrix are fused through a gating mechanism to form a hybrid feature. A hybrid prediction network is called to generate prediction information by predicting the future usage of multi-dimensional resources based on the hybrid feature. An MLP is used to perform nonlinear transformation on the hybrid feature, and multiple sets of prediction samples are generated by random inactivation through Dropout. Confidence intervals are calculated. The feature contribution degree (e.g., 30% of the impact of a big promotion event on the prediction result) is quantified based on the attention weight matrix to generate explainability information. The prediction information, confidence intervals, and explainability information are combined as the prediction result.

[0092] It should be noted that the confidence interval represents a quantitative evaluation of the uncertainty of the prediction result, i.e., under a certain probability, the true value falls within the interval. For example, a 90% confidence interval of [60%, 75%] indicates that there is a 90% probability that the CPU usage at a certain time in the future will be between 60% and 75%. Explainability information refers to the attribution analysis and logical explanation of the prediction result, which is used to answer why the model makes such a prediction. It mainly includes understandable information such as feature importance and model focus.

[0093] The prediction result generation method provided in this embodiment fuses time sequence, business, and event features through a multi-branch architecture, breaking the limitations of single-dimensional prediction and capturing complex correlations between resource indicators. At the same time, through the attention mechanism, key time points (such as sudden traffic moments) are automatically selected, and the time sequence dependence modeling of LSTM is combined to improve the prediction capability in abnormal load scenarios. In addition, confidence intervals and explainability information are generated to facilitate the trust and adjustment of operation and maintenance personnel, support operation and maintenance decisions, and solve the pain points of low prediction accuracy, poor explainability, and weak adaptability in the prior art.

[0094] Embodiment Three

[0095] The specific data type division method and the corresponding feature extraction method in the modal splitting of the multi-dimensional preprocessed features in the above embodiments are not limited, and the process of calling the feature extraction network of the multi-branch architecture to separately extract features of type-matched features through each branch and then fuse them can refer to the following steps:

[0096] Step S211, splitting the multi-dimensional preprocessed features into multi-modal data including time sequence data, system indicator data, event data, and business data according to data types;

[0097] The pre-processed multi-dimensional features are divided into four types of modalities according to the original data types:

[0098] Time series data: continuous indicators such as CPU / memory usage, network traffic, etc. that change over time;

[0099] System indicator data: performance counters, alarm logs, and other system state data;

[0100] Event data: discrete events such as system upgrades, promotional activities, etc.

[0101] Business data: transaction volume, user activity, and other business behavior indicators.

[0102] Step S212, according to the data type, call the corresponding feature processing flow to process the multi-modal data, and get the time series feature, system feature, event feature and business feature;

[0103] Step S213, fusion time series feature, system feature, event feature and business feature, generate cross-modal fusion feature.

[0104] The four types of modality features are concatenated by dimension (such as time series feature matrix + system feature vector + event Embedding + business feature tensor), forming a comprehensive feature vector containing time regularity + system state + event influence + business driving". By fusing the complementary information of different modalities, a more complete resource demand driving model is constructed (such as considering the influence of CPU historical trend, system upgrade event, and large promotion business volume on memory usage).

[0105] The feature extraction and fusion method provided in this embodiment can capture the periodic information of time series data and the nonlinear influence of business data through modality processing. At the same time, through the pipeline of type splitting, exclusive processing and cross-modal fusion, a feature system covering four-dimensional driving factors of time, system, event and business is constructed, which adapts to the physical characteristics and business meanings of different data modalities, avoids one-size-fits-all feature extraction, fuses multi-dimensional correlation information, captures the complex driving mechanism of resource usage, and provides a clear feature source tracing path for subsequent attribution analysis.

[0106] Embodiment four:

[0107] The above embodiment does not limit the specific feature extraction process of each type corresponding to step S212 of processing multi-modal data according to the data type to call the corresponding feature processing flow. In this embodiment, a step is proposed according to the characteristics of different data modalities, which uses exclusive algorithms to extract features, Figure 2 As shown in a feature processing and fusion diagram for multiple data modalities.

[0108] Step S212 can be performed according to the following substeps:

[0109] Step S2121, multi-scale features are extracted from the time series data by using a multi-layer 1D-CNN to obtain time series features;

[0110] The time series data (such as CPU usage and network traffic) are processed by using a multi-layer 1D convolutional neural network (such as a convolution kernel size of [3, 5, 7], and a filter number of [64, 128, 256]), and multi-granularity time features are captured by using convolution kernels of different sizes; a small convolution kernel (such as 3) can be used to capture short-term local fluctuations (such as CPU spikes within 5 minutes); and a large convolution kernel (such as 7) can be used to capture long-term trends or periodic patterns (such as weekly traffic rules).

[0111] The time series feature vectors containing multi-scale time patterns are extracted, and the time series data are decomposed into trend items, seasonal items and residual items by using a time domain decomposition technique.

[0112] Step S2122, the system index data are converted and reduced in dimension by using a fully connected network to obtain system features;

[0113] The multi-dimensional system index data (such as CPU, memory and disk utilization) are input into the fully connected network, and matrix transformation is realized.

[0114] The features are reduced in dimension to remove redundant features (such as the high correlation features of CPU and memory utilization);

[0115] The features are converted to map the original indexes to more representative comprehensive features (such as a “system load comprehensive index”).

[0116] Step S2123, discrete events are converted into dense vector representations by using an Embedding layer to obtain event features;

[0117] The discrete event data (such as “system upgrade” and “promotion activity”) are converted from sparse one-hot encoding into dense continuous vector representations by using the Embedding layer, so that semantically related events are closer in the vector space (such as the vectors of “big promotion activity” and “traffic surge” events are closer), and event feature vectors that preserve semantic associations are generated.

[0118] Step S2124, business data are called to extract internal association features by using a Transformer layer through a self-attention mechanism to obtain business features.

[0119] The self-attention mechanism of the Transformer layer is used to process business data (such as order volume, user activity, etc.), and by calculating the attention weight of each element in the sequence, the long-distance dependency relationship in the business process (such as the time sequence correlation of user ordering, payment, and logistics) is captured, and a business feature vector containing business logic correlation is extracted.

[0120] The feature processing procedure provided in this embodiment extracts a targeted algorithm for data feature extraction, which can improve feature quality and support accurate prediction in complex scenarios.

[0121] Embodiment five:

[0122] In one embodiment, the method for generating the prediction result in step S24 includes:

[0123] Step 241, calling a gating network to generate a fusion weight based on the importance score of the cross-modal feature;

[0124] Since fixed weight fusion cannot adapt to the scene switching of "daily load → burst traffic", relying on historical time sequence rules during a large promotion will cause prediction lag. In this embodiment, a gating network is called to generate a fusion weight between 0 and 1 based on the importance score (such as the dynamic weight of the time sequence feature and the business feature) of the cross-modal feature. The weight is used to measure the relative importance of the key features of attention screening and the basic features of time sequence modeling. For example, during a large promotion, the gating network detects that the importance of the business feature increases, and generates a fusion weight of 0.7 (attention feature proportion) and 0.3 (time sequence feature proportion), so that the model pays more attention to the resource peak related to the promotion.

[0125] By calculating the fusion weight in real time through the gating network, the model can focus on the time sequence features (80% proportion) to capture periodic patterns in stable scenarios (such as daily office work), and focus on attention features (90% proportion) to focus on event impact in burst scenarios (such as system upgrade), thereby dynamically adapting to different business scenarios.

[0126] Step 242, dynamically weighting and fusing the attention weight matrix and the time sequence feature sequence according to the fusion weight to obtain mixed features;

[0127] According to the fusion weight, the attention weight matrix (screened key features) and the time sequence feature sequence (basic trend features) are weighted and summed.

[0128] Step 243, performing nonlinear transformation and dimension mapping on the mixed features through a multilayer perceptron to generate a prediction result of multi-dimensional resources.

[0129] The nonlinear transformation capability of MLP can handle the coupled relationship between multiple resource indicators. In this step, MLP is called to transform the mixed features using a nonlinear activation function (such as ReLU), mining hidden associations between features (such as the nonlinear relationship between business volume and memory requirements), and mapping them to multi-dimensional resource indicators (such as future CPU, memory, and bandwidth usage) through the output layer.

[0130] The prediction result generation method provided in this embodiment adjusts the feature priority in real time through gating weights, which can significantly reduce the prediction error of burst traffic scenarios caused by dynamic weighted fusion, avoid information loss caused by one-size-fits-all fusion, and improve the average accuracy of multi-dimensional resource prediction.

[0131] Example 6:

[0132] Explanable information refers to the attribution analysis and logical explanation of the prediction results of the hybrid prediction network, which can be measured based on the feature contribution quantified by the attention weight matrix. In this embodiment, a method for generating highly reliable explainable information is proposed. Specifically, the method for generating explainable information in step S24 includes:

[0133] Step 244: Call the SHAP library to calculate the global contribution of each type of feature in the time series feature sequence to the prediction result;

[0134] The SHAP (SHapley Additive ExPlanations) library is called to calculate the global contribution of each input feature (such as time series features and business features) to the prediction result based on game theory. This step quantifies feature importance from a global perspective and addresses the black box problem of deep learning. For example, a SHAP value of 0.32 for a promotion feature indicates that this feature contributes 32% to the memory prediction result.

[0135] Step 245: extract key information of the hybrid prediction network in generating the prediction result;

[0136] Extract the weight matrix from the hybrid prediction network's attention mechanism to identify key time points or data patterns that the model focuses on when generating predictions. This step reveals the model's focus from a local perspective, providing direct evidence for the prediction results. For example, a heat map reveals that the model allocated over 80% of its attention weight to traffic features between 8:00 PM and 12:00 AM on Double 11.

[0137] Step 246: Perform gradient perturbation on the data corresponding to the key information and then input the hybrid prediction model to perform feature sensitivity analysis to obtain a sensitivity curve;

[0138] Gradient perturbation is performed on key features (such as promotion intensity) (such as ±10%, ±20%), the perturbed value is input into the mixed prediction model, the output change is observed, and a sensitivity curve is generated. For example, when the promotion intensity is increased from 50% to 70%, the memory prediction value is increased by 15%, and the curve shows a nonlinear growth. The correlation between the feature and the prediction result is verified by causal reasoning, and the influence of the feature change on the prediction is determined, thereby enhancing the credibility of the explanation.

[0139] Step 247, encapsulating the global contribution value, key information, and sensitivity curve as a structured information package as an interpretable information.

[0140] The SHAP value ranking table, attention key information (such as a heat map), sensitivity curve, etc. are integrated into a structured information package in JSON or XML format, which is convenient for front-end display or downstream system calling.

[0141] The interpretable information generation method provided in the embodiment combines global SHAP values and local attention key information to form a complete explanation chain of what is important global contribution + where it is important, and at the same time, the output change is observed by perturbing the feature, and the causal relationship can be verified. For example, after perturbing the promotion intensity feature, the memory prediction value fluctuates as expected, proving that there is a real causal relationship between the two, rather than a casual correlation, and the credibility of the explanation is enhanced through causal verification.

[0142] Embodiment seven:

[0143] The passive scheduling based on historical data or threshold triggering in the traditional method often leads to scheduling lag, and the method of early active planning and scheduling proposed in the embodiment can balance resource efficiency and service quality.

[0144] The process of generating a resource scheduling strategy based on the prediction result through a multi-objective optimization algorithm combined with a dynamic threshold in step S103 can be performed according to the following steps:

[0145] Step S31, extracting usage trend and fluctuation range features from the prediction result, and constructing an initial vector containing resource types and time granularity;

[0146] The prediction result is converted into a structured scheduling input, and the trend direction and uncertainty range of resource demand are determined, providing a quantitative basis for subsequent optimization. Two types of key information are extracted from the prediction result:

[0147] Usage trend: such as the growth slope of CPU usage rate in the next 24 hours, and the periodic fluctuation rule of memory demand;

[0148] Fluctuation range feature: obtain the upper and lower bounds of resource usage through the confidence interval (such as a 90% confidence interval of [60%, 75%]).

[0149] The above information is organized into an initial vector by resource type (CPU / memory / bandwidth) and time granularity (15 minutes / hour), for example: [CPU trend: +5% per hour, memory fluctuation range: [60%, 75%], bandwidth peak: 1.2 Gbps].

[0150] Step S32, input the initial vector into the target optimization model to solve by multi-objective optimization algorithm, to generate resource scheduling strategy data containing resource allocation priority and scheduling timing;

[0151] The initial vector is input into the multi-objective optimization algorithm (such as NSGA-II), and the weights of the three objectives of performance, cost and stability are dynamically adjusted according to the business stage, priority and historical feedback, to generate a scheduling strategy containing the following elements:

[0152] Resource allocation priority: determine the allocation order of CPU / memory / bandwidth (such as high priority of memory over CPU during big promotion);

[0153] Scheduling timing: calculate the optimal expansion / contraction time point (such as scheduling 1 hour in advance when predicting that CPU will exceed the threshold after 3 hours).

[0154] Among them, the multi-objective optimization model sets resource cost and service quality optimization targets based on prediction features, and is constructed and generated in combination with historical scheduling rules. The construction of the model can set double targets, resource cost optimization and service quality (QoS) optimization. Resource cost and service quality can be quantified according to state space (current load state of the system, business priority, historical decision effect, etc.) and action space (weight adjustment direction and amplitude, balance target achievement and system stability) state parameters.

[0155] Among them, resource cost optimization refers to minimizing idle resource occupation, which can avoid cost waste caused by over-allocation; service quality (QoS) optimization refers to ensuring that resource allocation meets business SLA, such as request response time < 50ms.

[0156] The optimization algorithm can adopt an improved multi-objective genetic algorithm (NSGA-II) to solve a Pareto optimal solution set. Specifically, an optimization algorithm is as follows: the population size is set to 100, and the maximum number of iterations is 200; the coding scheme adopts real number coding, each chromosome represents a set of resource configuration schemes, and the tournament selection is based on the Pareto level and the crowding distance; the simulated binary crossover (Simulated Binary Crossover) is selected for the crossover operation, and the crossover probability is 0.9; the polynomial mutation (PolynomialMutation) is selected for the mutation operation, and the mutation probability is 0.1; the optimal solution is reserved for the next generation to ensure the convergence of the algorithm, the historical data and the current system state are automatically analyzed, the thresholds of various types are dynamically calculated and optimized, the time-weighted moving average method is used to calculate the dynamic baseline, the Bayesian optimization method is used to find the optimal threshold parameter based on the historical alarm accuracy data.

[0157] The method for generating a resource scheduling strategy provided in this embodiment automatically adjusts the threshold according to the predicted fluctuation range, the dynamic threshold can adapt to changes in the business scenario, and the resource efficiency and the service quality are balanced through multi-objective weight adjustment, so as to meet the dynamic balance requirements of different scenarios.

[0158] Embodiment Eight

[0159] After obtaining the resource scheduling strategy, in order to ensure the automatic execution of resource dynamic allocation and the stability of key businesses, a resource dynamic allocation strategy is provided in this embodiment, and the process of resource dynamic allocation according to the resource scheduling strategy can be executed according to the following steps:

[0160] Step S41, the resource scheduling strategy is automatically disassembled by using a strategy analysis engine to generate a resource allocation instruction set with priorities;

[0161] The strategy analysis engine automatically disassembles the scheduling strategy generated by multi-objective optimization (such as preferentially expanding memory during a large promotion and allocating 85% of resources 2 hours in advance) into an executable instruction set, converts the abstract scheduling strategy into specific operation instructions that can be recognized by a container orchestration system, and sorts the operation instructions according to priorities. For example:

[0162] High-priority instruction: immediately allocate 200 GB of memory to the e-commerce business cluster;

[0163] Medium-priority instruction: expand the CPU core number to 400 cores after 30 minutes;

[0164] Low-priority instruction: release 100 GB of storage for non-critical services after the scheduling is completed.

[0165] Step S42, based on the intelligent resource scheduling middleware, the container orchestration interface is called according to the priority of the resource allocation instruction set to perform elastic allocation of computing resources and dynamic mounting of storage resources.

[0166] The middleware calls the container orchestration interface (such as Kubernetes API) according to the instruction set priority to implement:

[0167] Elastic allocation of computing resources: dynamically adjust the number of Pod replicas, CPU / memory limits through the scale interface (such as expanding the number of Web service replicas from 10 to 50);

[0168] Dynamic mounting of storage resources: temporarily mount NFS storage volumes for business containers through the volume attach interface (such as adding 1TB of storage for log services during a big promotion).

[0169] The container orchestration interface (such as Kubernetes) is a standard resource management tool for cloud-native environments, and unified calling through the middleware can avoid the complexity of directly operating the underlying infrastructure (such as physical machines).

[0170] The resource dynamic allocation method provided by the embodiment realizes the full-process automation of policy generation -> instruction disassembly -> resource allocation through automatic disassembly by the engine + automatic execution by the middleware, which can avoid human intervention errors; at the same time, the allocation instructions are sorted according to the importance of the business, ensuring that critical resources are allocated first and non-critical resources are executed later, ensuring the stability of critical business.

[0171] In order to further realize the self-evolution of the resource scheduling system, the following steps can be continued during the process of dynamically allocating resources according to the resource scheduling strategy:

[0172] Step S43, real-time monitoring of system running indicators and running states to generate system monitoring data;

[0173] Through monitoring tools such as Prometheus and Grafana, system running data is collected in real time, including resource indicators (CPU / memory usage, disk I / O, network traffic), business indicators (request response time, throughput, error rate), scheduling execution status (container creation / destroy time, resource allocation delay), etc. are stored as structured data in time series to generate system monitoring data.

[0174] Step S44, effect evaluation according to the change of system monitoring data before and after allocation, to generate evaluation results;

[0175] The scheduling effect can be evaluated from the effect evaluation indexes (resource utilization improvement, response time improvement, cost saving rate) and verification result classification (fully consistent with expectations, partially consistent with expectations, not consistent with expectations, and negative impact, etc.) to verify the actual effect of the scheduling strategy.

[0176] Step S45, feedback of the parameters of the optimized hybrid deep learning model and the dynamic threshold value according to the evaluation result.

[0177] Through the gradient descent algorithm, the hybrid deep learning model is fine-tuned with the evaluation result (such as prediction error) as the loss function, and the mapping relationship between the threshold value and the business scenario (such as large promotion scenario→ safety coefficient 1.8) is established. The mapping rules are automatically updated through historical evaluation data.

[0178] It should be noted that the above feedback optimization process can set a three-level feedback optimization mechanism of short cycle feedback (fast feedback of execution results to the generation model of the resource scheduling strategy, adjustment of current decision parameters, feedback period: 5-15 minutes, adjustment object: threshold value parameter, optimization target weight), medium cycle feedback (feedback of execution effect to the hybrid deep learning model, optimization of the prediction model, feedback period: 1-24 hours, adjustment object: model parameter, feature weight) and long cycle feedback (system overall efficiency feedback to the data acquisition end, optimization of data acquisition and feature engineering, feedback period: 1-7 days, adjustment object: data acquisition strategy, feature extraction method). The short cycle feedback adjusts the threshold value and the optimization target weight in real time to adapt to the sudden changes in resource scheduling and improve the decision response efficiency. The medium cycle feedback relies on the optimization of model parameters and feature weights to continuously polish the prediction accuracy of the hybrid deep learning model and strengthen the long-term prediction reliability. The long cycle feedback optimizes the acquisition strategy and feature engineering from the data source, systematically improves the data quality of the whole link, and enhances the dynamic adaptation ability of the system to the business scenario. The three-level feedback mechanism builds an intelligent closed loop from bottom to top through the hierarchical design of real-time response-model evolution-system reconstruction: the short cycle ensures the timeliness of decision-making, the medium cycle guarantees the accuracy of the model, and the long cycle consolidates the data foundation. The three work together to make the resource scheduling system agile, growing and robust, and ultimately achieve the goal of dynamically adapting to business needs, continuously optimizing scheduling efficiency, and systematically improving resource utilization.

[0179] As shown in Figure 3 is a whole flowchart of operation and maintenance resource allocation, the data source provides original data (such as server indicators, business logs), which is cleaned and formatted through data acquisition and preprocessing, becoming available data. Artificial experience data (such as additional resources needed for large promotion) can also be input through manual intervention interface. The preprocessed data is input into deep learning prediction to calculate future resource demand, and the prediction result is transmitted to dynamic threshold adjustment (adjusting alarm rules) and multi-objective decision optimization.

[0180] The optimized strategy data is input to resource allocation execution, directly operating the target resource pool, and the effect data after execution is fed back to effect evaluation feedback, which is used to correct the deep learning prediction and multi-objective decision-making.

[0181] From data collection, through prediction-decision-making-execution, and finally feedback optimization, a closed-loop driving intelligent operation and maintenance is formed.

[0182] The feedback optimization method provided in the embodiment can verify the effectiveness of the strategy through actual operation data, reduce scheduling risks, continuously optimize the model and threshold, automatically adapt to the dynamic changes of the business scenario, and reduce the operation and maintenance burden.

[0183] Embodiment nine:

[0184] Further reference Figure 4 It shows an exemplary structure block diagram of an intelligent operation and maintenance resource allocation device according to an embodiment of the application, which mainly includes a data collection and preprocessing module, a deep learning prediction module, an intelligent decision-making module, and a dynamic resource allocation execution module. The intelligent operation and maintenance resource allocation device adopts modular design and realizes multi-dimensional fusion prediction and dynamic allocation through four core units.

[0185] The data collection and preprocessing module is mainly used for collecting multi-source data of the full-link business scenario and performing data preprocessing to obtain multi-dimensional preprocessed features.

[0186] The deep learning prediction module is mainly used for calling a hybrid deep learning model to perform multi-dimensional resource future usage prediction according to the multi-dimensional preprocessed features, and generate a prediction result containing a confidence interval and explainability information.

[0187] The intelligent decision-making module is mainly used for generating a resource scheduling strategy based on the prediction result through a multi-objective optimization algorithm combined with a dynamic threshold.

[0188] The dynamic resource allocation execution module is mainly used for performing dynamic resource allocation according to the resource scheduling strategy.

[0189] In an embodiment, the intelligent operation and maintenance resource allocation device can further be configured with an intelligent threshold management module, which continuously optimizes the threshold through a continuous effect evaluation and feedback mechanism, and adjusts the threshold more accurately in combination with the business priority strategy and the deep learning prediction result.

[0190] In an embodiment, the intelligent operation and maintenance resource allocation device can further be configured with an effect feedback and verification module, which ensures the effectiveness of the resource allocation strategy and the continuous evolution ability of the system through systematic effect evaluation, feedback optimization, and closed-loop verification, and ensures continuous self-optimization of the system.

[0191] In an embodiment, the intelligent operation and maintenance resource deployment device can be further configured with a visual and interactive interface, supporting the visual display of prediction results, interpretable information, and monitoring information of resource deployment state, and can also support dynamic adjustment of policy parameters, manual intervention in abnormal scenarios, query and analysis of historical data, etc.

[0192] The flowcharts and block diagrams in the drawings illustrate the possible architectural, functional, and operational scenarios of systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment, or a portion of code that contains one or more executable instructions for implementing the specified functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in a different order than that shown in the figures. For example, two blocks connected in series can actually be executed in parallel, or they can be executed in reverse order, depending on the functionality involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0193] The units or modules described in the embodiments of the present application can be implemented in software or in hardware. The described units or modules can also be arranged in a processor. In some cases, the names of the units or modules do not constitute a limitation on the units or modules themselves.

[0194] The above description is merely preferred embodiments of the present application and a description of the principles of the technology used. Those skilled in the art should understand that the disclosed scope of the present application is not limited to the technical solutions formed by the specific combinations of the above technical features, and also covers other technical solutions formed by any combination of the above technical features or equivalent features without departing from the above disclosed concept. For example, the above features can be replaced with other technical features disclosed in the present application (but not limited to) having similar functions to form technical solutions.

Claims

1. A method for allocating intelligent operation and maintenance resources, characterized in that: include: Collect multi-source data from full-link business scenarios and perform data preprocessing to obtain multi-dimensional preprocessing features; Calling the hybrid deep learning model to predict the future usage of multi-dimensional resources based on the multi-dimensional preprocessing features, and generating a prediction result including confidence intervals and interpretability information; Based on the prediction results, a resource scheduling strategy is generated by combining a multi-objective optimization algorithm with a dynamic threshold; Dynamic resource allocation is performed according to the resource scheduling strategy.

2. The method according to claim 1, wherein The calling of the hybrid deep learning model to predict the future usage of the multi-dimensional resources according to the multi-dimensional preprocessing features includes: The feature extraction network of the multi-branch architecture is called to extract the features of the type matching separately through each branch and then fuse them to obtain cross-modal fusion features; Inputting the cross-modal fusion features into a temporal modeling network to perform temporal dependency modeling to obtain a temporal feature sequence; Perform feature weight calculation on the temporal feature sequence through the attention mechanism network to generate an attention weight matrix; The hybrid prediction network is called based on the attention weight matrix and the temporal feature sequence, combined with multiple groups of prediction sample distributions generated by random inactivation and feature contributions quantified based on the attention weight matrix, to generate prediction results including confidence intervals and interpretability information.

3. The method according to claim 2, wherein The feature extraction network calling the multi-branch architecture extracts the features of the type matching separately through each branch and then fuses them, including: Splitting the multi-dimensional preprocessing features into multimodal data including time series data, system indicator data, event data, and business data according to data type; Calling a corresponding feature processing flow to perform feature processing on the multimodal data according to the data type to obtain time series features, system features, event features, and business features; The time series features, the system features, the event features, and the business features are fused to generate cross-modal fusion features.

4. The method according to claim 3, wherein Calling a corresponding feature processing flow to perform feature processing on the multimodal data according to the data type includes: A multi-layer 1D-CNN is used to extract multi-scale features from the time series data to obtain time series features; Performing feature conversion and dimensionality reduction on the system indicator data using a fully connected network to obtain system features; The event data is subjected to an Embedding layer to convert discrete events into dense vector representations to obtain event features; The Transformer layer is called on the business data to extract internal correlation features through the self-attention mechanism to obtain business features.

5. The method according to claim 2, wherein The calling hybrid prediction network is based on the attention weight matrix and the temporal feature sequence, combined with multiple groups of prediction sample distributions generated by random dropout and feature contributions quantified based on the attention weight matrix, and the prediction results including confidence intervals and interpretability information include: The gating network is called to generate fusion weights based on the importance scores of cross-modal features; According to the fusion weight, dynamically weighted fusion is performed on the attention weight matrix and the temporal feature sequence to obtain a hybrid feature; The mixed features are subjected to nonlinear transformation and dimensional mapping by a multi-layer perceptron to generate prediction results of multi-dimensional resources.

6. The method according to claim 2, wherein The method for generating the explainability information includes: Call the SHAP library to calculate the global contribution of each type of feature in the time series feature sequence to the prediction result; Extracting key information of the hybrid prediction network in generating the prediction result; Performing gradient perturbation on the data corresponding to the key information and inputting the data into the hybrid prediction model to perform feature sensitivity analysis to obtain a sensitivity curve; The global contribution value, the key information, and the sensitivity curve are encapsulated into a structured information package as the interpretability information.

7. The method according to claim 1, wherein Based on the prediction results, a resource scheduling strategy is generated through a multi-objective optimization algorithm combined with dynamic thresholds, including: Extracting usage trends and fluctuation range features from the prediction results, and constructing an initial vector including resource type and time granularity; The initial vector is input into the target optimization model and solved by a multi-objective optimization algorithm to generate resource scheduling policy data including resource allocation priority and scheduling timing; wherein, the multi-objective optimization model sets resource cost and service quality optimization targets based on prediction characteristics and is constructed and generated in combination with historical scheduling rules.

8. The method according to claim 1, wherein Dynamically allocating resources according to the resource scheduling strategy includes: Using a policy parsing engine to automatically disassemble the resource scheduling policy and generate a resource allocation instruction set with priority; Based on the intelligent resource scheduling middleware, the container orchestration interface is called according to the priority of the resource allocation instruction set to perform elastic allocation of computing resources and dynamic mounting of storage resources.

9. The method according to claim 8, wherein Also includes: Monitor system operating indicators and operating status in real time and generate system monitoring data; Evaluate the effect based on the changes in the system monitoring data before and after deployment and generate evaluation results; The parameters of the hybrid deep learning model and the dynamic threshold are optimized based on the feedback of the evaluation results.

10. An intelligent operation and maintenance resource allocation device, characterized in that: include: The data collection and preprocessing module is used to collect multi-source data from full-link business scenarios and perform data preprocessing to obtain multi-dimensional preprocessing features; A deep learning prediction module is used to call a hybrid deep learning model to predict the future usage of multi-dimensional resources based on the multi-dimensional preprocessing features, and generate a prediction result including confidence intervals and interpretability information; An intelligent decision-making module, configured to generate a resource scheduling strategy based on the prediction results by using a multi-objective optimization algorithm combined with dynamic thresholds; The dynamic resource allocation execution module is used to dynamically allocate resources according to the resource scheduling strategy.

Images